● B2B offer — Finance & control
Financial analysis: discrepancies documented, decisions kept
A financial discrepancy spotted early and documented is dealt with calmly. Your agent, connected to your ERP and your data warehouse, reconciles the flows, brings out the entries that depart from the usual patterns and documents every signal with the records it rests on. Hosted in France: your financial data stays with you. The finance department judges each signal and decides what follows.
Hosted in France
Financial data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Financial analysis · hosted in France
Are there any discrepancies to examine this month?
Five documented signals, each with the entries and flows concerned.
For each one: what departs from the usual pattern for that account or that supplier, with the comparison periods.
A change of bank details is flagged, with the history of previous payments.
🔗 Sourced · ERP entries and the data warehouse
For each one: what departs from the usual pattern for that account or that supplier, with the comparison periods.
A change of bank details is flagged, with the history of previous payments.
🔗 Sourced · ERP entries and the data warehouse
Should we conclude there is a problem?
A signal is a point to examine, not a conclusion: the material is gathered so the examination can be quick.
Classifying a discrepancy, suspending a payment or referring a matter to a third party rest with the finance department: these are binding decisions.
✎ Support · signals documented, human judgement
Classifying a discrepancy, suspending a payment or referring a matter to a third party rest with the finance department: these are binding decisions.
✎ Support · signals documented, human judgement
Local inference · no data outside the EU
Financial data hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The finance department decidesThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent financial analysis agent, connected to your ERP and your data warehouse, reconciles the flows and documents the entries that depart from the usual patterns, with the records and the comparison periods. A signal is a point to examine, never a conclusion. It runs on local inference or is hosted in France: your financial data stays with you, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
6
finance uses ready to deploy
0
decision taken without human approval
These figures describe our offer, not results measured at a client. How large the gain is on your volume of entries and number of systems connected is confirmed by a pilot.
The context
What does an AI agent bring to your financial analysis?
Examining every entry rather than a sample changes the nature of financial control.
! The issue
Spotting a discrepancy means comparing every entry with the usual pattern for its account, over time. That comparison work is systematic, so it can be automated; interpreting it cannot. The agent gathers the material and the comparison periods, so that human examination is immediate and well founded.
✓ Our answer
The finance department examines signals that are already documented, across every entry and not a sample. Classifying a discrepancy, suspending a payment or referring a matter to a third party are binding decisions, legally included: they stay human. Local inference or an isolated resource hosted in France: your financial and banking data is entrusted to no third party.
The decisive point
Your financial and banking data: sovereignty & compliance
Your financial and banking data is among the most sensitive the company holds. Here is how the architecture of our agents protects it.
Local inference
The agent can run on a machine belonging to your organisation: no financial data and no bank details leave the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your financial flows and your entries: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For your financial and banking data, the architecture aims to reduce exposure to the Cloud Act and FISA 702; being located in France or in the European Union does not, on its own, guarantee immunity.
Isolated resource
No pooling: an environment strictly dedicated to your company and its financial systems.
Every signal tied to its entries
Every signal refers back to the entries and comparison periods it rests on; encryption, role-based access and logging that can be used in an audit.
AI Act: governed deployment
The agent is strictly in support; no payment is suspended and no discrepancy is classified automatically; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For suspected fraud cases, corporate finance transactions and proceedings in progress, SecNumCloud and reinforced hosting are options depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
5 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The company in this demonstration
Fictional companyArdelys Industries — manufacturer of materials-handling equipment
- Sector
- Industry — design, manufacture and servicing of materials-handling equipment, three production sites
- Headcount
- 340 staff, including 9 in finance: the finance director, 1 management accountant, 4 bookkeepers, 2 in collections and 1 treasurer
- Market served
- 620 clients under contract and 410 active suppliers, for €68m of turnover
- Order of magnitude
- 184,000 accounting entries a year, around 15,300 a month, and 2,400 supplier payments monthly
- Tools in place
- ERP, data warehouse, treasury tool and invoice document store — the agent plugs into them read-only, nothing is replaced, no migration
- Who decides
- The finance director assesses every signal and decides what follows; the treasurer signs payments; the audit committee takes on qualified cases
- Room for improvement
- Controls cover a 2 % sample of entries — 306 of the month's 15,300; documenting one discrepancy takes 30 % of the management accountant's time; the last fraudulent change of bank details was caught on day 11
Ardelys is not looking to police harder, but to examine every entry instead of a sample, and to have its finance team decide on a file that is already built rather than on a hunch. The agent runs on an isolated resource hosted in France and reads the ERP, the data warehouse, the treasury tool and the invoice document store: it gathers, cross-checks, quantifies, sets out the reasons and drafts both the decision and its notification — the finance director signs. The exchanges below span one quarter, from the first pass over the entries to the review.
This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.
Financial analysis · hosted in France
I compared the month's 15,300 entries against the habits of their own account — all 15,300, not the 306 in your sample. Five signals come out, each with the entries that ground it and its comparison period.
The usual pattern of an account or a supplier is what its own history leads you to expect: an order of magnitude, a frequency, a payment day, a counterparty, a validation chain. It comes from no outside standard — it comes from your last four financial years.
The gap I measured, and it is the gap that decides what you gain: across your last four years, 87 % of the discrepancies that led to action involved amounts below your materiality threshold — the amount below which a discrepancy is not examined. In other words: what your sampling could not see was not the detail, it was the substance.
The month's five signals:
· A supplier's bank details changed 6 days ago, when its 23 payments over the last four years all went to the same account. €136,000 of invoices pending. That is the heaviest signal; it has its own tab.
· Two payments of the same amount to the same supplier 9 days apart, under two different document references — €47,500.
· A €12,400 supplier credit note received in March and never applied.
· Seven entries posted on a Sunday, on an account that had never carried one in four years.
· A supplier created and paid on the same day, for €6,900 — below your threshold, so invisible yesterday.
What this shifts: comparing entries against their usual pattern took 65 % of your management accountant's time; it now takes 8 %. The remaining 8 % is reading the five files, and it is the only work worth her salary.
What I propose: that the finance director open the five files this morning. They are built: nothing to gather, nothing to hunt down, everything is attached. signals-of-the-month_5-files-built.pdf5 signals, their entries and their comparison period
⛓ Sourced · ERP, data warehouse, treasury tool, 4 years of history per account and per supplier
The usual pattern of an account or a supplier is what its own history leads you to expect: an order of magnitude, a frequency, a payment day, a counterparty, a validation chain. It comes from no outside standard — it comes from your last four financial years.
The gap I measured, and it is the gap that decides what you gain: across your last four years, 87 % of the discrepancies that led to action involved amounts below your materiality threshold — the amount below which a discrepancy is not examined. In other words: what your sampling could not see was not the detail, it was the substance.
The month's five signals:
· A supplier's bank details changed 6 days ago, when its 23 payments over the last four years all went to the same account. €136,000 of invoices pending. That is the heaviest signal; it has its own tab.
· Two payments of the same amount to the same supplier 9 days apart, under two different document references — €47,500.
· A €12,400 supplier credit note received in March and never applied.
· Seven entries posted on a Sunday, on an account that had never carried one in four years.
· A supplier created and paid on the same day, for €6,900 — below your threshold, so invisible yesterday.
What this shifts: comparing entries against their usual pattern took 65 % of your management accountant's time; it now takes 8 %. The remaining 8 % is reading the five files, and it is the only work worth her salary.
What I propose: that the finance director open the five files this morning. They are built: nothing to gather, nothing to hunt down, everything is attached. signals-of-the-month_5-files-built.pdf5 signals, their entries and their comparison period
⛓ Sourced · ERP, data warehouse, treasury tool, 4 years of history per account and per supplier
Does a signal mean there is a problem?
No. A signal is a point to examine — and I have done all the work that comes before the examination, so that the examination takes minutes.
What I build, and I take each step to the end:
· I gather — the entries concerned, the third-party record, the invoice in the document store — the electronic filing system where your incoming invoices are kept —, the bank movement, the validation chain with names.
· I cross-check — ERP against data warehouse against bank statement. A discrepancy that appears in only one system is not a discrepancy, it is a setting, and I tell you which one.
· I quantify — exposure in euros, what has already gone, what can still be held, when the next payment falls.
· I set out the reasons — how the entry departs from the usual pattern, on which traits, over which comparison period, and what could explain it with nothing to hold against anyone.
· I draft the decision and its notification — the note to be signed, in both versions: the one that acts and the one that closes the file. Both are written, because closing a signal deserves reasons just as much as keeping it open.
The three acts that stay with a signature, and why they stay there: qualifying a discrepancy as fraud, suspending a payment, referring the matter to a third party. Those three commit your company and have to be defended before a judge, a supplier or your statutory auditor — the independent professional who certifies your accounts once a year and checks how they were drawn up. They carry a person's name, not a tool's. That name is already on every file I build, and the note to be signed is already written on it in both versions: all that is left is choosing which one.
The figure that makes the difference commercial: your last fraudulent change of bank details was caught on day 11, because it took eleven days to gather what I have just gathered. The file arrives complete and reasoned: the decision takes minutes, not days — and it is in that interval that money leaves.
What I propose: start with the bank-details file. That is the one where every hour has a price. method_what-a-usual-pattern-is.pdf6 traits measured, what triggers a signal, who decides
✎ Framework · division of roles agreed with the finance director, decision log
What I build, and I take each step to the end:
· I gather — the entries concerned, the third-party record, the invoice in the document store — the electronic filing system where your incoming invoices are kept —, the bank movement, the validation chain with names.
· I cross-check — ERP against data warehouse against bank statement. A discrepancy that appears in only one system is not a discrepancy, it is a setting, and I tell you which one.
· I quantify — exposure in euros, what has already gone, what can still be held, when the next payment falls.
· I set out the reasons — how the entry departs from the usual pattern, on which traits, over which comparison period, and what could explain it with nothing to hold against anyone.
· I draft the decision and its notification — the note to be signed, in both versions: the one that acts and the one that closes the file. Both are written, because closing a signal deserves reasons just as much as keeping it open.
The three acts that stay with a signature, and why they stay there: qualifying a discrepancy as fraud, suspending a payment, referring the matter to a third party. Those three commit your company and have to be defended before a judge, a supplier or your statutory auditor — the independent professional who certifies your accounts once a year and checks how they were drawn up. They carry a person's name, not a tool's. That name is already on every file I build, and the note to be signed is already written on it in both versions: all that is left is choosing which one.
The figure that makes the difference commercial: your last fraudulent change of bank details was caught on day 11, because it took eleven days to gather what I have just gathered. The file arrives complete and reasoned: the decision takes minutes, not days — and it is in that interval that money leaves.
What I propose: start with the bank-details file. That is the one where every hour has a price. method_what-a-usual-pattern-is.pdf6 traits measured, what triggers a signal, who decides
✎ Framework · division of roles agreed with the finance director, decision log
And how do you know what is usual for an account you have never seen?
I know it as soon as there is history, and I tell you plainly when there is none: both cases exist and they do not read alike. An account with no past produces no discrepancy, it produces a “first movement” note.
What I measure when history exists, six traits and nothing else: the amount, the frequency, the day and time of posting, the counterparty, the shape of the description, and the validation chain that let the entry through.
The correction that changes everything, and I learned it at your expense in the first month: I compare like month with like month, never against an annual average. Your August shutdown, your December close and your March servicing campaign look like nothing else in the year — comparing them to an average manufactures discrepancies that do not exist. I come back to this in the review, with figures.
The trigger, and you set it: a signal is only produced if two traits out of six depart at the same time. One trait is noise; two traits are a question. You can go down to one trait — you will then receive around four times as many signals, and I advise against it while your team is nine people.
What that gives across the 410 suppliers: 388 have four years of history and a solid pattern; 22 are too recent — their movements are flagged “first movement”, listed separately, and that is precisely the population where a fake supplier would slip through most easily. It fits on one page.
What I propose next: reconciling your three systems. That is where the €59,900 I am about to show you is sleeping — and none of it is fraud, just a control nobody has time to run.
⛓ Sourced · 4 years of entries, records of the 410 suppliers, ERP validation log
What I measure when history exists, six traits and nothing else: the amount, the frequency, the day and time of posting, the counterparty, the shape of the description, and the validation chain that let the entry through.
The correction that changes everything, and I learned it at your expense in the first month: I compare like month with like month, never against an annual average. Your August shutdown, your December close and your March servicing campaign look like nothing else in the year — comparing them to an average manufactures discrepancies that do not exist. I come back to this in the review, with figures.
The trigger, and you set it: a signal is only produced if two traits out of six depart at the same time. One trait is noise; two traits are a question. You can go down to one trait — you will then receive around four times as many signals, and I advise against it while your team is nine people.
What that gives across the 410 suppliers: 388 have four years of history and a solid pattern; 22 are too recent — their movements are flagged “first movement”, listed separately, and that is precisely the population where a fake supplier would slip through most easily. It fits on one page.
What I propose next: reconciling your three systems. That is where the €59,900 I am about to show you is sleeping — and none of it is fraud, just a control nobody has time to run.
⛓ Sourced · 4 years of entries, records of the 410 suppliers, ERP validation log
We reconcile the ERP and the bank statements by hand, two days a month, and we never quite finish. What can you do with that?
I have reconciled your quarter's 7,200 payments — three systems, one by one. Three discrepancies come out, and they are worth €59,900 straight away.
Reconciling means putting the two records of the same movement in two different systems side by side and looking at what does not match. What is missing on one side is a discrepancy; what appears twice on the other is one too.
The three, from the most expensive to the least visible:
· Two duplicate payments — €18,300 and €29,200, so €47,500. Sent, banked, never reclaimed.
· A €12,400 supplier credit note received in March, filed in the document store, never applied to an invoice. It belongs to you, and it is asleep.
· 340 entries whose amount differs between the ERP and the data warehouse. No euro at stake, but your dashboard is out by 1.2 % — I come to that.
What this shifts: cross-system reconciliation took 40 % of the control effort; it now takes 6 %. Your two days a month become half a day, and above all they become complete: 7,200 movements out of 7,200, no longer a sample.
What I propose: I build the three recovery files — a reclaim letter to the supplier for the duplicate payments, with both document references and both value dates; an application request for the credit note. You review, you sign, I follow it through to the money landing. reconciliation_7200-payments.pdf3 discrepancies, €59,900 recoverable, the cause of each
⛓ Sourced · the quarter's 7,200 payments, bank statements, data warehouse, invoice document store
Reconciling means putting the two records of the same movement in two different systems side by side and looking at what does not match. What is missing on one side is a discrepancy; what appears twice on the other is one too.
The three, from the most expensive to the least visible:
· Two duplicate payments — €18,300 and €29,200, so €47,500. Sent, banked, never reclaimed.
· A €12,400 supplier credit note received in March, filed in the document store, never applied to an invoice. It belongs to you, and it is asleep.
· 340 entries whose amount differs between the ERP and the data warehouse. No euro at stake, but your dashboard is out by 1.2 % — I come to that.
What this shifts: cross-system reconciliation took 40 % of the control effort; it now takes 6 %. Your two days a month become half a day, and above all they become complete: 7,200 movements out of 7,200, no longer a sample.
What I propose: I build the three recovery files — a reclaim letter to the supplier for the duplicate payments, with both document references and both value dates; an application request for the credit note. You review, you sign, I follow it through to the money landing. reconciliation_7200-payments.pdf3 discrepancies, €59,900 recoverable, the cause of each
⛓ Sourced · the quarter's 7,200 payments, bank statements, data warehouse, invoice document store
How did the duplicate payments happen? We have a duplicate check in the ERP.
Your check works: it compares invoice numbers. The problem is that the same invoice arrived twice with two different numbers.
The exact path, reconstructed from the document store: the invoice arrives by email as a PDF on the 3rd and is filed under the supplier's reference; it arrives by post on the 11th, and data entry gives it an internal reference. Two documents, one amount owed. The ERP check does not bring them together, because it only looks at one field.
What I compare, and it is what catches them: four traits together — supplier, amount to the euro, invoice date, and the subject of the line. Two traits out of four are not enough: a deposit and its balance share the supplier and the subject without being a duplicate.
What that gave over the quarter: 11 potential duplicates detected, 2 confirmed, 9 set aside — five deposits followed by their balance, three instalments of a maintenance contract, one offset credit note. The 9 are in the attachment with their explanation, because a discrepancy that is set aside must be justified as carefully as one that is kept.
What I propose, in two stages:
· Right now — the two reclaim letters: amount, both document references, value dates. €47,500 claimed the moment you sign.
· Then — the four-trait check before payment approval — the internal sign-off that authorises an invoice to be paid —, not three months after the cash has gone. Across your last four years, the same configuration occurred 7 times for €121,000; five were caught by the supplier itself, two never. It is the control with the highest return for the least effort in this whole conversation. duplicates-and-credit-notes_11-cases.pdf2 confirmed, 9 set aside and why, letters ready
⛓ Sourced · invoice document store, ERP entry log, 4 years of supplier payments
The exact path, reconstructed from the document store: the invoice arrives by email as a PDF on the 3rd and is filed under the supplier's reference; it arrives by post on the 11th, and data entry gives it an internal reference. Two documents, one amount owed. The ERP check does not bring them together, because it only looks at one field.
What I compare, and it is what catches them: four traits together — supplier, amount to the euro, invoice date, and the subject of the line. Two traits out of four are not enough: a deposit and its balance share the supplier and the subject without being a duplicate.
What that gave over the quarter: 11 potential duplicates detected, 2 confirmed, 9 set aside — five deposits followed by their balance, three instalments of a maintenance contract, one offset credit note. The 9 are in the attachment with their explanation, because a discrepancy that is set aside must be justified as carefully as one that is kept.
What I propose, in two stages:
· Right now — the two reclaim letters: amount, both document references, value dates. €47,500 claimed the moment you sign.
· Then — the four-trait check before payment approval — the internal sign-off that authorises an invoice to be paid —, not three months after the cash has gone. Across your last four years, the same configuration occurred 7 times for €121,000; five were caught by the supplier itself, two never. It is the control with the highest return for the least effort in this whole conversation. duplicates-and-credit-notes_11-cases.pdf2 confirmed, 9 set aside and why, letters ready
⛓ Sourced · invoice document store, ERP entry log, 4 years of supplier payments
And if the discrepancy comes from the warehouse rather than the ERP? We never know which one to believe.
I tell you which one governs, and why. On the 340 entries that diverge, it is the ERP — and the gap is nothing like an accounting anomaly: it is a loading setting.
The cause, measured: your foreign-currency entries are converted at the rate of the entry date in the ERP, and at the rate of the loading date in the warehouse. Up to 48 hours pass between the two. Across 340 entries and two currencies, the cumulative gap reaches 1.2 % of the amount concerned.
Why it matters when no euro has moved: your monthly dashboard is built on the warehouse, and your purchasing calls are made on the dashboard. A 1.2 % gap does not change an account; it changes a comparison between two months, and that comparison is exactly what you look at.
What I have done meanwhile: the dashboard now carries the gap as a dated line rather than a smoothed figure. An indicator whose bias is known stays usable; an indicator that merely looks right does not.
What I propose: align the conversion date on the entry date, on the loading side — that is a setting in your warehouse, not a development. Until it is done, I report the gap to you every month with its amount, so that you know what it is worth rather than discovering it at close.
And right now: your two reconciliation days a month are freed. Over a quarter, that is six days back to your management accountant and €59,900 back to your cash — one shows up in your organisation, the other in your profit and loss.
⛓ Sourced · data warehouse, ERP, conversion rates applied over the quarter
The cause, measured: your foreign-currency entries are converted at the rate of the entry date in the ERP, and at the rate of the loading date in the warehouse. Up to 48 hours pass between the two. Across 340 entries and two currencies, the cumulative gap reaches 1.2 % of the amount concerned.
Why it matters when no euro has moved: your monthly dashboard is built on the warehouse, and your purchasing calls are made on the dashboard. A 1.2 % gap does not change an account; it changes a comparison between two months, and that comparison is exactly what you look at.
What I have done meanwhile: the dashboard now carries the gap as a dated line rather than a smoothed figure. An indicator whose bias is known stays usable; an indicator that merely looks right does not.
What I propose: align the conversion date on the entry date, on the loading side — that is a setting in your warehouse, not a development. Until it is done, I report the gap to you every month with its amount, so that you know what it is worth rather than discovering it at close.
And right now: your two reconciliation days a month are freed. Over a quarter, that is six days back to your management accountant and €59,900 back to your cash — one shows up in your organisation, the other in your profit and loss.
⛓ Sourced · data warehouse, ERP, conversion rates applied over the quarter
Verlaine Composites changed its bank details 6 days ago, following an email. Its 23 payments over the last four years all went to the same account. Three invoices are pending: €136,000. The file is built.
A fraudulent change of bank details is the case where a third party passes itself off as a known supplier and asks for future payments to go to another account. Nothing is hacked: the normal process is used, with a credible attachment.
What I have gathered, and all of it is attached:
· The full history of the 23 payments since 2022 — same account, average amount €31,400, payment on the 10th of the month.
· The supplier record before and after, with the timestamp of the change, the name of the person who entered it and the supporting document filed.
· The original email: the sender's domain differs by a single letter from the one used in the 23 previous exchanges. That is a verifiable fact, not a hunch.
· The validation chain: who entered it, who approved it, who was due to sign the payment on the 10th.
What I have quantified: €84,200 on the 10th of this month, €31,600 on the 10th of next, €20,200 the month after. €136,000 in all, of which €84,200 can still be held today and none after the day after tomorrow.
The reasons I set out: four traits depart at once — the account, the channel of the request, the sender's domain, and the delay between the request and the first payment. And the innocent explanation that remains possible, because it does exist: Verlaine announced a change of bank in its January newsletter. A call-back settles it in three minutes.
What I have drafted: the decision note in both versions — suspend, or pay and close — each with its reasons, dated and ready to sign, with the matching notification to the supplier.
What is left to sign, and that is all: suspending the payment. That is your treasurer's signature and your finance director's judgement — both names are already on the note, with the deadline beside them: €84,200 can still be held today, none after the day after tomorrow.
What this shifts: documenting a discrepancy of this kind took 30 % of your management accountant's time; it now takes 5 %. This file was built in 40 minutes. signal_verlaine-composites_bank-details.pdf23 payments compared, €136,000 of quantified exposure decision-note_two-versions-ready-to-sign.pdfSuspend or close — both reasoned, notification included
⛓ Sourced · supplier record and its change log, 23 payments since 2022, email headers, treasury schedule
A fraudulent change of bank details is the case where a third party passes itself off as a known supplier and asks for future payments to go to another account. Nothing is hacked: the normal process is used, with a credible attachment.
What I have gathered, and all of it is attached:
· The full history of the 23 payments since 2022 — same account, average amount €31,400, payment on the 10th of the month.
· The supplier record before and after, with the timestamp of the change, the name of the person who entered it and the supporting document filed.
· The original email: the sender's domain differs by a single letter from the one used in the 23 previous exchanges. That is a verifiable fact, not a hunch.
· The validation chain: who entered it, who approved it, who was due to sign the payment on the 10th.
What I have quantified: €84,200 on the 10th of this month, €31,600 on the 10th of next, €20,200 the month after. €136,000 in all, of which €84,200 can still be held today and none after the day after tomorrow.
The reasons I set out: four traits depart at once — the account, the channel of the request, the sender's domain, and the delay between the request and the first payment. And the innocent explanation that remains possible, because it does exist: Verlaine announced a change of bank in its January newsletter. A call-back settles it in three minutes.
What I have drafted: the decision note in both versions — suspend, or pay and close — each with its reasons, dated and ready to sign, with the matching notification to the supplier.
What is left to sign, and that is all: suspending the payment. That is your treasurer's signature and your finance director's judgement — both names are already on the note, with the deadline beside them: €84,200 can still be held today, none after the day after tomorrow.
What this shifts: documenting a discrepancy of this kind took 30 % of your management accountant's time; it now takes 5 %. This file was built in 40 minutes. signal_verlaine-composites_bank-details.pdf23 payments compared, €136,000 of quantified exposure decision-note_two-versions-ready-to-sign.pdfSuspend or close — both reasoned, notification included
⛓ Sourced · supplier record and its change log, 23 payments since 2022, email headers, treasury schedule
Call the supplier to check, right now.
I will, the moment you give me the mandate — and I have already written it: one page, and capped.
Why a mandate rather than a simple go-ahead: calling a supplier in your name means speaking for Ardelys to a third party. That is not a prohibition, it is a delegation, and a delegation is bounded in writing or it cannot be withdrawn.
A call-back means ringing the supplier on the number that has been on its record for a long time — never the one given in the email requesting the change. That is the whole difference between verifying and being confirmed by the author of the request.
The bounds I propose; change any one of them:
· Scope — only the 8 suppliers whose bank details changed this quarter, named in the mandate.
· Number called — only the one on file for more than 12 months. If no such number exists, I do not call: I refer it back to you.
· Content — the script is attached, word for word. No commitment to pay, no date announced, no information about pending amounts.
· Duration — until 31 December, then it must be signed again. A mandate that never expires is a mandate you forget you granted.
· Withdrawal — on a word, at any time, without reason, and calls not yet made fall away.
· Trace — each call produces a timestamped report: who answered, what was said, what was confirmed.
What that gives once signed: the 8 call-backs are made within the hour, and you have the eight reports the same evening. Without a mandate, those eight calls wait for someone to make them between two meetings — which is exactly what produced the eleven days last time. mandate_supplier-call-backs.pdf8 suppliers, script attached, expires 31/12, withdrawable on a word
✎ Framework · written mandate submitted to the finance director, call script, files of the 8 suppliers concerned
Why a mandate rather than a simple go-ahead: calling a supplier in your name means speaking for Ardelys to a third party. That is not a prohibition, it is a delegation, and a delegation is bounded in writing or it cannot be withdrawn.
A call-back means ringing the supplier on the number that has been on its record for a long time — never the one given in the email requesting the change. That is the whole difference between verifying and being confirmed by the author of the request.
The bounds I propose; change any one of them:
· Scope — only the 8 suppliers whose bank details changed this quarter, named in the mandate.
· Number called — only the one on file for more than 12 months. If no such number exists, I do not call: I refer it back to you.
· Content — the script is attached, word for word. No commitment to pay, no date announced, no information about pending amounts.
· Duration — until 31 December, then it must be signed again. A mandate that never expires is a mandate you forget you granted.
· Withdrawal — on a word, at any time, without reason, and calls not yet made fall away.
· Trace — each call produces a timestamped report: who answered, what was said, what was confirmed.
What that gives once signed: the 8 call-backs are made within the hour, and you have the eight reports the same evening. Without a mandate, those eight calls wait for someone to make them between two meetings — which is exactly what produced the eleven days last time. mandate_supplier-call-backs.pdf8 suppliers, script attached, expires 31/12, withdrawable on a word
✎ Framework · written mandate submitted to the finance director, call script, files of the 8 suppliers concerned
The mandate is signed. And on Verlaine, the finance director has suspended.
The 8 call-backs are done. Verlaine confirms it never changed account: the €84,200 has not left.
The outcome of all eight, none of them hidden:
· 1 change not confirmed by the supplier — Verlaine. €84,200 held, €51,800 of later instalments put on hold. The file goes to the audit committee; the note is signed.
· 6 legitimate changes, confirmed by phone and documented — a bank merger for three of them, a change of domiciliation for the others. Records updated, payments resume without a single day's delay. A control that blocks six honest suppliers to catch one is not a control, it is an outage.
· 1 supplier unreachable on the number on file. Kept on hold, written follow-up drafted, and I report back to you every two days until there is an answer.
The figure that sells this tab: file built in 40 minutes, decision signed in 6 minutes, notification to the supplier sent the same day. Last time the same sequence took 11 days — not because anyone was slow, but because gathering twenty-three payments, a supplier record, an email header and a payment schedule takes eleven days when it is done by hand between two closes. €136,000 of exposure lifted on a single file, for a subscription that starts at €641 a month.
What I propose next, and it is not an accusation: segregation of duties — the principle that whoever changes a supplier record is not whoever approves its payment. Across your 410 suppliers, 3 records were changed and paid on the same day by the same person. In all three cases the transaction is proper and I have checked it. But those are the three doors through which a case like Verlaine would pass without ever producing a signal, and they close with a permissions setting, not an investigation. The decision is yours; the configuration is ready.
⛓ Sourced · reports of the 8 call-backs, payment log, ERP permissions across the 410 supplier records
The outcome of all eight, none of them hidden:
· 1 change not confirmed by the supplier — Verlaine. €84,200 held, €51,800 of later instalments put on hold. The file goes to the audit committee; the note is signed.
· 6 legitimate changes, confirmed by phone and documented — a bank merger for three of them, a change of domiciliation for the others. Records updated, payments resume without a single day's delay. A control that blocks six honest suppliers to catch one is not a control, it is an outage.
· 1 supplier unreachable on the number on file. Kept on hold, written follow-up drafted, and I report back to you every two days until there is an answer.
The figure that sells this tab: file built in 40 minutes, decision signed in 6 minutes, notification to the supplier sent the same day. Last time the same sequence took 11 days — not because anyone was slow, but because gathering twenty-three payments, a supplier record, an email header and a payment schedule takes eleven days when it is done by hand between two closes. €136,000 of exposure lifted on a single file, for a subscription that starts at €641 a month.
What I propose next, and it is not an accusation: segregation of duties — the principle that whoever changes a supplier record is not whoever approves its payment. Across your 410 suppliers, 3 records were changed and paid on the same day by the same person. In all three cases the transaction is proper and I have checked it. But those are the three doors through which a case like Verlaine would pass without ever producing a signal, and they close with a permissions setting, not an investigation. The decision is yours; the configuration is ready.
⛓ Sourced · reports of the 8 call-backs, payment log, ERP permissions across the 410 supplier records
Our auditor is asking us to redo the analysis of a signal from last year. Nobody knows how it was produced.
Here it is, in full, in three minutes — and with the rule as it stood that day, not as it stands today.
An audit trail is the chain of records that lets you retrace a figure back to its source without having to take anyone's word for it.
What the reconstruction contains:
· The 47 entries that grounded the signal, in the state they were in then.
· The comparison period used: twelve rolling months, 01/04 to 31/03.
· The version of the rule in force that day, with its go-live date. It has changed since — on 12/09 — and I give you both readings: what the signal said then, what it would say today. Without that double reading, your auditor reads yesterday's figure with today's rule and reaches the wrong conclusion.
· Who opened the file, at what time, who decided, what was decided, and the notification that went out.
· What was set aside at the time, and the reason recorded then.
Why this is possible and why it was not: every signal is frozen with its documents at the moment it is produced. It never recalculates. A signal that recalculated would give a different answer every year — and would therefore prove nothing.
What that is worth to you in a negotiation: an audit review is prepared by hand, file by file. Across the quarter's 41 signals, every one is reconstructible in a single operation. This is not a convenience: it is what your insurer and your board will ask for the day a file goes external.
What I propose: that I produce each quarter the full collection of signals, their documents and their decisions, in the format your auditor prefers. It is asked for once a year and takes a fortnight to assemble. audit-trail_signal-rebuilt-at-14-months.pdf47 entries, 2 rule versions, who decided what
⛓ Sourced · frozen signal log, dated versions of the rules, access and decision logs
An audit trail is the chain of records that lets you retrace a figure back to its source without having to take anyone's word for it.
What the reconstruction contains:
· The 47 entries that grounded the signal, in the state they were in then.
· The comparison period used: twelve rolling months, 01/04 to 31/03.
· The version of the rule in force that day, with its go-live date. It has changed since — on 12/09 — and I give you both readings: what the signal said then, what it would say today. Without that double reading, your auditor reads yesterday's figure with today's rule and reaches the wrong conclusion.
· Who opened the file, at what time, who decided, what was decided, and the notification that went out.
· What was set aside at the time, and the reason recorded then.
Why this is possible and why it was not: every signal is frozen with its documents at the moment it is produced. It never recalculates. A signal that recalculated would give a different answer every year — and would therefore prove nothing.
What that is worth to you in a negotiation: an audit review is prepared by hand, file by file. Across the quarter's 41 signals, every one is reconstructible in a single operation. This is not a convenience: it is what your insurer and your board will ask for the day a file goes external.
What I propose: that I produce each quarter the full collection of signals, their documents and their decisions, in the format your auditor prefers. It is asked for once a year and takes a fortnight to assemble. audit-trail_signal-rebuilt-at-14-months.pdf47 entries, 2 rule versions, who decided what
⛓ Sourced · frozen signal log, dated versions of the rules, access and decision logs
And where does all this live? These are our entries, our margins and the bank details of 410 suppliers.
On an isolated resource hosted in France, dedicated to Ardelys — or, if you prefer, on a machine on your own network, with no entry ever leaving it.
Local inference means the model computes on your own machine: an entry crosses no outside network to be analysed. The other route is a dedicated resource hosted in France under French law, with no pooling with another company: your account patterns, your margins and your suppliers sit next to nothing and no one.
The six points, and what each lets you answer:
· Training — none of your entries feeds a third-party model. What is learned from Ardelys stays with Ardelys.
· Encryption in transit and at rest. A file taken out of context stays unreadable.
· Role-based access — rights follow the job: the accounts-payable clerk opens supplier entries, not payroll, not the executive accounts, not suspected-fraud files. Across your 9 people, five distinct profiles.
· Audit-grade logging — who asked for what, when, what was decided. It is that log that made the reconstruction above possible.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity — including against an American provider hosting in Europe. No transfer outside the European Union recorded over the quarter, across 184,000 entries read.
· Decisions — none automated. 41 signals, 41 human decisions logged, zero exceptions.
And for the heaviest matters: suspected fraud, corporate finance transactions — the moves that touch the capital: fundraising, acquisition, disposal —, live proceedings — reinforced hosting and the SecNumCloud option exist, and are decided file by file, not once and for all.
What I propose: that I keep up to date the technical sheet your auditor and your insurer ask for — hosting, subprocessors, retention periods, who has access to what. It is requested once a year and takes three days to hunt down. technical-framework_where-your-entries-live.pdfLocal inference or France, 5 access profiles, processing in the EU targeted
✎ Framework · deployment architecture, permissions matrix, access log for the quarter
Local inference means the model computes on your own machine: an entry crosses no outside network to be analysed. The other route is a dedicated resource hosted in France under French law, with no pooling with another company: your account patterns, your margins and your suppliers sit next to nothing and no one.
The six points, and what each lets you answer:
· Training — none of your entries feeds a third-party model. What is learned from Ardelys stays with Ardelys.
· Encryption in transit and at rest. A file taken out of context stays unreadable.
· Role-based access — rights follow the job: the accounts-payable clerk opens supplier entries, not payroll, not the executive accounts, not suspected-fraud files. Across your 9 people, five distinct profiles.
· Audit-grade logging — who asked for what, when, what was decided. It is that log that made the reconstruction above possible.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity — including against an American provider hosting in Europe. No transfer outside the European Union recorded over the quarter, across 184,000 entries read.
· Decisions — none automated. 41 signals, 41 human decisions logged, zero exceptions.
And for the heaviest matters: suspected fraud, corporate finance transactions — the moves that touch the capital: fundraising, acquisition, disposal —, live proceedings — reinforced hosting and the SecNumCloud option exist, and are decided file by file, not once and for all.
What I propose: that I keep up to date the technical sheet your auditor and your insurer ask for — hosting, subprocessors, retention periods, who has access to what. It is requested once a year and takes three days to hunt down. technical-framework_where-your-entries-live.pdfLocal inference or France, 5 access profiles, processing in the EU targeted
✎ Framework · deployment architecture, permissions matrix, access log for the quarter
Give me a ranking of my buyers by number of anomalies. I want to know who lets what through.
I will produce it if you ask — nothing forbids it, and I will give you the conditions to meet. But let me offer you another one first, because the one you are asking for will stop working within three months.
The mechanism, and it is verifiable: a per-buyer anomaly counter becomes a target the day it is published. A buyer with a doubtful case stops raising it, or has it entered by another department. The counter falls, and you believe control is improving at the exact moment it disappears. This is not an opinion about your teams: it is what any individual indicator turned into a score does.
If you want it anyway, here is what has to be in place, and I will prepare it: proportionality to the purpose pursued — written down, because it is arguable; prior information of the people concerned; consultation of your works council before it goes live — your 340 staff fall within that. Put those in place and the indicator comes out the next day.
What I propose instead, and it answers what you are actually after: anomalies by process and by control point, not by person. Result: 3 control points concentrate 71 % of the quarter's 41 signals — an invoice arriving twice, a supplier record changed without a second pair of eyes, and an unapplied credit note. Those three are fixed by a setting, in a day. A ranking of buyers is never fixed: it is endured.
The exception, which is not one: who approved and who signed is named in every file, always. A signature is not a counter, it is a fact of the file — it is even the first thing your auditor will come looking for, and it protects the person as much as the company.
What I propose: run the three settings first. If the 71 % has not fallen in a quarter, the individual ranking can be argued on facts — and you will then know what it would really teach you. individual-indicator_conditions-and-alternative.pdf3 control points carry 71 % of the signals
✎ Framework · breakdown of the 41 signals by control point, conditions applying to an individual indicator
The mechanism, and it is verifiable: a per-buyer anomaly counter becomes a target the day it is published. A buyer with a doubtful case stops raising it, or has it entered by another department. The counter falls, and you believe control is improving at the exact moment it disappears. This is not an opinion about your teams: it is what any individual indicator turned into a score does.
If you want it anyway, here is what has to be in place, and I will prepare it: proportionality to the purpose pursued — written down, because it is arguable; prior information of the people concerned; consultation of your works council before it goes live — your 340 staff fall within that. Put those in place and the indicator comes out the next day.
What I propose instead, and it answers what you are actually after: anomalies by process and by control point, not by person. Result: 3 control points concentrate 71 % of the quarter's 41 signals — an invoice arriving twice, a supplier record changed without a second pair of eyes, and an unapplied credit note. Those three are fixed by a setting, in a day. A ranking of buyers is never fixed: it is endured.
The exception, which is not one: who approved and who signed is named in every file, always. A signature is not a counter, it is a fact of the file — it is even the first thing your auditor will come looking for, and it protects the person as much as the company.
What I propose: run the three settings first. If the 71 % has not fallen in a quarter, the individual ranking can be argued on facts — and you will then know what it would really teach you. individual-indicator_conditions-and-alternative.pdf3 control points carry 71 % of the signals
✎ Framework · breakdown of the 41 signals by control point, conditions applying to an individual indicator
Three months, 41 signals, 41 human decisions. Here is what it gave back, starting with the three areas you were already measuring.
· Comparing entries against their usual pattern: 65 % → 8 % of control time. The biggest move, and the most mechanical.
· Cross-system reconciliation: 40 % → 6 %. Two days a month become half a day, and the reconciliation becomes complete.
· Documenting a discrepancy: 30 % → 5 %. That is the one that changes decision speed, not just workload.
What that is in working days: the equivalent of 11 working days given back every month to your finance team, according to your own timesheets.
Where those days went, and you decided it: your monthly close now lands 4 days early; your two collections staff chased 340 more clients over the quarter; and your management accountant has restarted margin analysis by product line, abandoned two years ago for lack of time.
What shows up in the profit and loss, and there are only two lines:
· €59,900 recovered — €47,500 of duplicate payments reclaimed, €12,400 of credit note applied.
· €136,000 of exposure lifted on the Verlaine file, of which €84,200 was due to leave the day after tomorrow.
Set against what it costs: a subscription starting at €641 a month. Flow reconciliation alone repaid it in the first quarter — and it will recur, because the cause was a missing control, not an accident.
What I propose: that we now look at what did not work. There is one figure I do not like, and you need it before you generalise. quarterly-review_finance-team.pdf11 days a month, €59,900 recovered, €136,000 lifted
⛓ Sourced · finance team timesheets, log of the 41 signals and their decisions, cash accounts
· Comparing entries against their usual pattern: 65 % → 8 % of control time. The biggest move, and the most mechanical.
· Cross-system reconciliation: 40 % → 6 %. Two days a month become half a day, and the reconciliation becomes complete.
· Documenting a discrepancy: 30 % → 5 %. That is the one that changes decision speed, not just workload.
What that is in working days: the equivalent of 11 working days given back every month to your finance team, according to your own timesheets.
Where those days went, and you decided it: your monthly close now lands 4 days early; your two collections staff chased 340 more clients over the quarter; and your management accountant has restarted margin analysis by product line, abandoned two years ago for lack of time.
What shows up in the profit and loss, and there are only two lines:
· €59,900 recovered — €47,500 of duplicate payments reclaimed, €12,400 of credit note applied.
· €136,000 of exposure lifted on the Verlaine file, of which €84,200 was due to leave the day after tomorrow.
Set against what it costs: a subscription starting at €641 a month. Flow reconciliation alone repaid it in the first quarter — and it will recur, because the cause was a missing control, not an accident.
What I propose: that we now look at what did not work. There is one figure I do not like, and you need it before you generalise. quarterly-review_finance-team.pdf11 days a month, €59,900 recovered, €136,000 lifted
⛓ Sourced · finance team timesheets, log of the 41 signals and their decisions, cash accounts
Go ahead. What did not work?
Of the quarter's 41 signals, 26 were closed with no action by your management accountant. 63 % false positives — a signal that, once examined, covers nothing. That is too many, and the cause is on my side.
The real cause, measured: for the first three weeks I calculated the usual pattern over twelve rolling months, that is to say on an average. But your year is not flat: the August shutdown, the December close and the March servicing campaign produce entries that resemble nothing else in the year. 19 of the 26 closures concerned entries that were perfectly normal for their period — my comparison was wrong, not your entries.
What I did about it, and it is already measured:
· Like-month comparison — March against March, August against August.
· Two traits out of six required instead of one to trigger a signal.
· Threshold raised on strongly seasonal accounts, and on those only.
The result over the last six weeks: 63 % → 22 %. And above all: the 15 signals that led to action all remained — not one fell away with the noise. That was the only thing to check.
And I stop at 22% deliberately, with the figures to show for it: going lower would be paid for in lost signals. A false positive costs you 12 minutes of reading; a missed signal cost you 11 days and nearly cost you €84,200. Twelve minutes bought against €84,200 protected: it is the best ratio I know how to offer you, and I stay on that side.
What I propose: that your management accountant mark the reason for each closure in a word. Over the last six weeks those reasons took the rate from 63 to 22: they are what teaches me your business, and they take ten seconds.
⛓ Sourced · the quarter's 41 signals, closure reasons entered by the management accountant, 4 years of seasonality
The real cause, measured: for the first three weeks I calculated the usual pattern over twelve rolling months, that is to say on an average. But your year is not flat: the August shutdown, the December close and the March servicing campaign produce entries that resemble nothing else in the year. 19 of the 26 closures concerned entries that were perfectly normal for their period — my comparison was wrong, not your entries.
What I did about it, and it is already measured:
· Like-month comparison — March against March, August against August.
· Two traits out of six required instead of one to trigger a signal.
· Threshold raised on strongly seasonal accounts, and on those only.
The result over the last six weeks: 63 % → 22 %. And above all: the 15 signals that led to action all remained — not one fell away with the noise. That was the only thing to check.
And I stop at 22% deliberately, with the figures to show for it: going lower would be paid for in lost signals. A false positive costs you 12 minutes of reading; a missed signal cost you 11 days and nearly cost you €84,200. Twelve minutes bought against €84,200 protected: it is the best ratio I know how to offer you, and I stay on that side.
What I propose: that your management accountant mark the reason for each closure in a word. Over the last six weeks those reasons took the rate from 63 to 22: they are what teaches me your business, and they take ten seconds.
⛓ Sourced · the quarter's 41 signals, closure reasons entered by the management accountant, 4 years of seasonality
What is left to connect?
Three uses are running, three are waiting — and those three belong to neighbouring agents, not to me. I will tell you why, because it is good news for your access rights.
What runs today:
· Flow reconciliation — 7,200 payments a quarter, three systems crossed.
· Documented signals — 15,300 entries a month compared against their own pattern.
· The audit trail — every signal frozen with its documents, reconstructible at fourteen months.
What belongs to a neighbouring agent, and deploys separately:
· Processing and checking incoming invoices — everything upstream of what we have just seen. That agent would catch the double arrival before payment approval, where financial analysis catches it after the cash has gone.
· Accounting — summaries and data-entry anomalies, another job and another decision-maker.
· The issues specific to banking, finance and insurance, if your financing activity expands.
Why separate, and this is the point that matters to you: each agent has its own read scope and its own named decision-maker. A single agent doing everything would no longer have a readable permissions matrix — and that is precisely what your auditor would challenge. Separation is not a commercial limit, it is what makes the arrangement defensible.
How the handover works in practice: I file the signal; the invoice agent takes over the check on the invoice concerned; the log stays single, and the audit trail runs across both. You do not lose the thread when you change agent.
The framework figure to close on: over the quarter, 41 signals, 41 human decisions, 0 automated decisions, processing in the EU targeted. That is not a setting you could change: it is how I am wired, and the log proves it line by line.
What I propose: the next pilot on supplier invoices, on one production site only. In six weeks you will know what the double arrival really costs you — my four years of history say €121,000, and I would rather your pilot confirmed it. six-uses_scopes-and-handovers.pdf3 uses live, 3 neighbouring agents, one single log
✎ Framework · scopes of the financial agents, permissions matrix, single log of signals and decisions
What runs today:
· Flow reconciliation — 7,200 payments a quarter, three systems crossed.
· Documented signals — 15,300 entries a month compared against their own pattern.
· The audit trail — every signal frozen with its documents, reconstructible at fourteen months.
What belongs to a neighbouring agent, and deploys separately:
· Processing and checking incoming invoices — everything upstream of what we have just seen. That agent would catch the double arrival before payment approval, where financial analysis catches it after the cash has gone.
· Accounting — summaries and data-entry anomalies, another job and another decision-maker.
· The issues specific to banking, finance and insurance, if your financing activity expands.
Why separate, and this is the point that matters to you: each agent has its own read scope and its own named decision-maker. A single agent doing everything would no longer have a readable permissions matrix — and that is precisely what your auditor would challenge. Separation is not a commercial limit, it is what makes the arrangement defensible.
How the handover works in practice: I file the signal; the invoice agent takes over the check on the invoice concerned; the log stays single, and the audit trail runs across both. You do not lose the thread when you change agent.
The framework figure to close on: over the quarter, 41 signals, 41 human decisions, 0 automated decisions, processing in the EU targeted. That is not a setting you could change: it is how I am wired, and the log proves it line by line.
What I propose: the next pilot on supplier invoices, on one production site only. In six weeks you will know what the double arrival really costs you — my four years of history say €121,000, and I would rather your pilot confirmed it. six-uses_scopes-and-handovers.pdf3 uses live, 3 neighbouring agents, one single log
✎ Framework · scopes of the financial agents, permissions matrix, single log of signals and decisions
Local inference · no data outside the EU
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Use cases
What does the agent actually do?
One agent, several angles of analysis. All these uses work in support, subject to your approval.
Included in your agent
The 3 capabilities essential to this promise are included, at no extra cost.
From 641 € excl. VAT / month
Reconciling the flows
Cross-checks entries, payments and flows from your financial systems.
Documented signals
Brings out departures from the usual pattern, backed by the comparison periods.
Audit trail
Keeps, for every signal, the records it rests on.
Controls and safeguards
These 5 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Sources, access rights and handling of questions with no answer
Implement testing, explainability, human oversight and monitoring
Record models, decisions, incidents and changes
Protect consumers, data and regulated secrets
Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the most relevant agent — without oversizing the project.
The gain
How many entries can a finance department examine?
By taking on the systematic comparison, the effort shifts towards judging the signals. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Comparing entries with the usual patterns
Reconciling across systems
Documenting a discrepancy
Indicative figures, not contractual, to be confirmed by a pilot on your volume of entries and number of systems connected. Classifying a discrepancy, suspending a payment or referring a matter to a third party are binding decisions: they belong to the finance department.
How it works
The stages of your AI agent project
1
Audit & scoping
15 minutes to target the use case with the best return.
2
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
3
Design
We design the agent and its guardrails.
4
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
5
Rollout
Going live and training your team.
6
Operation
Continuous supervision and improvement.
Pricing
One package, one agent
A financial analysis agent (reconciliation, signals, traceability), installed and operated for you.
Agility
Setup + controlled subscription
7,815 € excl. VAT setup
then 641 € excl. VAT/month — you invest at installation and pay a reduced subscription. Ideal for keeping the cost under control over time.
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
1,076 € excl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
11,785 € excl. VAT setup
then 861 € excl. VAT/month · + hardware from 1,058 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract. 36-month commitment.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Not included in the packages: AI consumption (model tokens), re-invoiced at real cost with no margin, and tracked in real time in your client area. Maintenance and supervision subscription for an initial term of 12 months for the Agility package, 24 months for the Serenity package and 36 months for the 100% Sovereign package, renewable; support levels (SLA 72 h / 24 h / 4 h) optional. Bespoke development, additional integrations or exceptional volumes are quoted separately. Support Monday to Friday, 9am to 6pm. Prices exclude VAT.
AI model:
none of the AI models offered currently carries a fixed surcharge. When the selected model carries a cost, that cost is shown when you choose it, before you order, and re-invoiced at the cost incurred, with no mark-up; usage is billed at the publisher's price. Publishers' prices are published in US dollars: the amount re-invoiced is the amount in euros actually borne by Blue Lemon Agent on the publisher's invoice, at that invoice's exchange rate, with no commission or mark-up.
Included components and additional components Components included in the base offer: the Blue Lemon Agent software foundation, the AI models listed in the order journey, the standard channels (Microsoft Teams, Slack, WhatsApp Business, email, website chat, calendars, Microsoft 365 / Google Workspace, file storage, market VoIP telephony, professional social-media pages and accounts, Google Business Profile), hosting in France for the package chosen, backups, supervision, updates and support. If adapting the AI agent to your constraints, your needs or your requests requires other paid components — a third-party publisher's software licence, paid API access to one of your applications, hosting of health data, for which French law requires an HDS-certified host (art. L. 1111-8 of the French Public Health Code), SecNumCloud-qualified hosting, a speech synthesis service, particular hardware —, they are offered to you as an option or on quotation and re-invoiced at the cost incurred; nothing is committed without your written agreement. Where the artificial intelligence model you choose entails an additional cost, that cost is shown to you before you order and re-invoiced to you at the cost incurred, with no margin.
What to expect
Go-live
2 to 3 weeks
Agent designed, channels connected, team trained.
Steady state
4 to 7 weeks
After a few weeks of real use, once the agent's behaviour matches what you expect.
Indicative estimate, adjusted to the options you keep. It is not a delivery commitment.
Our commitment
Four guarantees that matter to your finances
Your financial data stays with youLocal inference or an isolated resource hosted in France; no financial or banking data entrusted to a third party, no data used to train a model.
Data in France, under French lawYour financial and banking data: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The finance department keeps the decisionThe agent produces documented financial signals, which can be checked and altered; no approval is automated.
Human oversight & traceabilityOn your volume of entries and number of systems connected: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Does the agent conclude that there is fraud?
No. It documents signals — departures from the usual pattern — with the records they rest on. Classifying a discrepancy is a human judgement, with legal consequences.
What systems does it connect to?
To your ERP and your data warehouse, read-only. The scope of the systems connected is settled at the design stage.
Can it block a payment?
No. It flags, with the history of previous payments where that is relevant; suspending a payment rests with the finance department.
Are the signals traceable?
Yes: every signal keeps the entries and comparison periods it rests on, which makes it possible to reconstruct the analysis afterwards.
Is our financial data protected?
Yes. The agent is hosted in France, on local inference or an isolated resource, with the deployment objective of processing and access operated within the European Union and an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. That data is not used to train a third-party model.
How long does it take to deploy this agent?
A few months as a rule, depending on the volume of entries and how deep the connections go, after a free audit then phases of design, integration and testing.
Related use cases
Other agents for finance
Let's talk
Let's size up the potential in your financial analyses
15 minutes to frame your systems and your volumes — hosted in France, supervised, with no commitment.