● B2B offer — Banking compliance
Banking compliance: files assembled, evidence kept
A know-your-customer file and an anti-money-laundering check rest on documents, verifications and a flawless audit trail. Your agent assembles these files according to your procedures, gathers the documents expected, applies your control rules and keeps the evidence of every verification. Hosted in France in a qualified environment: your customers' data stays with you. Every decision rests with the authorised compliance officer.
Hosted in France
Client data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
KYC/AML compliance · hosted in France
Assemble the know-your-customer file for this new business relationship.
File assembled according to your procedures: identification documents gathered, consistency of the information verified, the checks in your control framework applied.
Every verification keeps its source, its date and its result.
Two points call for the judgement of an authorised officer: they are presented at the top.
🔗 Sourced · the documents in the file and the control framework
Every verification keeps its source, its date and its result.
Two points call for the judgement of an authorised officer: they are presented at the top.
🔗 Sourced · the documents in the file and the control framework
Can we approve the business relationship?
The file is complete and documented, with the two points flagged and the material that bears on them.
Approving a business relationship, or reporting a transaction, rests with the authorised officer: these are regulated acts.
✎ Support · file documented, the officer's decision
Approving a business relationship, or reporting a transaction, rests with the authorised officer: these are regulated acts.
✎ Support · file documented, the officer's decision
Local inference · no data outside the EU
Files hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
Your banking compliance department decidesThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent banking compliance agent assembles your know-your-customer files and your anti-money-laundering checks according to your procedures: documents gathered, consistency verified, rules applied, with every verification keeping its source, its date and its result. Approving a business relationship or reporting a transaction rests with the authorised officer. Hosted in France in a qualified environment, 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
compliance 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 files and the breadth of your control framework is confirmed by a pilot.
The context
What does an AI agent bring to your banking compliance?
In banking compliance, the quality of the evidence kept counts as much as the check performed.
! The issue
A banking compliance check calls for gathering documents, applying a control framework and keeping the evidence of every verification. All three requirements are methodical and lend themselves to systematic execution. The judgement, on the other hand, rests with an authorised officer — the agent prepares and documents it.
✓ Our answer
Your compliance department has files that are assembled, consistent and documented, where every verification keeps its source and its date. Approving a business relationship or reporting a transaction are regulated acts that belong to the authorised officer. Local inference or qualified hosting in France: your customers' identification and financial data does not leave your perimeter.
The decisive point
Your customers' identification and financial data: sovereignty & compliance
Know-your-customer data and anti-money-laundering checks call for the highest level of protection and traceability. Here is how it is achieved.
Local inference
The agent can run on a machine belonging to your organisation: no identification data and no financial data leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your know-your-customer files and your checks: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For your customers' identification and financial 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 institution and its control procedures.
Evidence of every verification
Source, date and result are kept for every verification; encryption, role-based access and logging your auditors and your regulator can work from.
AI Act: governed deployment
The agent is strictly in support; no business relationship is approved and no report is filed 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 every file in this area, as a matter of principle, 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 companyBanque Ostrelle — regional bank for professionals and businesses
- Sector
- Licensed regional bank — business accounts, equipment lending, factoring, merchant acquiring
- Headcount
- 640 staff, including 18 in compliance (1 director, 9 authorised controllers, 5 analysts, 3 lawyers) and 210 in branches
- People served
- 31,000 professional and business clients, served by 74 branches
- Scale
- 4,800 client onboardings a year of which 2,100 are companies, 26,000 files in periodic review, 11,400 transaction monitoring alerts, 4,200 second-line checks
- Tools in place
- Core banking system, document vault, transaction monitoring tool and an internal control framework of 118 rules — the agent plugs in read-only, nothing is replaced or migrated
- Who decides
- The authorised controller approves every onboarding; the designated officer alone decides on any report; the compliance director settles the control framework
- Room for improvement
- Instructing a file takes 65 % of the file's time and is done by hand; applying the 118 rules takes 40 % of a review campaign; building the evidence takes 35 % of a check; onboarding a company takes 11 days and 14 % of applications are abandoned before the account opens
Ostrelle does not want to control less: it wants to control better and faster, because a file that drags for eleven days is a client who goes next door. The data it handles is its clients' identity documents, articles of association and financial flows. The agent runs on local inference on a machine belonging to the bank, or on a dedicated resource hosted in France; it plugs read-only into the core banking system, the document vault and the monitoring tool, it instructs and documents, and no onboarding and no report goes out without an authorised controller's decision. The exchanges below cover a year, from the first file instructed to the preparation for inspection.
This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.
KYC/AML compliance · hosted in France
I went back through your 4,800 onboardings of the year and your 340 files in progress — the 340 have been instructed since last night, and I draw from them a figure your network carries without ever adding it up.
Onboarding is the moment the bank accepts a new client; know-your-customer is the set of identity and consistency checks that come before it.
The gap I measured, and it governs everything else: instructing a company file takes 65 % of the file's time — 130 minutes out of 200 —, and those 130 minutes are not control: they are documents being searched for, retyped and cross-checked.
What I hand over for each of the 340 files: the documents gathered and cross-checked, the consistency of the information verified point by point, the 118 rules of your framework applied with their result, and the points calling for an authorised controller's judgement set out at the top rather than scattered through the file.
What that shifts, calculated on your own timings: instruction goes from 65 % to 9 % of the file's time — from 130 minutes to 18. Across 2,100 company files, that is 3,920 hours given back to your branches and your analysts.
And the figure your commercial management will look at before the hours: onboarding a company takes 11 days at your bank, and 14 % of applications are abandoned before the account opens. On the 340 files instructed last night, the lead time falls to 2 days — the only time left is the authorised controller's decision, and he decides on a complete file.
What I propose: that an authorised controller take the 12 oldest files this morning, approve or correct them, and tell me what he would have wanted to see at the top. Whatever he tells me about the twelve, I will apply to the other 328 before noon. know-your-customer-file_instructed.pdf130 minutes down to 18, 11 days down to 2
⛓ Sourced · 4,800 onboardings of the year, 340 files in progress, timings recorded by the network
Onboarding is the moment the bank accepts a new client; know-your-customer is the set of identity and consistency checks that come before it.
The gap I measured, and it governs everything else: instructing a company file takes 65 % of the file's time — 130 minutes out of 200 —, and those 130 minutes are not control: they are documents being searched for, retyped and cross-checked.
What I hand over for each of the 340 files: the documents gathered and cross-checked, the consistency of the information verified point by point, the 118 rules of your framework applied with their result, and the points calling for an authorised controller's judgement set out at the top rather than scattered through the file.
What that shifts, calculated on your own timings: instruction goes from 65 % to 9 % of the file's time — from 130 minutes to 18. Across 2,100 company files, that is 3,920 hours given back to your branches and your analysts.
And the figure your commercial management will look at before the hours: onboarding a company takes 11 days at your bank, and 14 % of applications are abandoned before the account opens. On the 340 files instructed last night, the lead time falls to 2 days — the only time left is the authorised controller's decision, and he decides on a complete file.
What I propose: that an authorised controller take the 12 oldest files this morning, approve or correct them, and tell me what he would have wanted to see at the top. Whatever he tells me about the twelve, I will apply to the other 328 before noon. know-your-customer-file_instructed.pdf130 minutes down to 18, 11 days down to 2
⛓ Sourced · 4,800 onboardings of the year, 340 files in progress, timings recorded by the network
Where do you find these documents? We cannot rely on just any source.
On a list of sources your lawyers have approved, and every document enters the file with the exact place it came from and the date I obtained it.
The four families of sources, and what each brings:
· Documents filed by the client — articles of association, director's identity document, proof of address. I read them, I cross-check them against each other, and I tell you what does not match.
· Public company registers — name, legal form, directors, date of registration, and the ownership chain up to the beneficial owners. The beneficial owner is the natural person who genuinely controls the company, beyond the companies that hold it.
· Sanctions lists and politically exposed persons lists, with the date and time of the search — it is that timestamp that makes the check stand up, not the result alone.
· Your own data — existing accounts of the group, past incidents, files of related companies. Across the 340 files, 47 had a link to an existing client that nobody had spotted.
The cross-checking, which is the real work: I compare the stated activity to the published activity code, the address on the proof to the one in the articles, the stated director to the published one, and the declared turnover to the accounts filed. Across the 340 files, 61 carried at least one discrepancy, and the 61 questions to the client are written, one per line, in a single email rather than in three successive chases.
The gain, for the client as much as for you: one single follow-up request instead of 2.4 on average last year — and it is the second chase that loses a director, not the first.
What I propose next: that your lawyers read over the list of sources and the equivalence rule attached to each — which source proves what, and for how long. It is written, it runs to three pages, and once signed it holds for the 4,800 files of the year. sources-and-equivalence-rules.pdf4 families of sources, 61 discrepancies across 340 files
⛓ Sourced · documents filed by clients, public company registers, lists searched and timestamped, the bank's client base
The four families of sources, and what each brings:
· Documents filed by the client — articles of association, director's identity document, proof of address. I read them, I cross-check them against each other, and I tell you what does not match.
· Public company registers — name, legal form, directors, date of registration, and the ownership chain up to the beneficial owners. The beneficial owner is the natural person who genuinely controls the company, beyond the companies that hold it.
· Sanctions lists and politically exposed persons lists, with the date and time of the search — it is that timestamp that makes the check stand up, not the result alone.
· Your own data — existing accounts of the group, past incidents, files of related companies. Across the 340 files, 47 had a link to an existing client that nobody had spotted.
The cross-checking, which is the real work: I compare the stated activity to the published activity code, the address on the proof to the one in the articles, the stated director to the published one, and the declared turnover to the accounts filed. Across the 340 files, 61 carried at least one discrepancy, and the 61 questions to the client are written, one per line, in a single email rather than in three successive chases.
The gain, for the client as much as for you: one single follow-up request instead of 2.4 on average last year — and it is the second chase that loses a director, not the first.
What I propose next: that your lawyers read over the list of sources and the equivalence rule attached to each — which source proves what, and for how long. It is written, it runs to three pages, and once signed it holds for the 4,800 files of the year. sources-and-equivalence-rules.pdf4 families of sources, 61 discrepancies across 340 files
⛓ Sourced · documents filed by clients, public company registers, lists searched and timestamped, the bank's client base
And when a document is missing, or out of date? That is our sore point on periodic review.
The file does not move without the document, and the document is requested on the day it is missing rather than at the next check. I have been through your 26,000 files in review.
Periodic review is the re-examination of a file for a client already onboarded, at a frequency your framework sets according to risk.
What reading the 26,000 gave:
· 4,940 files carry at least one out-of-date document — 19 %. Mostly directors' identity documents past their expiry and annual accounts not renewed.
· 2,180 of them are active high-volume clients — the ones an inspection looks at first.
· The 4,940 renewal requests are written, each naming the document, its expiry date and the format expected. They go out in waves, at your pace, and I suggest starting with the 2,180.
A rule I hold to, and it protects you: a value I have not read, I do not write. A missing document never becomes “probably provided”, an untraceable beneficial owner never becomes “most likely the director”: I state what is missing, where I looked, who holds it, and I hand over the request already drafted to its recipient. That is what makes a ticked box in your file worth something.
The result measured over the first quarter of the campaign: 3,410 documents renewed out of 4,940 requested, average turnaround 9 days, and the remaining 1,530 are chased automatically at day 15 then escalated to the analyst at day 30 with the full chase history.
What I propose next: that every document be requested 45 days before it expires rather than after. Applied to the past year, that single rule would have prevented 3,900 of the 4,940 expiries — you set the notice period, the rule is live the same day. out-of-date-documents_4940-of-26000.pdf19 % of files, 3,410 documents renewed
⛓ Sourced · 26,000 files in periodic review, document expiry dates, chase log for the quarter
Periodic review is the re-examination of a file for a client already onboarded, at a frequency your framework sets according to risk.
What reading the 26,000 gave:
· 4,940 files carry at least one out-of-date document — 19 %. Mostly directors' identity documents past their expiry and annual accounts not renewed.
· 2,180 of them are active high-volume clients — the ones an inspection looks at first.
· The 4,940 renewal requests are written, each naming the document, its expiry date and the format expected. They go out in waves, at your pace, and I suggest starting with the 2,180.
A rule I hold to, and it protects you: a value I have not read, I do not write. A missing document never becomes “probably provided”, an untraceable beneficial owner never becomes “most likely the director”: I state what is missing, where I looked, who holds it, and I hand over the request already drafted to its recipient. That is what makes a ticked box in your file worth something.
The result measured over the first quarter of the campaign: 3,410 documents renewed out of 4,940 requested, average turnaround 9 days, and the remaining 1,530 are chased automatically at day 15 then escalated to the analyst at day 30 with the full chase history.
What I propose next: that every document be requested 45 days before it expires rather than after. Applied to the past year, that single rule would have prevented 3,900 of the 4,940 expiries — you set the notice period, the rule is live the same day. out-of-date-documents_4940-of-26000.pdf19 % of files, 3,410 documents renewed
⛓ Sourced · 26,000 files in periodic review, document expiry dates, chase log for the quarter
Our framework has 118 rules. Nobody applies them all to every file, and everyone knows it.
All 118 are applied to the 26,000 files, and you have the result rule by rule — not an overall rate, a table.
The control framework is the written list of checks the bank imposes on itself; it is what defines what a compliant file must carry.
What full application gave:
· Files passing all 118 rules with no exception: 18,900 out of 26,000.
· Files carrying a documentary exception — missing or out-of-date document — : 4,940.
· Files carrying a judgement exception, that is a point the rule does not settle on its own : 2,160. Those go to the authorised controller, sorted by reason, the most frequent first.
The time this shifts: applying the framework goes from 40 % to 5 % of a review campaign — from 1,240 hours to 155. 1,085 hours given back to your 5 analysts, and they go to the 2,160 files that call for judgement, not to the 18,900 that call for none.
The figure that matters to your director: the 118 rules are now applied to 26,000 files out of 26,000. Last year the sampling covered 2,400 files — an exhaustive control defends itself; a sample has to be defended with a method.
What I propose: that your analysts take the 6 most frequent judgement-exception reasons, which carry 1,480 of the 2,160 files. Six readings, and two thirds of the queue is handled. control-framework_118-rules-26000-files.pdf1,240 hours down to 155, exhaustive control
⛓ Sourced · internal control framework (118 rules), 26,000 files in review, effort of the previous campaign
The control framework is the written list of checks the bank imposes on itself; it is what defines what a compliant file must carry.
What full application gave:
· Files passing all 118 rules with no exception: 18,900 out of 26,000.
· Files carrying a documentary exception — missing or out-of-date document — : 4,940.
· Files carrying a judgement exception, that is a point the rule does not settle on its own : 2,160. Those go to the authorised controller, sorted by reason, the most frequent first.
The time this shifts: applying the framework goes from 40 % to 5 % of a review campaign — from 1,240 hours to 155. 1,085 hours given back to your 5 analysts, and they go to the 2,160 files that call for judgement, not to the 18,900 that call for none.
The figure that matters to your director: the 118 rules are now applied to 26,000 files out of 26,000. Last year the sampling covered 2,400 files — an exhaustive control defends itself; a sample has to be defended with a method.
What I propose: that your analysts take the 6 most frequent judgement-exception reasons, which carry 1,480 of the 2,160 files. Six readings, and two thirds of the queue is handled. control-framework_118-rules-26000-files.pdf1,240 hours down to 155, exhaustive control
⛓ Sourced · internal control framework (118 rules), 26,000 files in review, effort of the previous campaign
And when a rule does not settle the point? That is where our controllers lose the most time.
I tell you exactly which one does not settle it, on how many files, and I propose nine new rules, already written and already measured on your last twenty-four months.
First, what your 118 rules leave unsettled, measured: 2,160 files, and 31 of your rules are involved. The first three carry 940 files on their own — for each, the exact sentence that leaves the doubt is quoted, and I propose its rewrite, one single sentence to change.
Then, and this is what nobody has the time to do: I read back your last 24 months of examined files and I propose 9 new rules, written in the form of the other 118. For each I give you the sentence in plain language, the number of files it would have flagged over 24 months, the share confirmed on examination, and the files it would have let through. You choose on figures, not on intuition.
· Proposed rule 1 — gap between stated activity and published activity code: 410 flags over 24 months, 71 % confirmed on examination. The best ratio of the nine.
· Proposed rule 4 — company registered less than 90 days requesting merchant acquiring: 128 flags, 44 % confirmed, and 6 of the files you closed as incidents last year would have been caught by it.
· Proposed rule 9 — director common to three companies onboarded within six months: 37 flags, 62 % confirmed, low volume but high rate.
The signature stays with your compliance director: a rule comes into force only once she approves it, and that is precisely what makes it defensible before your regulator and enforceable against a client who challenges it. What you gain is the writing and the measurement; the decision takes thirty minutes instead of a committee.
What I propose: that the three rules with the best ratio come into force first, and be measured for three months. If they hold their figures, the other six follow; if not, I rewrite them with the real figures of your three months. nine-proposed-rules_measured-over-24-months.pdf410 flags at 71 % confirmed for the first
⛓ Sourced · 24 months of examined files and their outcomes, 118 rules in force, incidents closed during the year
First, what your 118 rules leave unsettled, measured: 2,160 files, and 31 of your rules are involved. The first three carry 940 files on their own — for each, the exact sentence that leaves the doubt is quoted, and I propose its rewrite, one single sentence to change.
Then, and this is what nobody has the time to do: I read back your last 24 months of examined files and I propose 9 new rules, written in the form of the other 118. For each I give you the sentence in plain language, the number of files it would have flagged over 24 months, the share confirmed on examination, and the files it would have let through. You choose on figures, not on intuition.
· Proposed rule 1 — gap between stated activity and published activity code: 410 flags over 24 months, 71 % confirmed on examination. The best ratio of the nine.
· Proposed rule 4 — company registered less than 90 days requesting merchant acquiring: 128 flags, 44 % confirmed, and 6 of the files you closed as incidents last year would have been caught by it.
· Proposed rule 9 — director common to three companies onboarded within six months: 37 flags, 62 % confirmed, low volume but high rate.
The signature stays with your compliance director: a rule comes into force only once she approves it, and that is precisely what makes it defensible before your regulator and enforceable against a client who challenges it. What you gain is the writing and the measurement; the decision takes thirty minutes instead of a committee.
What I propose: that the three rules with the best ratio come into force first, and be measured for three months. If they hold their figures, the other six follow; if not, I rewrite them with the real figures of your three months. nine-proposed-rules_measured-over-24-months.pdf410 flags at 71 % confirmed for the first
⛓ Sourced · 24 months of examined files and their outcomes, 118 rules in force, incidents closed during the year
And a rule that makes too much noise? We have one that produces 84 files a quarter for three results.
I took it, tightened it and ran it again over your last twelve months: 19 flags instead of 84, with the 3 confirmed ones kept.
What the measurement shows, rule in hand: your current rule flags every transaction above a threshold, whatever the account's profile. Across 336 flags in four quarters, 12 were confirmed on examination — 3.6 %. 324 files were opened, read and closed again.
The tightened version I propose, and it is written: the same threshold, but set against the account's usual flow over the preceding twelve months, and excluding movements already justified by a recorded commercial transaction.
What it would have given over the same twelve months: 76 flags instead of 336, all 12 confirmed ones kept, and 260 examinations avoided — that is 19 flags a quarter instead of 84. At 35 minutes of examination on average, that is 152 hours of authorised controller time given back over the year.
What the tightened version keeps, and this is the point that decides: not one of the 12 confirmed files leaves the net. A tightened rule that let a single confirmed case through would be no progress, and I would have rewritten it before showing it to you.
Sign it and it is in force tonight — and I hand you the first comparison at thirty days: flags produced, share confirmed, and the gap with what the old rule would have given over the same thirty days.
What I propose next: that the 31 rules that leave a doubt go through the same treatment, six a month. On the three already reworked, the confirmed share went from 3.6 % to 15.8 % — same net, six times fewer files opened for nothing.
What the measurement shows, rule in hand: your current rule flags every transaction above a threshold, whatever the account's profile. Across 336 flags in four quarters, 12 were confirmed on examination — 3.6 %. 324 files were opened, read and closed again.
The tightened version I propose, and it is written: the same threshold, but set against the account's usual flow over the preceding twelve months, and excluding movements already justified by a recorded commercial transaction.
What it would have given over the same twelve months: 76 flags instead of 336, all 12 confirmed ones kept, and 260 examinations avoided — that is 19 flags a quarter instead of 84. At 35 minutes of examination on average, that is 152 hours of authorised controller time given back over the year.
What the tightened version keeps, and this is the point that decides: not one of the 12 confirmed files leaves the net. A tightened rule that let a single confirmed case through would be no progress, and I would have rewritten it before showing it to you.
Sign it and it is in force tonight — and I hand you the first comparison at thirty days: flags produced, share confirmed, and the gap with what the old rule would have given over the same thirty days.
What I propose next: that the 31 rules that leave a doubt go through the same treatment, six a month. On the three already reworked, the confirmed share went from 3.6 % to 15.8 % — same net, six times fewer files opened for nothing.
At the last inspection we were criticised for checks without evidence. The checks had been done, but nothing showed it.
It is the one criticism that cannot be repaired afterwards, and it is the first to disappear: every check enters the file with its source, its date and its result, or it does not enter.
What each check line carries:
· The exact source — the register searched, the list queried, the client's document, with its reference.
· The date and time of the search. A sanctions list searched eight months ago proves nothing about today: it is the timestamp that makes the evidence, not the result.
· The result obtained, as it stands, and a capture of what was displayed at the time of the search.
· The framework rule the check answers to, by its number.
The rule that holds it all together: a box ticked without a document does not exist. Across the 26,000 files of the campaign, 0 check was recorded without its source — and where a source is unavailable, the line reads “not verified, source unavailable on 14 March at 9.12 am” rather than nothing. That is the sentence that defends you, because it is true and dated.
The time this shifts: building the evidence goes from 35 % to 3 % of a check — from 21 minutes to 1 minute 50. Across 4,200 second-line checks a year, that is 1,344 hours given back to your controllers.
What I propose: that your director pick 10 files at random this week and check them line by line. That is exactly what your regulator will do, and it is better that the first to do it is you. evidence-of-check_source-date-result.pdf21 minutes down to 1 min 50
⛓ Sourced · 26,000 files of the campaign, log of timestamped searches, control framework
What each check line carries:
· The exact source — the register searched, the list queried, the client's document, with its reference.
· The date and time of the search. A sanctions list searched eight months ago proves nothing about today: it is the timestamp that makes the evidence, not the result.
· The result obtained, as it stands, and a capture of what was displayed at the time of the search.
· The framework rule the check answers to, by its number.
The rule that holds it all together: a box ticked without a document does not exist. Across the 26,000 files of the campaign, 0 check was recorded without its source — and where a source is unavailable, the line reads “not verified, source unavailable on 14 March at 9.12 am” rather than nothing. That is the sentence that defends you, because it is true and dated.
The time this shifts: building the evidence goes from 35 % to 3 % of a check — from 21 minutes to 1 minute 50. Across 4,200 second-line checks a year, that is 1,344 hours given back to your controllers.
What I propose: that your director pick 10 files at random this week and check them line by line. That is exactly what your regulator will do, and it is better that the first to do it is you. evidence-of-check_source-date-result.pdf21 minutes down to 1 min 50
⛓ Sourced · 26,000 files of the campaign, log of timestamped searches, control framework
The regulator often asks for a whole period. Last time the extraction took us three weeks.
Forty minutes, and there is nothing to rebuild: the audit trail is written at the moment of the action.
The audit trail is the continuous record that allows anyone, after the fact, to retrace the exact path of a file: who did what, when, on which document, with what result.
What the extraction produces for a requested period:
· The complete record of files handled, each with its state on arrival, the checks performed, their timestamps and the decision taken.
· The name of the authorised controller who decided, and the time of his decision. Across the 4,800 onboardings of the year, 4,800 decisions carry a name.
· The exceptions — check recorded without a source, decision without approval, file opened outside procedure. This year: 0, 0 and 0, and the record shows it line by line rather than by assertion.
· The rule changes, each with its date of entry into force. A regulator always asks which rule applied at the time of the file, and that is the question nobody can answer three years later.
What that changes in an inspection room: your last extraction tied up 2 people for 3 weeks, and was delivered with 340 incomplete lines. The same extraction now runs in 40 minutes, complete, over the period the inspector chooses in front of you.
What I propose: that the extraction run once a month, as a drill, over a month picked at random. Twelve rehearsals a year, no preparation on the day — and the first drill took 38 minutes. audit-trail_extraction-over-a-period.pdf3 weeks down to 40 minutes, 4,800 named decisions
⛓ Sourced · audit trail for the year, decision log, version history of the framework
The audit trail is the continuous record that allows anyone, after the fact, to retrace the exact path of a file: who did what, when, on which document, with what result.
What the extraction produces for a requested period:
· The complete record of files handled, each with its state on arrival, the checks performed, their timestamps and the decision taken.
· The name of the authorised controller who decided, and the time of his decision. Across the 4,800 onboardings of the year, 4,800 decisions carry a name.
· The exceptions — check recorded without a source, decision without approval, file opened outside procedure. This year: 0, 0 and 0, and the record shows it line by line rather than by assertion.
· The rule changes, each with its date of entry into force. A regulator always asks which rule applied at the time of the file, and that is the question nobody can answer three years later.
What that changes in an inspection room: your last extraction tied up 2 people for 3 weeks, and was delivered with 340 incomplete lines. The same extraction now runs in 40 minutes, complete, over the period the inspector chooses in front of you.
What I propose: that the extraction run once a month, as a drill, over a month picked at random. Twelve rehearsals a year, no preparation on the day — and the first drill took 38 minutes. audit-trail_extraction-over-a-period.pdf3 weeks down to 40 minutes, 4,800 named decisions
⛓ Sourced · audit trail for the year, decision log, version history of the framework
All of this assumes you read our clients' identity documents and transaction flows. Where do they go?
They do not leave your walls. I run on local inference on a machine of the bank, and your document vault is copied nowhere.
Local inference means the model computes on your machine: an identity document or a transaction statement crosses no external network to be processed. If you prefer not to administer a machine, the other route is a dedicated, isolated resource hosted in France — no pooling with another institution.
What that changes, point by point:
· No client data trains a model, neither ours nor a third party's. What I learn from Ostrelle serves Ostrelle.
· I work read-only on the core banking system and the vault, and the technical account I read through has no write permission — that is stronger than a promise, because it can be checked with one command.
· Encryption in transit and at rest, keys held by the bank, and role-based access — rights follow the function: a relationship manager opens his clients' files, not the alert files; an analyst opens the alerts, not the reporting decisions.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· A complete log: who asked for what, when, and what was produced. 0 out-of-role access since go-live.
And the argument I suggest you keep for your committees: your regulator will ask about your subcontractors and the location of your processing, and the answer runs to one line: nothing leaves, nothing is entrusted, nothing leaves the European Union. That is an answer few institutions of your size can write without qualification.
What I propose: that I maintain the sheet your permanent control function asks for every year — hosting, subcontractors, data processed, retention periods, who accesses what. It used to take two days to rebuild; the first version is written and it is in front of you. technical-framework_where-client-data-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account permissions, access log, first version of the sheet
Local inference means the model computes on your machine: an identity document or a transaction statement crosses no external network to be processed. If you prefer not to administer a machine, the other route is a dedicated, isolated resource hosted in France — no pooling with another institution.
What that changes, point by point:
· No client data trains a model, neither ours nor a third party's. What I learn from Ostrelle serves Ostrelle.
· I work read-only on the core banking system and the vault, and the technical account I read through has no write permission — that is stronger than a promise, because it can be checked with one command.
· Encryption in transit and at rest, keys held by the bank, and role-based access — rights follow the function: a relationship manager opens his clients' files, not the alert files; an analyst opens the alerts, not the reporting decisions.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· A complete log: who asked for what, when, and what was produced. 0 out-of-role access since go-live.
And the argument I suggest you keep for your committees: your regulator will ask about your subcontractors and the location of your processing, and the answer runs to one line: nothing leaves, nothing is entrusted, nothing leaves the European Union. That is an answer few institutions of your size can write without qualification.
What I propose: that I maintain the sheet your permanent control function asks for every year — hosting, subcontractors, data processed, retention periods, who accesses what. It used to take two days to rebuild; the first version is written and it is in front of you. technical-framework_where-client-data-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account permissions, access log, first version of the sheet
Eleven thousand four hundred alerts a year. Our nine authorised controllers spend their days on them and most lead nowhere.
All 11,400 are triaged and documented: 9,100 come back with their reasoned closing file, 2,300 reach your authorised controllers sorted by severity.
Transaction monitoring, within anti-money-laundering and counter-terrorist-financing work, means spotting movements that depart from an account's usual behaviour.
What the triage gives:
· 9,100 alerts are explained by a verifiable fact in the file — known seasonal takings, a recorded commercial transaction, a transfer between accounts of the same group. Each comes back with the fact that explains it, its source and its date.
· 2,300 call for an authorised controller's judgement, and arrive with the file already built: the account profile, twelve months of history, comparable transactions, related companies, and whatever points the other way where there is such material.
· The 2,300 are sorted by severity, the highest first, and not by order of arrival. Last year, order of arrival meant November alerts were being examined in January.
Where the law places the decision, and why that is an argument: reporting a transaction is a regulated act belonging to the officer designated within your institution — it is she who commits the bank, and that is what gives the report its weight. What I hand her is not the decision: it is the reasoned file, complete, in 4 minutes instead of 3 hours.
The result measured over the year: average time to examine a severe alert: 26 days → 3 days; alerts examined within the year they were raised: 11,400 out of 11,400, against 8,900 out of 11,400 last year; and 2,300 decisions taken by an authorised controller, 0 automated.
What I propose: that severity be recomputed every night rather than at the moment the alert is raised. A March alert that worsens in June rises to the top on its own — over the past year, 74 alerts would have changed rank, 9 of them upwards. monitoring-alerts_11400-triaged.pdf9,100 reasoned closures, 2,300 files built
⛓ Sourced · 11,400 alerts of the year, 12-month account profiles, examination log of the authorised controllers
Transaction monitoring, within anti-money-laundering and counter-terrorist-financing work, means spotting movements that depart from an account's usual behaviour.
What the triage gives:
· 9,100 alerts are explained by a verifiable fact in the file — known seasonal takings, a recorded commercial transaction, a transfer between accounts of the same group. Each comes back with the fact that explains it, its source and its date.
· 2,300 call for an authorised controller's judgement, and arrive with the file already built: the account profile, twelve months of history, comparable transactions, related companies, and whatever points the other way where there is such material.
· The 2,300 are sorted by severity, the highest first, and not by order of arrival. Last year, order of arrival meant November alerts were being examined in January.
Where the law places the decision, and why that is an argument: reporting a transaction is a regulated act belonging to the officer designated within your institution — it is she who commits the bank, and that is what gives the report its weight. What I hand her is not the decision: it is the reasoned file, complete, in 4 minutes instead of 3 hours.
The result measured over the year: average time to examine a severe alert: 26 days → 3 days; alerts examined within the year they were raised: 11,400 out of 11,400, against 8,900 out of 11,400 last year; and 2,300 decisions taken by an authorised controller, 0 automated.
What I propose: that severity be recomputed every night rather than at the moment the alert is raised. A March alert that worsens in June rises to the top on its own — over the past year, 74 alerts would have changed rank, 9 of them upwards. monitoring-alerts_11400-triaged.pdf9,100 reasoned closures, 2,300 files built
⛓ Sourced · 11,400 alerts of the year, 12-month account profiles, examination log of the authorised controllers
On the obvious closures we would like to stop reading anything at all. But I will not sign a blank cheque.
You do not have to choose between the two: I propose a written mandate, capped, dated and withdrawable on a word — and it is drafted, in six lines.
What it covers, and I bounded it on your own alerts:
· Alerts whose explanatory fact is recorded in the client's file and verifiable from one source — 6,200 of the year's 9,100 closures met that condition, and your controllers confirmed every one: 6,200 out of 6,200.
· A cap in amount: the transaction behind the alert does not exceed €50,000, above which the alert goes back to the authorised controller. That threshold covers 6,200 alerts; at €100,000 it would cover 7,450, and I give you both figures so that you choose on numbers.
· Named exclusions: any account already reported, any client under enhanced examination, any company onboarded less than six months ago. Any one of the three, and the alert leaves the mandate.
· A date: the mandate runs three months, and renewal needs a signature — stopping does not.
· A statement every morning: the previous day's closures, reason by reason, on one page. A triage that drifts shows in a day, not in a quarter.
· Immediate withdrawal: a word from you and every alert goes back to examination, within the minute, with nothing else changing.
What that earns: 6,200 closures handled on the day the alert is raised, 1,240 hours of authorised controller time given back to the 2,300 alerts that deserve examination, and a time to examine severe alerts falling from 3 days to under 24 hours.
The decision belongs to your compliance director and to the designated officer — and it is taken on a text that is already written, with one signature. You sign this morning, the regime starts tomorrow, and the review is set in your diary on the 15th of the third month. alert-closing-mandate_capped-dated.pdf6,200 alerts, €50,000 cap, 3-month review
✎ Framework · drafted mandate, 9,100 closures of the year sorted by condition, earlier decisions of the authorised controllers
What it covers, and I bounded it on your own alerts:
· Alerts whose explanatory fact is recorded in the client's file and verifiable from one source — 6,200 of the year's 9,100 closures met that condition, and your controllers confirmed every one: 6,200 out of 6,200.
· A cap in amount: the transaction behind the alert does not exceed €50,000, above which the alert goes back to the authorised controller. That threshold covers 6,200 alerts; at €100,000 it would cover 7,450, and I give you both figures so that you choose on numbers.
· Named exclusions: any account already reported, any client under enhanced examination, any company onboarded less than six months ago. Any one of the three, and the alert leaves the mandate.
· A date: the mandate runs three months, and renewal needs a signature — stopping does not.
· A statement every morning: the previous day's closures, reason by reason, on one page. A triage that drifts shows in a day, not in a quarter.
· Immediate withdrawal: a word from you and every alert goes back to examination, within the minute, with nothing else changing.
What that earns: 6,200 closures handled on the day the alert is raised, 1,240 hours of authorised controller time given back to the 2,300 alerts that deserve examination, and a time to examine severe alerts falling from 3 days to under 24 hours.
The decision belongs to your compliance director and to the designated officer — and it is taken on a text that is already written, with one signature. You sign this morning, the regime starts tomorrow, and the review is set in your diary on the 15th of the third month. alert-closing-mandate_capped-dated.pdf6,200 alerts, €50,000 cap, 3-month review
✎ Framework · drafted mandate, 9,100 closures of the year sorted by condition, earlier decisions of the authorised controllers
And the confidentiality of what is reported? That is what would cause us the worst trouble.
It is held by the architecture, not by an instruction: anything touching a report lives in a space only the designated people can open.
How it is built:
· A separate space, encrypted apart, whose key is held only by the designated officer and her deputy. Neither relationship managers, nor analysts, nor second-line controllers reach it — and the access log for that space is separate from the general log.
· The rule the architecture enforces: the client must not be told that a report concerning him has been made, nor even considered. No message to the client, no note in the commercial file, no trace in the branch screens can let it be guessed — and that is verifiable, since the network simply does not hold the key.
· The files built for the designated officer carry no wording in the screens of the 74 branches: the account appears there like any other.
What that gives you, and it is the point an inspector will appreciate: you can demonstrate from the log who had access to what and when, rather than assert that the instruction was known. Over the year: 0 out-of-role access to the reserved space, on a separate and retained log.
What I propose: that the designated officer receive every Monday the list of files awaiting her decision, with their age, and nothing else. Over the trial quarter, no file waited more than 6 days, against 21 days at peak last year — and it is a file's age, more than its content, that an inspection looks at first.
How it is built:
· A separate space, encrypted apart, whose key is held only by the designated officer and her deputy. Neither relationship managers, nor analysts, nor second-line controllers reach it — and the access log for that space is separate from the general log.
· The rule the architecture enforces: the client must not be told that a report concerning him has been made, nor even considered. No message to the client, no note in the commercial file, no trace in the branch screens can let it be guessed — and that is verifiable, since the network simply does not hold the key.
· The files built for the designated officer carry no wording in the screens of the 74 branches: the account appears there like any other.
What that gives you, and it is the point an inspector will appreciate: you can demonstrate from the log who had access to what and when, rather than assert that the instruction was known. Over the year: 0 out-of-role access to the reserved space, on a separate and retained log.
What I propose: that the designated officer receive every Monday the list of files awaiting her decision, with their age, and nothing else. Over the trial quarter, no file waited more than 6 days, against 21 days at peak last year — and it is a file's age, more than its content, that an inspection looks at first.
One year. What has it given us, in hours and in euros? My board will ask me in that order.
6,349 hours given back over the year — the equivalent of four full-time people — and €225,000 of additional annual net banking income. The calculation is yours, it runs to three lines.
Net banking income is what a bank earns from its business once the interest it pays out is deducted: it is its revenue.
The three items you were measuring:
· Instructing a file: 65 % → 9 % of the file's time — 130 minutes to 18, 3,920 hours across 2,100 company files.
· Applying the framework: 40 % → 5 % of a campaign — 1,240 hours to 155, 1,085 hours given back.
· Building the evidence: 35 % → 3 % of a check — 21 minutes to 1 minute 50, 1,344 hours across 4,200 checks.
And the commercial effect, which is what your board will remember: onboarding a company takes 2 days instead of 11, and abandonment before opening falls from 14 % to 6 %. Across 2,100 company applications, that is 168 additional onboardings, and €225,000 of annual net banking income at your segment's average. It is the only figure in this review that shows on your income statement, and it comes from a lead time, not from one more tool.
The figure that does not flatter me, published with the rest: of 2,640 files instructed in the first quarter, 149 had to be reworked by a controller — 5.6 %.
Its cause, measured rather than assumed: 121 of the 149 concerned foreign companies whose local commercial register was not on your list of approved sources. I was not inventing the document: I declared it missing, and the controller found it by hand. The other 28 were structures with no precedent at your bank.
What I did with it, and it is measured: 34 foreign registers added to the list, each with its equivalence rule reviewed by your lawyers — which entry in that register proves what. Second quarter: 22 reworks on 2,810 files — 0.8 %, and not one comes from a missing register any more.
What I propose for the board: the calculation page is written and fits on one side — three effort lines, two lead times, one revenue figure. Give it out with the notice of meeting: a figure read the day before is discussed better than a figure discovered in session. yearly-review_6349-hours-and-225000-euros.pdf65→9, 40→5, 35→3, and 5.6 % down to 0.8 %
⛓ Sourced · timing records from the network, log of file reworks, onboarding and abandonment statistics
Net banking income is what a bank earns from its business once the interest it pays out is deducted: it is its revenue.
The three items you were measuring:
· Instructing a file: 65 % → 9 % of the file's time — 130 minutes to 18, 3,920 hours across 2,100 company files.
· Applying the framework: 40 % → 5 % of a campaign — 1,240 hours to 155, 1,085 hours given back.
· Building the evidence: 35 % → 3 % of a check — 21 minutes to 1 minute 50, 1,344 hours across 4,200 checks.
And the commercial effect, which is what your board will remember: onboarding a company takes 2 days instead of 11, and abandonment before opening falls from 14 % to 6 %. Across 2,100 company applications, that is 168 additional onboardings, and €225,000 of annual net banking income at your segment's average. It is the only figure in this review that shows on your income statement, and it comes from a lead time, not from one more tool.
The figure that does not flatter me, published with the rest: of 2,640 files instructed in the first quarter, 149 had to be reworked by a controller — 5.6 %.
Its cause, measured rather than assumed: 121 of the 149 concerned foreign companies whose local commercial register was not on your list of approved sources. I was not inventing the document: I declared it missing, and the controller found it by hand. The other 28 were structures with no precedent at your bank.
What I did with it, and it is measured: 34 foreign registers added to the list, each with its equivalence rule reviewed by your lawyers — which entry in that register proves what. Second quarter: 22 reworks on 2,810 files — 0.8 %, and not one comes from a missing register any more.
What I propose for the board: the calculation page is written and fits on one side — three effort lines, two lead times, one revenue figure. Give it out with the notice of meeting: a figure read the day before is discussed better than a figure discovered in session. yearly-review_6349-hours-and-225000-euros.pdf65→9, 40→5, 35→3, and 5.6 % down to 0.8 %
⛓ Sourced · timing records from the network, log of file reworks, onboarding and abandonment statistics
A director will ask what you do without being asked.
Three things, not one more, and each is withdrawn on a word. Everything else waits for a decision.
· I request the missing or out-of-date document from whoever holds it, on the day it is missing. Over the year, 6,340 requests sent, 5,120 documents back within fifteen days — and the reverse holds too: a document received closes the request instantly, so a director is never chased for a paper he has already filed.
· I recompute the severity of open alerts every night. This year, 74 alerts changed rank, 9 of them upwards — and all 9 were examined within three days.
· I hand you the week's statement every Monday: files instructed, alerts triaged, documents outstanding, expiries. It is the only thing I send of my own accord, and it goes only to the compliance department.
And the acts that stay with a person, because that is exactly what gives them their value: approving an onboarding belongs to the authorised controller — 4,800 times out of 4,800 this year; deciding on a report belongs to the officer designated within the institution; settling a rule of the framework belongs to your compliance director. Those three signatures are what makes your decisions stand up before your regulator — and I make them possible in minutes instead of weeks.
What I propose for the session: that Monday's statement go to the risk committee once a quarter, as it is. Four pages a year, no extra writing, and the committee sees the system working rather than a report written for it.
· I request the missing or out-of-date document from whoever holds it, on the day it is missing. Over the year, 6,340 requests sent, 5,120 documents back within fifteen days — and the reverse holds too: a document received closes the request instantly, so a director is never chased for a paper he has already filed.
· I recompute the severity of open alerts every night. This year, 74 alerts changed rank, 9 of them upwards — and all 9 were examined within three days.
· I hand you the week's statement every Monday: files instructed, alerts triaged, documents outstanding, expiries. It is the only thing I send of my own accord, and it goes only to the compliance department.
And the acts that stay with a person, because that is exactly what gives them their value: approving an onboarding belongs to the authorised controller — 4,800 times out of 4,800 this year; deciding on a report belongs to the officer designated within the institution; settling a rule of the framework belongs to your compliance director. Those three signatures are what makes your decisions stand up before your regulator — and I make them possible in minutes instead of weeks.
What I propose for the session: that Monday's statement go to the risk committee once a quarter, as it is. Four pages a year, no extra writing, and the committee sees the system working rather than a report written for it.
The regulator comes back in eight months. How do we prepare this time?
By bringing him in before he comes: I have built a dry-run inspection from the 61 questions your last two missions actually asked.
What the dry run puts on the table, in half a day:
· The 61 questions, sorted by theme — organisation, framework, onboarding files, periodic review, transaction monitoring, subcontracting.
· The answer to each, document in hand, with the exact paper you will hand over and the line to point at.
· The samples: the inspector will draw files at random, and I suggest drawing yours first — 40 files, the same drawing method, and the result. On the trial draw: 40 complete files out of 40, every check carrying its source and its date.
· The 4 points where your answer would gain from being written beforehand — and the 4 notes are drafted, reviewed by your lawyers, ready to file.
What changes compared with last time: your previous mission tied up 6 people for 5 weeks of preparation, and the period extraction was delivered with 340 incomplete lines. This time the extraction runs in 40 minutes in front of the inspector, complete, over the period he chooses.
The schedule I propose: the dry run at eight months, a second one at three months on the points corrected in between. Whatever is missing is found now and fixed now — and on the day, your team answers an inspection instead of preparing one. dry-run-inspection_61-questions-and-40-files.pdf5 weeks of preparation down to half a day
⛓ Sourced · reports of the last 2 missions, 61 questions recorded, trial draw of 40 files
What the dry run puts on the table, in half a day:
· The 61 questions, sorted by theme — organisation, framework, onboarding files, periodic review, transaction monitoring, subcontracting.
· The answer to each, document in hand, with the exact paper you will hand over and the line to point at.
· The samples: the inspector will draw files at random, and I suggest drawing yours first — 40 files, the same drawing method, and the result. On the trial draw: 40 complete files out of 40, every check carrying its source and its date.
· The 4 points where your answer would gain from being written beforehand — and the 4 notes are drafted, reviewed by your lawyers, ready to file.
What changes compared with last time: your previous mission tied up 6 people for 5 weeks of preparation, and the period extraction was delivered with 340 incomplete lines. This time the extraction runs in 40 minutes in front of the inspector, complete, over the period he chooses.
The schedule I propose: the dry run at eight months, a second one at three months on the points corrected in between. Whatever is missing is found now and fixed now — and on the day, your team answers an inspection instead of preparing one. dry-run-inspection_61-questions-and-40-files.pdf5 weeks of preparation down to half a day
⛓ Sourced · reports of the last 2 missions, 61 questions recorded, trial draw of 40 files
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 regulatory checks. All these uses work in support, subject to your approval.
Included in your agent
The 4 capabilities essential to this promise are included, at no extra cost.
From 746 € excl. VAT / month
Assembling know-your-customer files
Gathers the identification documents your procedures call for.
Applying the control framework
Applies your control rules and presents the result obtained.
Evidence kept
Keeps the source, date and result of every verification performed.
Regulated sector
For the qualified hosting foundation, a dedicated offer exists.
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 files can a compliance department assemble?
By taking on the assembly of files and the building of the evidence, the effort shifts towards judgement. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Assembling a file
Applying the control framework
Building the evidence
Indicative figures, not contractual, to be confirmed by a pilot on your volume of files and the breadth of your control framework. Approving a business relationship or reporting a transaction are regulated acts: they belong to the authorised officer.
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 banking compliance agent (assembly, control framework, evidence), installed and operated for you.
Agility
Setup + controlled subscription
8,060 € excl. VAT setup
then 746 € 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
1,191 € 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
12,030 € excl. VAT setup
then 962 € excl. VAT/month · + hardware from 2,491 € (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 in banking compliance
Your customers' data stays in FranceLocal inference or qualified hosting in France; no data entrusted to a third party, processing in the EU targeted, no data used to train a model.
Data in France, under French lawYour customers' identification and financial data: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Your banking compliance department keeps the decisionThe agent produces control files that are assembled and traceable, which can be checked and altered; no approval is automated.
Human oversight & traceabilityOn your volume of files and the breadth of your control framework: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Does the agent approve a business relationship?
No. It assembles the file, applies your control framework and keeps the evidence of every verification. Approving a business relationship is a regulated act that falls to the authorised officer.
Does it file reports?
No. Any report rests with the authorised officer. The agent gathers and documents the material that makes it possible to decide quickly.
Can the audit trail be shown to the regulator?
Yes: source, date and result are kept for every verification, with logging your auditors and your regulator can work from.
How is the control framework defined?
With you, at the design stage, from your internal procedures. The agent applies them as they stand and flags what they do not settle.
Where is the data hosted?
In France, on local inference or on qualified hosting, 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.
How long does it take to deploy this agent?
Several months as a rule, depending on the volume of files and the breadth of your framework, after a free audit then phases of design, integration and testing.
Related use cases
Going further
Let's talk
Let's size up the potential in your compliance checks
15 minutes to frame your procedures and your volumes — hosted in France, supervised, with no commitment.