Advertising campaigns: your budget trade-offs documented
Deciding how to allocate a media budget means comparing actual performance, audience by audience and format by format. Your agent gathers that data from your platforms, presents the differences in performance and proposes several costed trade-offs, each with what it entails. Hosted in France: your budgets and your performance stay with you. Marketing management decides on any change of budget.
Updated on
Three notable differences against last week, with the volumes and costs that explain them.
The comparison scope is stated for each difference.
🔗 Sourced · data from the advertising platforms
Changing a media budget commits spending: the decision belongs to marketing management.
✎ Support · trade-offs proposed, budget decision
A Blue Lemon Agent advertising campaign agent gathers your performance from your platforms, presents the differences by audience and by format with the volumes and costs that explain them, and proposes several costed trade-offs. No change of budget is applied automatically. It runs on local inference or is hosted in France: your budgets stay with you, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
These figures describe our offer, not results measured at a client. How large the gain is on your number of platforms and campaigns tracked is confirmed by a pilot.
What does an AI agent bring to your advertising campaigns?
A budget trade-off taken on consolidated, recent data beats one taken on a hunch.
! The issue
Allocating a media budget means comparing several platforms, several audiences and several formats over the same period. Consolidating that data by hand takes up a share of the time that should go to the decision. The agent gathers it, presents the differences with their comparison scope and costs several possible trade-offs.
✓ Our answer
Marketing management compares complete trade-offs — budget moved, expected effect, what is left aside — and decides. Changing a media budget commits real spending: that decision stays human. Local inference or an isolated resource hosted in France: your budgets and your performance by audience, competitively sensitive information, do not leave the company.
Your budgets and your performance by audience: sovereignty & compliance
Your media budgets and your performance by audience are competitively sensitive information. Here is how they are protected.
Local inference
The agent can run on a machine belonging to your organisation: no budget and no performance data leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your advertising campaigns and your media budgets: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For your budgets and your performance by audience, 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 advertising platforms.
Comparison scope shown
Every difference states the period and the scope compared; encryption, role-based access and logging of the trade-offs proposed.
AI Act: governed deployment
The agent is strictly in support; no budget is changed and no campaign is paused 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.
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 companyMaison Verlaine — French home linen brand, sold online and in 3 shops
- Sector
- Retail — home linen woven in France, sold online and in 3 shops
- Headcount
- 62 staff, including 5 in marketing: 1 marketing director, 2 in acquisition, 1 in CRM, 1 in content
- Market
- Consumers in France and Belgium — average basket of €128, 1,150 orders attributed to media every month
- Volume
- €47,000 of media budget a month across 4 platforms, 38 live campaigns, 118 audiences, 214 ads running
- Tools in place
- The 4 ad platform accounts, the e-commerce platform, the measurement tool and the monthly trade-off spreadsheet — the agent plugs in read-only, nothing is replaced
- Who decides
- The marketing director settles every budget move; the executive board approves anything above €10,000; the two acquisition managers execute in the accounts
- Room for improvement
- Consolidating the 4 platforms takes one person 2 days a month; average cost per acquisition has gone from €34 to €41 in a year; 12 of the 38 campaigns have not been reopened in six months
Maison Verlaine is not looking to spend more: it wants to know what each of its €47,000 a month actually buys, platform by platform, audience by audience and format by format. The agent runs on a company machine and reads the 4 ad accounts, the e-commerce platform and the measurement tool: it consolidates, measures, costs and proposes; the marketing director decides. The exchanges below span a quarter, from the first reconciliation of the platforms to the closing review at constant budget.
This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.
An audience, in an ad account, is the population the ad is shown to: age, area, interests, pages already visited. A format is the shape of the ad: image, carousel, short video.
What I read: your ad accounts read-only, the e-commerce platform and the measurement tool. €141,000 of media budget over the quarter, €47,000 a month.
The gap I measured, and it decides everything that follows: the sum of what your four platforms claim for themselves comes to 1,402 orders last month. Your e-commerce platform counts 1,150 over the same period. 252 orders are counted twice — 22 %.
Why this is not a detail: your cost per acquisition — the budget spent divided by the number of orders obtained — reads €34 if you believe the platforms, €41 if you believe your own till. Same euros, same week. Seven euros of difference across 1,150 orders is €8,050 a month of decisions taken on the wrong figure.
What I now consolidate every morning: by campaign, by audience and by format, from the order actually recorded on your side, never from the sum of the platforms.
The time this moves: that consolidation took 2 days a month from one person, 60 % of the monthly trade-off cycle. It now takes 7 % — a little under two hours, the time to read back what I have assembled.
What I propose: that the marketing director approve a single reference figure — the e-commerce platform's — and that every trade-off this quarter rest on it. One decision, and the four spreadsheets stop contradicting each other. consolidation_4-platforms-90-days.pdf1,402 claimed, 1,150 real · 22 % double counting
⛓ Sourced · 4 ad accounts read-only, e-commerce platform, measurement tool, 90 days
The comparison scope is the period, the campaigns, the platforms and the counting rule applied on both sides of a comparison. Without it, two figures that differ prove nothing.
The three disagreements I settled, in order of what they cost:
· The attribution window — the delay after a click or a view during which an order is still credited to the campaign. Your four platforms were set to 7 days post-click, 28 days post-click, 7 days post-click plus 1 day post-view, and 30 days post-click. The platform set to 28 days credited itself 31 % more orders than the same campaign measured at 7 days. I have brought everything to 7 days post-click, and I keep the other readings alongside.
· Double counting. When two platforms touch the same order, each claims it in full. I reconcile on the order, not on the click: an order is counted once, and I say which platforms were disputing it. Those are the 252 orders from a moment ago.
· Post-view counting. One platform credited itself an order from someone who had seen the ad without ever clicking. 184 orders in the month. I do not erase them — I take them out of the reference figure and keep them on a separate line, because they say something, just not what they were being made to say.
What that gives you: a single, reproducible figure, one you can defend to your executive board line by line.
What I propose next: that every difference be published with its scope written above it. Of the last 6 trade-offs recorded in your spreadsheet, 4 compared two periods of different length — the conclusion may have been right, it was not demonstrated. comparison-scope_rules-applied.pdf4 windows brought to 7 days · 252 duplicates removed
⛓ Sourced · attribution settings of the 4 accounts, order-by-order reconciliation, trade-off spreadsheet
Local inference means the model computes on your machine: the amount of a budget or the definition of an audience never crosses an outside network to be processed. If you would rather not host a machine, the other route is an isolated resource hosted in France, under French law, dedicated to Maison Verlaine — no pooling with another brand.
What that protects, very concretely:
· Your budget split — €21,000 on the social platform, €16,000 on search, €6,000 on video, €4,000 on affiliation. That is your media buying strategy, readable in one line.
· Your performance by audience. Knowing that your « young parents » audience delivers at €27 while your « gift » audience delivers at €58 is knowing where to attack you without spending a euro to find out.
· Your unreleased creative and the calendar of your upcoming operations.
How it is held: encryption in transit and at rest; role-based access — rights follow the job: an acquisition manager opens the campaigns they own, not the overall split nor the margins; logging of every trade-off proposed, with its assumptions, its range and the decision taken; hosting in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
And none of your data trains a model: what I learn from Maison Verlaine serves Maison Verlaine.
The figure that makes this commercial rather than technical: your e-commerce platform and two of your four ad networks already require, in their terms, that you name your data sub-processors. The list fits on one line: processing in the EU targeted. You have no clause to negotiate and no transfer impact assessment to produce — the kind of file that takes three weeks when the answer is anything else.
What I propose: that I keep that technical sheet up to date — hosting, sub-processors, retention periods, who accesses what. It is asked for once a year and hunted for over three days. technical-framework_where-your-budgets-live.pdfLocal inference, role-based access, processing in the EU targeted
✎ Framework · deployment architecture, access log, ad network and e-commerce platform terms
Difference 1 — cost per acquisition is rising on the social platform. €38 → €52, up 37 %. Scope: week of 5 to 11 August against week of 29 July to 4 August, same 11 campaigns, same platform, 7-day post-click window on both sides, two full weeks. Volume: 168 orders against 231. Cost of the difference for the week: €2,350. It is the most expensive of the three.
Difference 2 — one audience delivers at €11, and it is the figure I trust least. On search, the audience typing « maison verlaine » absorbs €4,900 a month for a cost per acquisition of €11. Three times better than your average — because those people were already looking for you. Scope: rolling 30 days, a single campaign, queries containing your brand name. I come back to it in a moment: this is not a difference to correct, it is a difference to measure.
Difference 3 — one format is pulling away. Short video delivers at €27 where the carousel delivers at €44, over the same month, the same 3 audiences and the same 2 campaigns — but it carries only €6,000 of the €47,000. The best format in the house is the least funded.
The time this moves: spotting and qualifying these differences took 30 % of the monthly trade-off cycle. It now takes 5 %, and you get them on Monday morning instead of the 5th of the following month.
What I propose: that we take the three in order of cost. The first bleeds €2,350 a week; the other two are seams, not leaks. weekly-differences_3-with-scope.pdf3 differences · scope shown · cost in euros
⛓ Sourced · 4 ad accounts, e-commerce platform, weeks of 29/07 and 05/08, rolling 30 days
A cost per acquisition always breaks down the same way, and each of the three storeys is measured separately:
· The cost per thousand impressions — what you pay for a thousand displays of the ad, that is, the price of the space — goes from €9.10 to €9.40. Up 3 %. The market explains nothing.
· The click-through rate goes from 1.4 % to 0.9 %. Down 36 %. The whole difference is here.
· The conversion rate holds at 2.8 % then 2.9 %. Your site is not at fault: those who arrive buy as much as before.
What that means: fewer people click, so each order carries more paid displays. The cause is in the ad or in the repetition, not in the price nor in the site.
The number that settles it: frequency — the average number of times one person sees your ad over the period — went from 2.1 to 4.7 in seven days on an audience of 240,000 people. Your 6 creatives have been running unchanged for 61 days. The audience has seen them, and it has stopped looking.
The comparison that sheds light, and I give it for what it is worth: in social acquisition, a click drop-off is commonly observed beyond 3 to 4 exposures a week; that is an indicative market order of magnitude, not a law. Your own history says the same thing more reliably: across your 4 creative refreshes in 2025, the click-through rate came back between 1.3 % and 1.6 % each time, in under six days.
What I propose, and it costs nothing in budget: replace the 6 saturated creatives before touching the split. If the click returns to 1.3 %, cost per acquisition falls back to €40 without moving a single euro — and we will know the difference really came from there. I have built the 6 new creatives from your current collection visuals; your content team decides. breakdown_38-to-52-eur.pdfPrice +3 %, click −36 %, conversion flat · frequency 2.1 → 4.7
⛓ Sourced · social platform, cost per thousand, click and conversion rates, 7-day frequency, 4 refreshes in 2025
What that number says: people who type « maison verlaine » buy. That is true.
What it does not say: how many of them would have bought anyway, clicking on your organic result two centimetres below. That is called incrementality — what the campaign actually added to the sales that would have happened regardless. A cost per acquisition never measures it: it attributes, it does not compare.
The test that settles it, and it is simple: I switch that campaign off in Belgium only, for three weeks, and compare the evolution of Belgian orders with French orders over the same period. Your two markets have tracked each other to within 4 % over the last 12 months — that is what makes the comparison valid.
What the test costs: nothing. Belgium accounts for €640 of the €4,900. If incrementality is nil, you recover €4,900 a month. If it is total, you have lost three weeks of coverage on a secondary market.
The three outcomes, costed in advance: if Belgian orders fall by less than 10 %, most of those €4,900 are buying orders you would have had — I propose redeploying them. Between 10 % and 40 %, we keep the campaign but cap it. Above 40 %, it is worth every euro and I will instead propose increasing it.
What is left to your go-ahead: the switch-off itself. That campaign carries 445 orders a month; stopping it for three weeks is a commercial decision, not a setting. Everything else is already done: the test is built, capped at the €640 Belgium accounts for, dated, with its automatic restart date — one word starts it, one word makes it unnecessary.
What I propose next, whatever the outcome: run the same test on the « gift » audience at €58, the one that performs worst. Two tests, one a month, and you will know what your budget really buys rather than what it claims.
⛓ Sourced · search platform, rolling 30 days, France/Belgium comparison over 12 months
Trade-off A — refresh before you move anything.
· Budget moved: €0. We replace the 6 saturated creatives, nothing else.
· Expected effect: click-through back to 1.3 %, social cost per acquisition €52 → €40, that is +40 orders a month at identical budget. The arithmetic is yours: €21,000 at €52 buys 404 orders, at €40 it buys 525 — I announce 40 because a refresh only holds for three to four weeks.
· What it leaves aside: it corrects nothing in the imbalance between platforms. An effective, short-lived dressing.
· What it costs elsewhere: two days of your content team.
Trade-off B — move €9,000 from saturated social to short video and generic search.
· Budget moved: €9,000, that is €5,000 to short video and €4,000 to generic search. The monthly total does not change.
· Expected effect: the €9,000 taken out of social at €52 were buying 173 orders. Reinvested, they buy 172 in video — I use €29 rather than €27, because a format whose budget triples degrades — and 111 in generic search at €36. Total 283.
· Net expected gain: +110 orders a month, average cost per acquisition €41 → €37, at a strictly constant budget. I state it as a range: +80 to +130.
· What it leaves aside: social loses reach among people who do not know you yet. That is the real risk, and it can be measured in three weeks on new-customer volume.
Trade-off C — open a fifth platform.
· Budget moved: €6,000.
· Expected effect: unknown at your house, and I would rather say so than estimate it. What your history says: across your two platform openings in 2024 and 2025, first-week cost per acquisition ran at 2.3 times target and it took 5 weeks to come back to the promised level.
· What it leaves aside: €6,000 taken from a line delivering at €36, that is 166 certain orders traded for a learning period.
The time this moves: costing a trade-off at this level took 40 % of the monthly cycle. It now takes 8 % — and all three land on the same day, which is the only way to compare them. three-trade-offs_constant-budget.pdf€0 · €9,000 · €6,000 — effect, arithmetic and blind spot
⛓ Sourced · performance by platform and format over 90 days, history of the 2 platform openings in 2024 and 2025
Why A first: it costs nothing in budget and it answers a question. If the click comes back, the difference was saturation; if it does not, it comes from the audience itself, and B changes shape. Doing both on the same day would make the result unreadable — we would not know which one worked.
Why B next: it is the only one of the three that moves money from a measured line to two measured lines. No unknowns: short video has been running at your house for 7 months, generic search for 3 years. I am not asking you to believe a promise, I am asking you to believe your own last 90 days.
Why I advise against C, and this is not caution: C would consume €6,000 over 5 weeks to learn what B measures in 3 weeks without spending anything extra. Cost of that learning, based on your two previous openings: roughly €8,400 of ramp-up overspend. That is not money lost if you open the platform one day — it is money badly placed this quarter, when a measured seam sits right next door. If you want to open C anyway, I will build it: just tell me what budget you accept to spend on learning.
The stop threshold, written before the test and not after — the cost-per-acquisition level beyond which the test stops, decided cold so that nobody has to argue about it hot:
· Short video stops if its cost per acquisition exceeds €45 three days running. At €45 it is no better than the line the money came from.
· The whole thing stops if the house cost per acquisition exceeds €43 over 7 rolling days. That is your worst month last year: beyond it, we have done worse than nothing.
· In both cases the return to the original split is built in advance and takes four minutes.
What I hand you to decide with: the three trade-offs on one page, with their arithmetic, their range and their blind spot. You do not have to take my word for it: every number links back to the account line it comes from.
⛓ Sourced · 90 days by platform and format, opening history, worst cost-per-acquisition month of 2025
What is sitting in your accounts as drafts this morning: the two new budget splits, the daily caps, the three video ads built from your collection visuals, the generic-search keyword list, the two stop thresholds already set and the return to the original split. Execution takes four minutes; it was the preparation that took two days, and it is done.
Why I do not click on my own today: moving €9,000 commits real spend, and your internal note of 14 January says every budget move goes through the marketing director. I have no opinion on that rule: I apply it because it is yours.
The three forms of mandate I know how to hold, if you want to move faster:
· Case by case — that is today. I propose, you approve, I execute within minutes.
· Capped mandate — you authorise me to move budget inside a constant monthly total, up to an amount you set, say €5,000. The monthly total itself never rises through me: no budget raised, no platform opened, no campaign created. Every move is notified to you within the hour with its arithmetic.
· Withdrawal-only mandate — I may bring a budget back to its previous value if a stop threshold is crossed, never raise it. The safest mandate: it can only reduce spend in progress.
In all three cases: the mandate is written, capped, dated, and you withdraw it with a word, without notice and with nothing left running. Every action is logged with its author, its amount and its reason — that log is what lets you answer your executive board six months later.
What I propose this time: let us keep case by case. Approve B this morning, I execute before noon, and I bring you the first reading at D+3 with both thresholds at the top of the page. who-decides-what_media-budget.pdf3 mandate forms · capped, dated, withdrawn with a word
✎ Framework · internal note of 14/01, mandate forms, log of trade-offs and executions
What was executed on Monday at 11:12: €5,000 to short video, €4,000 to generic search, €9,000 taken out of the 4 saturated social campaigns. Monthly total: €47,000 before, €47,000 after.
Where it stands at D+10:
· House cost per acquisition: €41 → €34. I had announced €37.
· Attributed orders: +49 over the period, against +37 expected pro rata of the announced range.
· Short video: €31, against €29 announced — the degradation I predicted did happen, slightly stronger than expected.
· Generic search: €34, against €36 announced.
· Social after withdrawal: €47 — it is coming down, frequency has fallen back to 3.2.
The two stop thresholds, since that is the first thing to look at: short video at €31 against a €45 threshold; house cost per acquisition at €34 over 7 rolling days against a €43 threshold. Neither is anywhere near.
What I am watching and have not settled: new-customer volume, the blind spot I announced. It is at −4 % over ten days. That is inside your usual variation — I draw no conclusion, and I will give it back to you at D+21 with last year's same week alongside.
What I propose now, costed: take short video from €5,000 to €8,000, drawn from the same social line. At a prudent €33 cost per acquisition, those €3,000 buy 91 orders against 64 where they sit: +27 orders a month. Beyond €11,000 I advise against it: your video audience is 310,000 people and frequency would pass 4 before month end — we would replay the exact difference of the week of 5 August.
The decision remains yours, and it is a small one: €3,000 inside a total that does not change. reading-D10_trade-off-B.pdf€41 → €34 · thresholds at €45 and €43, not reached
⛓ Sourced · execution Monday 11:12, 10 days of readings, e-commerce platform, 7-day frequency
The four to reopen: they were delivering at €29 cost per acquisition and were paused in February for a stock shortage on the washed linen range, resolved on 12 March. Nobody reopened them because nobody had noted them. €3,200 a month, around 110 orders — the best ratio in this whole file, and it requires no trade-off: it is an oversight, not a choice.
The six that cost you while asleep, and this is the counter-intuitive point: they spend nothing, but their audiences overlap more than 70 % with live campaigns. When you switched them on for two days in June, your own campaigns bid against each other on the same people — that is self-competition: two of your campaigns targeting the same person push up the price you pay, without either winning an extra order. Measured over those two days in June and scaled to the month: €1,340 of overbidding. I propose archiving them and keeping their creatives, which are good.
The two to test: they target Belgium on a range you were not yet selling there. The Belgian catalogue opened in April. €800 over three weeks, stop threshold at €50, drawn from the €3,200 the first four give back — the test funds itself.
What is yours, and it is all that is left: the tick box. Reopening commits spend; archiving erases a measurement history — two acts that carry your name, not mine. The twelve sheets are ready, each with its original pause reason, its historical cost per acquisition and my one-line recommendation. You tick, I execute within the minute — and the €3,200 from the first four start delivering tonight.
What I propose next: that no paused campaign stay without a due date. I will flag any pause going past 30 days with the reason recorded at the moment of pausing — the four washed-linen campaigns would have reopened on 12 March. 12-paused-campaigns_4-6-2.pdf4 to reopen at €29 · €1,340 of measured overbidding
⛓ Sourced · 12 paused campaigns, original pause reasons, audience overlap, log of the two days in June
The quarter's tally: 14 costed trade-offs, approved by you and executed. 11 inside the announced range. 3 below. 1 negative: −12 orders against +45 announced.
What the four have in common, and it is my error, not yours: all four reference periods contained a promotional operation on the site. I was comparing a week without promotion to a week with one and crediting the difference to media. On those weeks my average error is −18 %; on the other ten it is −2 %. The problem was not the model, it was the comparison scope — exactly what I have been asking you to check since day one.
The three corrections, in place for six weeks:
· I exclude from the reference calculation any week containing a promotional operation, and I flag it above each difference. Over 90 days that removes 3 weeks out of 13.
· I publish a range, not a single figure. « +110 » was wrong by construction; « +80 to +130 » is a forecast, and it can be assessed.
· Every trade-off ships with its stop threshold written beforehand. The negative trade-off would have been stopped at D+6 instead of D+30: the loss would have been 3 orders instead of 12.
Where that stands since: 5 trade-offs, 5 inside the range. That is too few to conclude, and I say so rather than turn it into proof.
Why I give you this figure when nobody asked: because you commit money on my arithmetic. A supplier who publishes only its successes lets you discover its failures on your budget, and you will discover them anyway. Better that it comes with the fix already in place.
What I propose: that this tally be attached to every quarterly review without your having to ask. It reads in thirty seconds and it tells you what my signature is worth this quarter. forecast-reliability_14-trade-offs.pdf11 held, 3 short, 1 negative · 3 corrections
⛓ Sourced · 14 executed trade-offs of the quarter, promotional calendar, readings at due date
The budget first, because it is the only way to read the rest: €47,000 a month on day one, €47,000 a month today. Nothing was added.
What that budget buys:
· Cost per acquisition: €41 → €32. You were at €34 a year ago; you are now below your best known level.
· Attributed orders: 1,150 → 1,470 a month. +320.
· Attributed revenue, at a constant €128 basket: €147,000 → €188,000 a month. That is the only number on this list your executive board will look at, and it comes from an unchanged budget.
Where the 320 orders come from, line by line: +110 from trade-off B, +40 from the creative refresh, +110 from the four reopened campaigns, +27 from the video reinforcement, +33 from stopping self-competition. No line is an announcement effect: each has its reading.
On the three time lines you were measuring:
· Consolidating the platform data: 60 % → 7 % of the monthly trade-off cycle.
· Costing a trade-off: 40 % → 8 %.
· Identifying the differences: 30 % → 5 %.
In days: those three tasks took 4 and a half days a month; they now take three quarters of a day. That time did not go into a spreadsheet: over the quarter your team ran 3 incrementality tests, something it had never had time to do.
And the figure nobody thinks to look at: zero budget change applied without human approval across 19 logged executions. That is not a cautious setting, it is how I am wired into your accounts. quarterly-review_constant-budget.pdf€47,000 unchanged · €41 → €32 · +320 orders
⛓ Sourced · 3 months of readings, e-commerce platform, log of the 19 executions, original trade-off spreadsheet
To the marketing-campaign agent — the one that works on your contact base: I stop at the door of your customer file. What I give it: the 1,470 orders of the month with the campaign, audience and format of origin. What that is for: last month 218 people received a « first order » offer a few days after ordering. With the origin of each order, they drop out automatically. 218 irritating messages fewer, and the offer reserved for those who have not yet bought.
To the qualification and follow-up agent — the one that handles leads: your campaigns also produce 96 trade quote requests a month, hotels and guest houses. They land in a shared inbox and wait 2.4 days for a first reply. I pass it the request with its campaign of origin and the stated budget; it qualifies, follows up and books the meetings. That is not my trade: I measure media buying, not conversations.
To the decision-support agent — the one that queries your data: I work on attributed revenue. Margin by product sits in your commercial management system, and I do not open it. That agent does. The question waiting for it is worth money: of your 6 most heavily funded products, two carry a margin 9 points below the catalogue average. A €27 cost per acquisition on a low-margin product is worth less than €38 on a high-margin one — I can raise the question, I cannot answer it alone, and an agent claiming otherwise would have you decide on a figure it does not hold.
Why three agents and not one: each accesses only what it needs. I do not see your customer file, the marketing agent does not see your budgets, the decision-support agent runs no campaign. That partitioning is what makes role-based access true rather than declared.
What I propose: start with the first handover, the cheapest and the most visible. The order-origin export is ready; it only needs a recipient.
⛓ Sourced · monthly orders with origin, trade quote inbox, catalogue and most heavily funded products
Action 1 — the alert on spend that buys nothing. When a landing page stops responding, clicks keep being charged and the visitor arrives on an error. I do not suspend the campaign: your page is explicit, no campaign is suspended automatically. I alert you within three minutes, with the suspension draft ready to approve and the amount leaving every minute. Two cases this quarter: suspension approved in 9 and 14 minutes, €3,100 of charged clicks avoided — against roughly €41,000 had the outage lasted the day.
Action 2 — excluding recent customers from acquisition audiences. You approved it once, on 3 March. People who ordered within 30 days drop out of campaigns hunting for new customers. Paying to be seen by your own customers is the most ordinary leak in an ad account: yours was worth €2,900 a month. And the reverse is true too: past 30 days, the person becomes eligible again automatically, with the date on record. An exclusion is not a closed door.
Everything else runs through a mandate — and that is exactly what makes it fast: raising a total budget, opening a platform, creating a campaign, changing a bid. Each of those is written, capped, dated and withdrawable with a word; the proposal arrives priced, with the expected range and the stop threshold, and once the mandate is signed I execute it in minutes, not in a meeting. Over the quarter: 19 executions, 19 human approvals, no exception — and a median of 26 minutes between my proposal and its execution.
What the auditor receives, and receives on one page: for each proposed trade-off — the date, the comparison scope, the assumptions, the announced range, the stop threshold, the decision, its author, the execution time and the measured result. That is what the AI Act expects of a system supporting a decision: human supervision that can be proved, not declared.
And the foundation, which only shows the day someone looks for it: Hosted in France, under French law. No data outside the European Union, therefore an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Six advertising use cases ready to deploy. Zero decision taken without human approval.
What I propose to close: that the mandate review be diarised once a quarter. A mandate nobody has reread in a year is no longer a decision, it is a habit — and it is the first thing an auditor looks at. decision-framework_trade-off-log.pdf2 reversible actions · 19 executions, 19 approvals
✎ Framework · 2 actions approved once, 19 logged executions, trade-off log, quarterly mandate review
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
What does the agent actually do?
One agent, several angles on the trade-off. All these uses work in support, subject to your approval.
Consolidating performance
Gathers the data from your platforms by campaign, audience and format.
Documented differences
Presents the notable differences with the volumes and costs that explain them.
Costed trade-offs
Proposes several shifts of budget with their expected effect.
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.
Search & campaigns (SEO/SEA)
Bidding buys visibility; earning it durably on search engines belongs to the SEO agent.
Search & campaigns (SEO/SEA) from 534 € excl. VAT / month Discover the agent →Visual creation & adaptations
This agent arbitrates creatives and budgets; it does not produce them. Creation is a dedicated agent's job.
Visual creation & adaptations from 432 € excl. VAT / month Discover the agent →In 15 minutes we identify the most relevant agent — without oversizing the project.
How many trade-offs can a team examine?
By taking on the consolidation and the costing, the effort shifts towards the media decision. How large the gain is depends on your volume and remains to be confirmed by a pilot.
The stages of your AI agent project
Audit & scoping
15 minutes to target the use case with the best return.
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
Design
We design the agent and its guardrails.
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
Rollout
Going live and training your team.
Operation
Continuous supervision and improvement.
One package, one agent
An advertising campaign agent (consolidation, differences, trade-offs), installed and operated for you.
Setup + controlled subscription
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
All inclusive, no setup fee
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
On site, you own it
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Four guarantees that matter to your media budgets
Related resources
Your questions, our answers
Does the agent change the budgets?
Where does the performance data come from?
How are the trade-offs costed?
How does this differ from marketing campaigns?
Are our budgets protected?
How long does it take to deploy this agent?
Other agents for marketing
Let's size up the potential in your campaigns
15 minutes to frame your platforms and your budgets — hosted in France, supervised, with no commitment.