A platform can find a cheaper conversion while simultaneously reducing commercial quality. It can generate a high-performing hook that breaches a brand rule. It can send traffic to an unexpected page because the objective it was given values only the immediate sale.
The algorithm did not disobey. It optimised an incomplete instruction.
Control starts here.
1. Definition: acquisition AI is a system of delegated decisions
AI applied to paid acquisition encompasses models that predict the probability or value of an outcome, adjust bids, select audiences and inventory, personalise or generate assets, allocate budgets and detect anomalies.
These functions use platform signals, reported conversions, the advertiser’s creative and pages, and sometimes first-party or aggregated data. Their scope, transparency and controls vary by product and change regularly.
Automation does not mean removing people. It changes their work: fewer micro-adjustments and more objective-setting, data architecture, creative, experimentation, compliance and financial arbitration.
The question is not “AI or manual”. It is about assigning each decision to the system with the appropriate information, speed and accountability.
2. Key figures: adoption, quality, value and brand built on fragile foundations
- IAB surveyed more than 500 professionals for its State of Data 2025. Only 30% of agencies, brands and publishers said they had fully integrated AI into the campaign lifecycle; half of those that had not yet done so planned to achieve it in 2026.
- Nearly two-thirds cited data quality and protection, together with tool fragmentation, among the main obstacles. Of eighteen proposed governance solutions, none was being used or planned by more than 49% of respondents.
- In the State of Data 2026, based on more than 400 decision-makers on the buy side, around half say they are already deploying AI at scale in advanced measurement; among the remainder, more than 70% plan to do so within one to two years.
- Around 40% of brand–agency or partner contracts contain AI-related clauses according to IAB 2026, and respondents expect that share to double within one to two years. Yet fewer than 40% say they have, or plan to introduce, responses to major legal, security, accuracy or quality risks.
- Google reports that AI Max typically delivers 14% more conversions or conversion value at a similar CPA or ROAS for non-retail advertisers, according to its internal 2025 data. Smart Bidding Exploration is reported to deliver, on average, 18% more unique converting query categories and 19% more conversions. These are vendor-reported results: they justify a controlled test, not a guaranteed projection.
The gap is clear. Capabilities are advancing faster than the mechanisms for evidence and accountability.
3. Level 0 — What AI can receive and the business does not delegate
Before activating anything, define which decisions remain human or require approval from a named person. An advertising system does not inherently know the minimum contribution, level of legal risk, brand commitments, strategic stock, tolerance for a poor experience or populations a business refuses to target.
3.1. The objective function
The platform receives an action and sometimes a value. If every request for a quote is worth €100, it seeks the expected volume—even if some areas are not served, existing prospects are already in the pipeline or one type of project requires three times as many resources.
Build a useful value signal: net sale, estimated contribution, probability of closing, new-customer status, category or a prudent combination of these elements. Predicted values must be recalibrated against observed cohorts and bounded so that an extreme estimate cannot monopolise allocation.
3.2. Non-negotiable constraints
Permitted geographies, legal age, contractual exclusions, prohibited products, pages not intended for acquisition, disclosures, vocabulary and spending caps are guardrails. Meta states, for example, that its Advantage+ audience allows strict criteria such as minimum age, location, language and certain exclusions to be set—but they still need to be configured and monitored.
Health, finance, employment, housing and advertising involving minors require particular care. A qualified legal team determines the applicable rules; a suggestion in an interface is no substitute for that analysis.
3.3. Accountability
Name an owner for each class of decision: value, data, audience, creative, budget, compliance and incidents. The provider operates a model. The advertiser retains responsibility for what it funds, transmits and distributes under its contractual and regulatory framework.
Nothing has been automated yet. That is intentional.
4. Level 1 — Support analysis without taking direct action
The first level uses AI to summarise, classify and detect while a person confirms the conclusion. It is often the best place to start because errors remain reversible and the criteria can be observed.
4.1. Useful use cases
- group queries by intent and identify new themes;
- classify comments or reasons for loss;
- detect a break in volume, tracking or cost;
- propose a creative taxonomy;
- summarise a change history;
- flag pages that are inconsistent with the advert;
- prepare budget scenarios for approval.
Evaluate quality on a reference sample. For query classification, have two specialists annotate a set, measure precision and recall by category, then pay particular attention to costly errors: confusing a support search with purchase intent is not equivalent to a punctuation mistake.
Retain the context and version. A summary produced after the data has changed cannot be reproduced without the date, source, prompt or relevant parameters. For sensitive decisions, the log must make it possible to understand what the operator actually saw.
This level frees up time. It does not transfer authority.
5. Level 2 — Automate bidding within a closed scope
Automated bidding estimates the probability and sometimes the value of a conversion for a given opportunity, then adjusts the bid in real time. It can incorporate more signals than a human team, but its performance depends closely on the objective, volume and stability of feedback.
5.1. Prepare the data
An event must be deduplicated, timely, consistently defined and connected to the business outcome. If sales arrive sixty days later, consider choosing an early indicator whose relationship with closing is measured, then import mature outcomes to recalibrate.
For the best quality in data-driven attribution, Google recommended 200 conversions and 2,000 interactions over thirty days. This benchmark is not a universal minimum for every bidding system, but it illustrates a principle: fragmenting a small volume across twenty campaigns reduces learning capacity.
Avoid circular values. If a lead score already depends on the channel and is then used to demonstrate that the channel produces the best leads, the algorithm reinforces the original preference without independent evidence.
5.2. Define the sandbox
Start within a scope whose pages, areas, budgets and conversions are reliable. Set a spending cap, a CPA or ROAS consistent with the economics, exclusions and a minimum period. Do not change three levers while the system is learning.
Monitor average and marginal outcomes, quality, the share of new customers, product mix and time lag. An improvement in CPA may come from a shift towards brand or retargeting while incremental prospecting contracts.
Automation wins at micro-allocation. Management retains control of the threshold.
6. Level 3 — Let AI expand audiences, queries and pages
At this level, the system no longer decides only how much to bid; it also explores where and with whom. Growth potential increases alongside the risk of dilution.
6.1. Understand what is expanding
AI Max for Search combines broad match, keywordless technology, text customisation and final URL expansion, among other features. Google states that enabling URL expansion may send traffic to the page it considers most relevant and that some pinned assets will then not be respected. Brand, location and URL inclusions and exclusions are available depending on the settings.
These details are strategic. A recruitment page, customer support page, old promotion or low-stock product must not become a destination simply because it contains the right words. An inventory of eligible URLs and monitoring of pages actually served are necessary.
On social platforms, broad audiences allow the model to exploit internal signals that are unavailable to the advertiser. Separate flexible suggestions from strict constraints, monitor the proportion of existing customers and measure genuinely new reach instead of drawing conclusions from attributed volume alone.
6.2. Test expansion
Google offers AI Max experiments that split part of a campaign’s traffic between control and treatment, reducing some configuration errors associated with copies. Before launch, define the primary metric, guardrail, duration and action for every possible outcome.
The analysis does not end with the total. Examine additional queries, categories, URLs, new customers, contribution, brand, commercial quality and the uncertainty interval. An overall increase may hide deterioration within the scope the business actually wanted to develop.
Expansion deserves its own economics. As long as it remains blended into the established core, historical results subsidise exploration and conceal its marginal cost.
Open does not mean blind.
7. Level 4 — Generate and assemble creative
Generative AI can produce copy, images, video, voice, variants and combinations. It accelerates execution and increases testable variety, but it does not replace a creative strategy or the verification of claims.
7.1. Build the machine brief
Provide the objective, audience, situation, tension, promise, authorised evidence, tone, prohibited words, disclosures, formats, rights, pages and brand examples. Add what must never be invented: price, certification, testimonial, comparison, medical result or contractual commitment.
A good library separates approved, conditional and prohibited assets. It records provenance, rights, expiry date, markets and modifications. Without this register, production volume quickly exceeds review capacity.
7.2. Organise validation in proportion to risk
Low-exposure informational copy can undergo automated checks followed by human sampling. A sensitive claim, image of a person, regulated translation or widely distributed message requires explicit approval before publication.
Check accuracy, brand, rights, representation, readability, accessibility and consistency with the page. Performance does not turn an unsupported claim into acceptable content.
7.3. Learn at concept level
Classify assets by opening angle, benefit, evidence, format and call to action. If the platform automatically assembles combinations, require reporting that is as close as possible to these dimensions and retain structured creative experiments to understand what transfers from one audience to another.
Generation increases supply. Judgement protects demand.
8. Level 5 — Orchestrate budgets and decisions across channels
The most advanced level recommends or executes reallocations across campaigns and channels, incorporating forecasts, constraints and outcomes. It is also the level where a definition error has the greatest financial impact.
8.1. Do not confuse the local optimum with the business optimum
Each platform has a privileged view of its own interactions and its own commercial function. A campaign may maximise attributed conversions by capturing more brand demand while the business needs new customers or future demand.
The orchestrator must therefore receive an independent reference set: net sales, contribution, new customers, capacity, stock, brand objectives and incremental outcomes. Platform data supports operations; it must not become the sole arbiter of cross-channel allocation.
8.2. Use several methods of evidence
Attribution for diagnosis, experiments for causality, cohorts for quality, MMM for aggregate allocation and finance for value: no single method answers every question. IAB 2026 also reports that 60% to 75% of users still consider advanced measurement inadequate on major dimensions.
AI recommendations must display assumptions, data, intervals, constraints and sensitivity. If two nearly equivalent scenarios lead to very different allocations, the right action may be to fund an experiment rather than choose arbitrarily.
8.3. Maintain a route back
Set limits by day, week, channel and type of decision. A tracking anomaly, incorrect promotion or cost increase triggers a freeze or return to a known configuration. Changes are versioned so that the previous state can be restored.
Maximum autonomy requires maximum observability. Otherwise, it simply becomes a faster rate of loss.
9. The Logiks delegation contract
Logiks recommends a record for each automated function. This internal framework must be adapted to the organisation’s risks and obligations.
- Delegated decision: bid, targeting, page, creative or budget.
- Objective: event, value, time horizon and financial threshold.
- Inputs: sources, quality, consent, delay and owner.
- Freedom: exploration scope and strict constraints.
- Evidence: baseline, control, metric and interval.
- Monitoring: anomalies, bias, quality and drift.
- Authority: who approves, stops, modifies or restores.
- Traceability: versions, changes, assets and decisions.
- Review: expiry date and renewal criteria.
The record fits on one page. It prevents an option enabled during a test from silently becoming the account’s permanent policy.
10. 90-day deployment programme
10.1. Days 1 to 20 — Inventory and set boundaries
List every AI function already active, including those enabled by default. Map objectives, data, pages, audiences, generations, contracts and owners. Disable or constrain uses without an owner.
10.2. Days 21 to 40 — Repair the objective function
Reconcile conversions, CRM and finance. Deduplicate, qualify, introduce useful values and create alerts. Establish strict constraints and the brand library.
10.3. Days 41 to 65 — Test two delegations
Choose a low-risk analytical task and a media optimisation with measurable impact. Establish control, treatment, primary metric, guardrail and rollback protocol.
10.4. Days 66 to 80 — Examine errors
Do not look only at the average. Examine false positives, degraded segments, unexpected pages, rejected creative, existing customers, contribution drift and difficult cases. Errors determine the level of supervision required.
10.5. Days 81 to 90 — Expand or withdraw
Expand if the evidence, quality and observability hold. Maintain under supervision if the gain is real but fragile. Withdraw if errors or control costs exceed the value created.
Progressive. And reversible.
11. Executive control table
| Area | Value indicator | Risk indicator | Decision |
|---|---|---|---|
| Bidding | marginal contribution | mix or CPA drift | expand, constrain, revert |
| Audiences | incremental new customers | exclusion, frequency, bias | open or narrow |
| Pages | net conversion | prohibited URL, inconsistency | allow or exclude |
| Creative | outcome by concept | claim, rights, brand | approve or withdraw |
| Measurement | lift / iROAS | uncertainty, contamination | accept, retest, suspend |
| Governance | time and quality gained | incident, access, missing audit trail | renew the contract |
The committee does not ask whether “AI works”. It examines which delegation produces which value, within which limits and with what level of confidence.
12. Logiks advice: ten automation mistakes
- optimise an easy event instead of the business outcome;
- send every conversion as a primary objective;
- open targeting and URLs without verified exclusions;
- judge a system during its learning phase;
- apply a vendor-reported uplift to the business plan;
- generate claims without a source or review;
- entrust the cross-channel budget solely to platform-reported results;
- hide intervals and failure cases;
- leave default options enabled without an inventory;
- deploy without a shutdown or rollback procedure.
AI reduces some decision costs. It increases the potential cost of a poor instruction.
13. FAQ
13.1. Does AI replace a media buyer?
It automates micro-decisions and part of the analysis. The role shifts towards objectives, data, creative, experimentation, governance and economics, where context and accountability remain essential.
13.2. Do you need many conversions?
More stable signal generally improves learning, but the requirement varies by product and function. Concentrate useful data, avoid fragmentation and measure stability before expanding.
13.3. Do broad audiences always work better?
No. They give the model more freedom, which can improve discovery or dilute quality. Test them with constraints, commercial segmentation and incremental measurement.
13.4. Can AI choose the landing pages?
Yes, within a cleaned scope, with exclusions, tracking, compliance and reporting on the pages served. Sensitive, outdated or non-transactional pages must be blocked.
13.5. How do you control generated creative?
Use a structured brief, a library of approved evidence and usage rights, prohibited-content rules, risk-proportionate validation, traceability and rapid withdrawal.
13.6. How do you prove the gain from AI?
Compare control and treatment on a business metric, with guardrails and sufficient duration. Also measure time saved, supervision cost, errors, quality and additional contribution.
14. Conclusion
Advertising AI is powerful precisely because it makes thousands of decisions. The right response is neither to reject it nor hand over everything, but to build an explicit delegation.
Retain strategy, limits and accountability. Automate analysis, then bidding. Expand audiences and pages through experiments. Industrialise creative with validation. Finally, orchestrate budgets against an independent source of truth.
Autonomy must be earned through evidence.
15. Main sources
- IAB, State of Data 2025: The Evolution of AI for Media Campaigns, 24 March 2025: https://www.iab.com/news/iab-state-of-data-report-2025/
- IAB, State of Data 2026: The AI-Powered Measurement Transformation, 2 February 2026: https://www.iab.com/insights/2026-state-of-data-report/
- IAB, press release Project Eidos, key 2026 findings: https://www.iab.com/news/iab-announces-project-eidos/
- Google Ads Help, About AI Max for Search campaigns: https://support.google.com/google-ads/answer/15910366
- Google Ads Help, How AI Max works: https://support.google.com/google-ads/answer/15910187
- Google Ads Help, About AI Max experiments: https://support.google.com/google-ads/answer/16450159
- Google Ads Help, Google Ads Highlights of 2025, Smart Bidding Exploration: https://support.google.com/google-ads/answer/16756291
- Google Ads Help, About data-driven attribution: https://support.google.com/google-ads/answer/6394265
- Meta for Business, Advantage+ audience: https://www.facebook.com/business/ads/meta-advantage-plus/audience