A Wrong Number in a Report Is Embarrassing. In an Optimization, It Moves Your Budget.

In an agency, an error in a client report almost always ends the same way. You apologize, you correct it, you send the right version. Provided you have spotted it. Most of the time, nobody does: the numbers are not recomputed live, the document goes out, it is read in a meeting, and the next month begins.

For a long time, that was not serious. The report described the past. Decisions were made elsewhere, by people who knew the account and could tell when a number felt off.

With agents, that safety net disappears. When an AI no longer just summarizes but recommends and optimizes, the wrong number no longer ends up in a PDF. It ends up in a decision. You optimize on bad data, so you decide badly, and the machine does it fast, without hesitation, across several accounts at once.

What marketers are living through

AI errors are already part of daily work. In the global study by KPMG and the University of Melbourne published in 2025, covering more than 48,000 people in 47 countries, 56% of employees who use AI say they have made mistakes in their work because of it. On the marketing side, a study by the agency NP Digital published in February 2026, among 565 US marketers, found that 47.1% encounter inaccuracies several times a week, and that more than a third (36.5%) have seen hallucinated or incorrect content go live publicly.

Above all, these errors get through because nobody checks. Two employees in three have relied on an AI answer without evaluating it, according to the same KPMG study. According to McKinsey, only 27% of organizations using generative AI have all of its content reviewed before use. And reviewing is not always enough: spreadsheet research, summarized by Raymond Panko, shows that a single reviewer catches only about 60% of errors on average.

The platforms themselves are not immune. In 2016, Facebook acknowledged calculation errors in four of its metrics, including an organic reach figure that added up daily numbers without deduplication and stayed in production for months. It later settled advertisers' class action over two miscalculated video metrics for 40 million dollars, without acknowledging any impact on their spending. Little wonder that, according to the IAB, 60 to 75% of advertisers and agencies say their advanced measurement falls short on rigor, timeliness, trust or efficiency.

All this checking comes on top of an already full week. According to NP Digital, more than 70% of marketers spend one to five hours a week fact-checking what AI produces. In our own study, run in 2026 with 48 media consultants and managers, 60% of the week goes to monitoring, reporting and client relations, against 11% for optimization. When time runs short, whatever has no deadline goes first: optimization, and checking the numbers.

McKinsey sums up what comes next: with agents, the risk is no longer only an AI that says something wrong, but an AI that does something wrong, by taking unintended actions. In its 2026 survey of about 500 organizations, 74% consider inaccuracy a highly relevant risk.

Why a language model gets a number wrong, with confidence

A language model does not read a table the way a spreadsheet does. It predicts the most likely continuation of a text. Most of the time, the likely continuation is also the right one. Sometimes it is not, and nothing in the answer says so.

OpenAI researchers put it plainly in September 2025: the way models are trained and evaluated rewards guessing over acknowledging uncertainty. A model that always answers scores better than one that says "I don’t know". So it learns to answer.

Calculation has long been the most visible weak point. In 2023, GPT-4 got only about 59% of three-digit by three-digit multiplications right. In 2024, Apple researchers showed that changing only the numbers in a problem lowered the performance of every model tested: that is exactly the situation of an ad account, the same questions every day with different numbers. The best models of 2026 calculate far better, but still without any guarantee, while one line of code returns the exact result every time.

The model makers say so themselves. Anthropic’s documentation states that Claude runs code for non-trivial, precision-sensitive calculations. OpenAI presents its code interpreter as the tool for data and math problems. And Anthropic writes, about hallucinations, that critical information should always be validated, especially for high-stakes decisions.

The same number never exists twice

There is a quieter source of error, and a more frequent one in paid media: the model receives numbers that are correct, but not comparable.

A conversion does not have the same definition from one platform to the next. Google Ads counts it, by default, on the date of the click. Meta counts it on the date the ad was seen. The two do not apply the same attribution window, and part of Meta’s conversions are estimated through modeling. Same word, same sales, two different numbers for the same week.

A model that adds these numbers together is not being clumsy. It is adding quantities that do not measure the same thing, and it does so with the same confidence as if it were right.

A number without its source, its period and its calculation rule means nothing.

In an optimization, the wrong number becomes a decision

This is where the stakes change scale. The platforms themselves separate the number that is displayed from the number that bids.

Google built an entire tool for this case: data exclusions. When conversion tracking has been broken, they stop Smart Bidding from learning from those days. According to Google, they only affect the data used for bidding, not reporting. And when a week or more of clicks is affected, performance may fluctuate for one to two conversion cycles. Wrong data does not stay in the past: it keeps acting.

Meta says the same thing in other words. Without conversion deduplication, numbers can be artificially inflated, which, according to Meta, can mislead optimization and budget allocation decisions. And every significant edit to an ad set can restart the learning phase, which usually ends after about fifty results in a week. Meta advises editing an ad set only when you have reason to believe the edit will improve performance.

Put those two facts side by side. An agent reads a wrong cost per acquisition. It concludes that a campaign is slipping, changes its targeting or cuts its budget sharply. The campaign goes back into learning, its cost rises for a few days, and the next number seems to prove the agent right. The agent insists, cuts again, and a campaign that was working ends up no longer working.

That is the whole difference with a report. An error in a report stays in the document. An error in an optimization backs a bad decision, which produces new numbers, which justify the next decision. The error does not correct itself: it feeds on itself.

Clean data, exact calculations: the rule we apply

At Laier, AI serves the optimization of your campaigns, across all your ad platforms. That is precisely why we never hand it a calculation.

The model reasons, the code calculates. Every number Laier displays or uses comes from a query on your platforms' data, never from an operation performed by the model. A sum, a ratio, a cost per acquisition or a return on ad spend is produced by code, identically every time it is recalculated.

Every number keeps its source. The platform, the period, the attribution rule and the time it was read stay attached to the number, and two quantities are only compared when they measure the same thing.

Correct numbers are not enough; the reasoning has to be right too. A campaign is judged on what it actually optimizes for, with enough volume to draw a conclusion, before any recommendation.

You approve before anything is written. Laier proposes the change and explains which numbers it rests on, you approve, Laier applies it on the connected platforms, then measures what it produced against the same definition.

The goal is simple: 100% of the calculations that lead to a decision are exact, and no optimization rests on a number nobody can trace.