A playbook for an AI agent in advertising: what it does on its own, where it asks and how to check it

Why an agent needs a written playbook

When a team first connects an agent to Yandex Direct or VK Ads, the conversation is almost always about the technology: how to grant access, which protocol to use, which fields to read. The rules come up later, usually after the first unpleasant morning, when the agent has spent the night redistributing budget in a way no human would ever have done. Formally the agent was right: the cost per enquiry in one campaign was lower. It simply did not know that this campaign led to a building where the flats had sold out a week earlier.

The playbook is precisely that part of the knowledge that lives in the marketer's head and has never made it into the ad account. What can be touched and what cannot, up to what amount a decision is taken without sign-off, who to inform and in what form. While a person ran the advertising, these rules did not need writing down: they remembered them. An agent remembers only what has been written for it.

A good playbook fits on two pages. It does not describe how to optimise advertising: the agent has been trained for that better than any document could manage. It describes boundaries: where the agent acts on its own, where it acts and reports, where it must stop and ask. In the article on the AI-first approach we called this the second step of implementation. Here we look at what that step consists of.

Three zones of authority

The first zone: the agent acts on its own and does not distract the person. Everything reversible and cheap goes here. Negative keywords for queries without conversions, bid adjustments within a set range, pausing an ad with zero conversions once enough statistics have accumulated, moving budget between ad groups within one campaign. A mistake in this zone is cheap and can be rolled back in a minute, so sign-off would only slow the work down.

The second zone: the agent acts and must report. Changes the person needs to see but which do not require their signature. Moving budget between campaigns for the same project, launching new ads in an existing ad group, switching the strategy from manual bids to automated and back, extending the geography within an agreed list of cities. The agent's morning summary starts with exactly this block: what was changed, on what data, what effect to expect.

The third zone: the agent prepares a decision but does not take it. A new campaign, a budget above the threshold, switching off a channel entirely, changing the landing page, any action on a project the sales team has flagged as problematic. Here the agent brings the person a calculation and a recommendation, and the person answers yes or no. Without an answer nothing happens.

The boundary between zones is set not by the complexity of the task but by the cost of a mistake and the speed of rollback. The more expensive the consequences and the longer they take to fix, the higher the zone. This principle is easier to explain to a team than a list of fifty exceptions, and it does not go out of date when new tools appear.

Thresholds in money and in time

Thresholds in the playbook are best set in roubles and days, because that is how the owner who ultimately pays for the advertising thinks. The wording "no more than 15 per cent of the budget" sounds tidy, but on a budget of three million that is 450,000 ₽ the agent can move on its own. The wording "no more than 50,000 ₽ a day without sign-off" is clear to everyone and does not depend on how the budget changed this month.

The first threshold: the daily limit on independent reallocations. The amount the agent may move between campaigns in a day without asking. For a project with a budget of one to five million a month we usually set 30,000 to 80,000 ₽ a day. The threshold is not about trust in the model, it is about how much money you are prepared to lose overnight if the data in the ad account turned out to be wrong.

The second threshold: the cost of a qualified enquiry at which the agent must stop the combination and report. Not the cost per click and not the cost of a form lead, but the cost of an enquiry the sales team has recognised as relevant. If a combination brings enquiries above the threshold three days in a row, the agent pauses it and raises the question with the person. If below, it continues and scales up.

The third threshold: time. How many days the agent waits before declaring a combination burnt out, and how many days it gives a new combination to gain traction. In real estate an enquiry takes a long time to mature, and an agent that switches a campaign off after two days without conversions will kill half of the working combinations. We set five to seven days to gain traction and three to five days to declare burnout, depending on the traffic volume.

The fourth threshold: total spend over a period. Even if every individual action fits within the limits, the agent must not spend more than the agreed amount in a week. This threshold protects against the situation where ten small correct decisions add up to one large wrong one.

What the agent must record

Every action the agent takes in the ad account must leave a record a person can read in ten seconds. What was changed: campaign, ad group, ad or bid, old and new value. On what data: period, number of enquiries, cost per enquiry, where the qualification data came from. What effect is expected and when to check it. Without the third point the log turns into a list of edits from which it is impossible to understand why they were made.

The log is not there for control for its own sake. It does two things. First: it turns the agent's work into training material for the team. A marketer who spends a month reading the agent's explanations starts to understand their own advertising better. Second: it removes the main question a client has for any contractor. When the owner asks what was done with their budget, the answer is not a presentation but a list of changes with reasons and results.

Separately, the agent records everything it wanted to do and did not do because of a threshold. This is the most useful part of the log. If the agent hit the daily reallocation limit ten times in a week, the threshold is too low and is holding back results. If it never did, the threshold can be left alone, but it is worth checking whether it works at all.

The summary format: words first, then a table. The first paragraph answers the question of what changed and why. The table goes as an appendix for those who want to check the figures. We have seen the opposite order many times, when a person is sent a forty-row export and stops opening it after a week. A summary that is not read does not exist.

The weekly review: how a person checks the agent

A playbook without review stops working within a month. Not because the agent starts breaking it, but because the market changes while the thresholds stay the same. So the playbook must have a built-in procedure: once a week the person sits down with the log and answers four questions.

The first question: which of the agent's actions would I have done differently. Not in terms of the result, which is not yet known, but in terms of logic. If there are no such actions, the threshold can be relaxed. If there are many, the threshold needs tightening, or the rule the agent misunderstood needs rewriting.

The second question: where did the agent hit a boundary and it cost us money. A combination the agent wanted to scale but could not because of the limit, and over the week it produced the best enquiries on the project. That happens, and it is a signal to raise the threshold for that specific type of action.

The third question: does the qualification of enquiries in the CRM match what the agent is looking at. If the sales team has stopped setting statuses, the agent optimises the advertising blindly and confidently. Checking this junction once a week saves you from the most expensive mistakes, because the agent will do exactly what follows from the data, including the wrong data.

The fourth question: what has changed on the project and in sales that the agent does not know about. The studios have sold out, the sales team is away for the holidays, the bank has raised the mortgage rate, a competitor has launched a promotion. These events do not reach the ad account by themselves. The person brings them to the agent as updated rules, and that is the only way to keep the playbook alive.

A sample playbook for residential complex campaigns

Here is an abridged sample playbook for advertising a residential complex with a budget of around three million roubles a month on Yandex Direct and VK Ads. Zone one, the agent acts on its own: negative keywords for queries with spend from 3,000 ₽ and no enquiries; bids within plus or minus 30 per cent of current levels; pausing ads with spend from 5,000 ₽ and no enquiries over 7 days; moving budget between ad groups within one campaign up to 40,000 ₽ a day.

Zone two, the agent acts and reports in the morning summary: moving budget between campaigns for the same project up to 60,000 ₽ a day; new ad copy in existing ad groups; changing the strategy within a campaign; pausing a combination if the cost of a qualified enquiry is above 12,000 ₽ three days in a row.

Zone three, the agent prepares and waits for a decision: a new campaign or a new landing page; reallocation above the daily limits; weekly spend above 800,000 ₽; switching off a channel entirely; any actions on buildings from the sales team's stop list. Traction period for a new combination: 7 days. Declaring burnout: 5 days. A log of all actions in the shared project channel, a summary every morning at 9:00, a review with the marketer on Mondays.

The figures in this example are not universal. The cost-per-enquiry threshold depends on the class of the project and the margin, the reallocation limits on the budget and the number of campaigns. What is universal is the structure: three zones, four thresholds, a log and a weekly review. With this structure the agent becomes an employee the person can answer for, because they understand exactly what it does and where it will stop. At Matveo we draw up such a playbook together with the client's team in the first week of work and rewrite it after every review until it stops changing.

Frequently asked questions

Can you do without a playbook if the agent is running in test mode?Test mode is the first version of the playbook: the agent only reads and suggests, changing nothing. It is worth spending one or two weeks like this to see which decisions the agent would have proposed, and to write the thresholds for the first zone on that material. Letting the agent make changes without written rules is not worth it even on a small budget.

How often should the thresholds in the playbook be changed?After every weekly review, if there are grounds: the agent hit a limit and it cost money, or the opposite, it did something the person would have done differently. Usually the playbook changes actively for the first four to six weeks, then stabilises and is adjusted once a month or when the budget or the project changes.

Who on the team is responsible for the playbook and the review?One person, the same one whose name is on the project. Usually the project marketer or the head of marketing. The review takes 30 to 60 minutes a week, and it is the most useful meeting in the working calendar, because it shows how the budget is really being spent.

What should you do if the agent has broken the playbook?An agent cannot break a written threshold if the threshold is implemented technically rather than described in text for information. So the money thresholds must sit in the agent's code or in the access rules, not in the document. If the agent did something unexpected within the thresholds, that is a fault in the rule, and it gets fixed at the review.