Failure budgets: deciding how much agent error is acceptable
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Marketing agents now draft, segment, schedule, and spend on real programs. Every one of them makes mistakes, and so does every person and every process they work beside.
The useful question is how often an agent fails, how badly, and whether that rate is acceptable for its work. Most teams never answer it in writing. They react to single mistakes with alarm, or they let patterns build until something serious happens.
Borrowing from reliability engineering, we recommend a failure budget for each agent: an agreed level of acceptable error, defined before the agent goes live and tracked like any other operating measure.
Why teams need a number
Without an agreed number, each mistake is judged on its own. A minor tone slip in a visible campaign can trigger a pause, while a slow rise in factual errors goes unnoticed because no single case looked bad. The response tracks how visible a mistake was, not how much it mattered.
A budget replaces that judgment with a comparison against an agreed level. A failure within the level gets logged, and the team moves on. A failure beyond it triggers the action agreed in advance. The emotional debate happens once, before launch, rather than after every incident.
Defining failure by severity
Start by defining what counts as a failure for this agent. We suggest three severity levels, written with the people who review the agent's work, since they know which errors cost the most.
- Critical: reaches a customer with a false claim, violates consent, spends money incorrectly, or exposes data.
- Major: requires significant rework before use, or misses an important requirement of the brief.
- Minor: needs small fixes, such as wording, formatting, or tone adjustments.
Give each level two or three concrete examples from your own work. Reviewers will disagree about borderline cases, so check agreement the same way you would for any rubric. Have two reviewers classify the same set of failures and resolve the differences in writing.
Setting the budget
For each severity level, set an acceptable rate over a fixed window, such as failures per 100 tasks per month. Rates are easier to compare than raw counts, because volume changes from month to month.
- Critical failures should have a budget of zero for customer-facing work. If one occurs, the agent steps down the autonomy ladder until the cause is fixed.
- Major failures should have a small, specific budget, informed by the error rate of the existing human process.
- Minor failures can have a larger budget, since normal review catches them.
The baseline matters. Measure the current manual process with the same severity definitions before you set the agent's budget. Demanding perfection from an agent when the human process is imperfect may be unreasonable. Accepting more serious errors than the baseline is unwise.
A worked example
Consider a hypothetical lifecycle email agent that drafts about 400 emails a month. Before launch, the team reviews a month of human-written emails using the new severity definitions. Suppose it finds a major error rate of about two per 100 emails.
The team sets the agent's budget at two major failures per 100 tasks, zero critical failures, and 10 minor failures per 100. It agrees that exceeding the major budget for two weeks in a row moves the agent back to full review.
In its third month, suppose the agent records three major failures per 100 tasks. Two of them share a cause: an outdated shipping cutoff in the reference material. The budget did its job. It turned a pattern into a decision, and the fix went to the data rather than the agent's instructions.
Measuring against the budget honestly
A budget is only as good as the count behind it. Two problems distort that count, and both need a method rather than goodwill.
The first is detection. You only see failures that someone checks for. If reviewers inspect every output, the count is close to complete. If they review a sample, you need a sampling plan.
Draw a random sample each week, classify every failure in it, and estimate the rate from the sample. Report the estimate as a rate with its sample size, so readers know how much weight it carries.
The second is small numbers. An agent that runs 50 tasks a month cannot reliably distinguish a 2% error rate from a 4% one. For low-volume agents, we suggest a longer window, such as a rolling quarter, or a budget expressed as a count over that period.
Critical failures need their own channel. Customers, sales colleagues, and support teams often find them first. Give those reports a direct route into the failure log, and trace each one through the run history described in our post on observability for marketing agents.
Using the budget
Once the budget is set, it needs a routine. We recommend three habits, each agreed before the agent goes live.
- Track it weekly: review results against the budget each week, with the trend over the last month.
- Act when it is spent: take the predefined action, whether that is more review, less autonomy, or a pause while the cause is found.
- Record stability: note each month the agent stays within budget, since that record supports decisions to expand its role.
The predefined action should be specific. "Increase review" is vague. "Return to full human review for two weeks, then resume sampling if the rate is back within budget" is a rule people can follow. Pair it with clear escalation paths, so the agent itself stops and asks when it meets cases near the failure categories.
Why explicit budgets help
Explicit budgets replace emotional reactions with agreed rules. When a minor mistake happens, the team can see it is within budget and move on. When a pattern of major mistakes emerges, the budget makes clear that action is required.
They also make conversations with leadership easier. A statement such as "this agent has stayed within its failure budget for three months, and its major error rate is below our manual baseline" is honest and checkable. It invites a decision rather than a debate.
What we are studying next
Budgets set at launch age as the work changes. A new product line or a revised legal standard can change what counts as major. We are studying how often budgets need revisiting, and whether a budget should tighten automatically as an agent earns more autonomy.
How to apply this: for one agent, define critical, major, and minor failures with its reviewers, measure your manual baseline with the same definitions, and set a monthly budget for each level. Agree in writing what happens when a budget is exceeded.
Related research
Leading indicators for agent programs: what to watch weekly
Business outcomes take months to show up. A few weekly leading indicators, checked against normal ranges, show much sooner whether an agent program is healthy.
Incrementality testing for agent-driven campaigns
The reliable way to know whether an agent changed results is to compare against what would have happened without it. Incrementality testing does that.
Golden datasets: building test sets for marketing agent evaluation
A small set of test cases with agreed good answers is the most useful evaluation tool a marketing team can build. It can be built in a week.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
