Token budgets and audit logs for GTM agents

Rise, Brainlabs and PMG have all built their own token caps and audit logs around media buying agents this year because no vendor shipped them. Gartner puts 60% of organisations using AI on course for cost overruns from usage nobody tracked. Media agencies hit the problem first because they run agents against client budgets. Every revenue team running agents has the same problem, whether or not the invoice has arrived yet.

A sequence enrolment costs the same on the thousandth run as on the first. An agent does not. It reads the account, decides it needs another source, calls a tool, re-reads the result, decides again. Each loop is tokens, and the loop count depends on the account, the model, and how well you wrote the brief. Ten accounts might cost £2. The eleventh, with a messy website and three subsidiaries, might cost £40 on its own. Nothing in a HubSpot workflow or an n8n canvas warns you.

Three controls fix this. Put all three on every agent before it goes live.

1. A per-run token budget

Set a hard ceiling on tokens per agent run and make the agent stop when it hits the ceiling, returning whatever it has. Start at three times the median cost of your test runs. A research agent that normally spends 40,000 tokens gets a 120,000 cap. When it hits the cap, that account gets flagged for a human rather than silently costing ten times the average.

In n8n, the pattern is a counter you increment after every model call, checked before the next one, with the run routed to a “needs review” branch when it trips. Anthropic added budget controls to managed agents on 1 September. If your harness exposes a budget parameter, set it at session creation, because the default is no limit.

The cap is also a quality signal. Runs that hit it are almost always accounts where the agent could not find what it was looking for, and those are the accounts where its output was going to be weakest anyway.

2. A per-day spend cap per workflow

The per-run budget stops one account from blowing up. It does nothing about a trigger that fires 4,000 times because someone re-imported a list. Give each agent workflow its own API key with a daily spend limit set at the provider, and set it low: twice what a normal day costs. When the key stops working, the workflow fails loudly and the error branch pages someone.

One key per workflow. A shared key means the account research agent, the reply classifier and the content repurposer all draw from one pool, and when the pool empties at 3pm you cannot tell which one drained it. Separate keys turn the provider’s billing page into a per-workflow cost report you did not have to build.

3. An audit log of every decision

For each run, write one row per tool call: timestamp, account, tool name, tokens in, tokens out, the model’s stated reason for the call, and the result it got back. A Google Sheet is fine for the first month. A Postgres table is fine after that. Most workflow tools give you the token counts on every model node already; the reason and the result are what you have to capture deliberately.

The log pays for itself the first time an agent does something odd. Without it, “why did it email the CFO of a churned account” is a guess. With it, you can see the tool call that returned a stale CRM stage and fix the enrichment order rather than the agent. It also answers the question finance will eventually ask, which is cost per opportunity by workflow. Tokens per run, times runs per opportunity, from the log, in one query.

Name the owner

Every agency in the Digiday piece attached a named human to every AI deliverable. Do the same for every agent workflow: one person whose name sits on the audit log and who reviews the runs that tripped the budget. Agents without an owner get tuned by nobody and switched off by finance.

What this gets you

With the three controls in place, an agent’s cost is a line you can put next to the pipeline it produced. That number decides whether the agent expands to the next segment or gets replaced by a cheaper model. The ROI metrics for automation tell you whether the agent works. These controls tell you what it cost to find out.