An AI agent can stay inside its budget and still spend too much.
It can buy on signals from known sources. It can act within the authority it was given. It can log every decision. But the evidence can still arrive after the money is committed.
In three previous pieces, I argued that agents need richer signals than segments, feedback fast enough to reprice the next impression and standards that establish authority rather than parse data. Even if we grant all three, there is still a hole in the plan, and it has a dollar figure attached.
The proof gap
Suppose a brand hands an agent $200,000 and the agent moves $1,000 an hour into a new mix of CTV and retail media.
The first sales readout fit to judge that change would take 48 hours to mature. By then, $48,000 has been committed to a decision nobody has assessed, and that first readout will not settle all $48,000 either. It will produce an estimate, with error bars, about a change that is already two days old.
Call it the proof gap: spend committed while the evidence needed to assess a change is still pending.
The proof gap is a line most media plans leave implicit. The campaign budget caps total spend. Pacing sets the rate. Neither, by itself, limits how much can ride on a single untested change while the test is open.
Faster measurement narrows the gap but never closes it. A model updates its estimate in milliseconds, but a business outcome takes weeks to become real. And in an agentic system, the gap compounds. An agent changes bids, audiences and channels mix at once, so the same weak guess ends up driving spend across several budgets.
Who holds the gap?
This industry has fought over the space between commitment and settlement before. In the early 1990s, the fight was about sequential liability: whether an agency owed the publisher before the brand had paid the agency.
Sequential liability was a fight about credit risk. Who eats the loss if the money never arrives?
The agentic version is about proof risk. Who eats the loss if the outcome never arrives?
Markets that trade ahead of proof lean on two tools to keep everyone accountable: a risk budget that limits what you commit while you learn and a contract that assigns a defined loss to someone else when a defined part of the test goes wrong. These are different instruments, and the mistake this industry is about to make is treating them as one.
The risk budget
Alongside the campaign budget, the brand should set a risk budget: the most an AI agent may commit under an unvalidated change before it must show the agreed evidence or obtain fresh approval. An agent might be granted $12,000 to test a new mix and nothing more until the readout clears a threshold both sides set in advance.
Three rules keep this limitation honest.
1. All parties agree to the evidence standard before the first bid: what counts as progress, which method measures it and what parameters need to be met before expanding the agent’s scope. A pixel records an event. A sound causal test estimates how much the ads changed outcomes, with stated uncertainty. The agent’s own confidence score is not independent validation.
2. The clock must match the goal. A car, a mortgage and a brand campaign do not prove themselves in 48 hours, and a short-term sales gate should not be the default for work designed to pay back over months. And when evidence is late, the default is to pause, not to grant fresh funds for an expansion.
3. The risk budget has to sit outside the agent’s objective. A ceiling the agent is rewarded for reaching stops being a ceiling. The agent cannot raise the cap, and it cannot reset the count by moving spend into a new channel or handing it to a second agent.
The underwriter
While a risk budget limits what the buyer commits while it learns, a guarantee assigns a defined financial loss to a counterparty when an agreed condition is met.
For that guarantee to hold, the contract has to say what triggers payment, how much is owed, what is excluded and who funds the promise.
A brand can halt new spend, though it cannot recover money already committed. A publisher still needs payment for the media it has delivered. An intermediary may have bought impressions that fail to earn back their cost. Those risks require separate terms. The seller may buy protection to support the promise it makes to the brand. But the buyer’s test allowance still stays in force.
Advertising has one working precedent for this approach: the make-good. If the seller delivers fewer impressions than expected, the cure is more inventory. But the edges of the make-good are now being probed as AI takes over.
Munich Re has backed AI vendors’ performance guarantees with a reinsurer’s balance sheet since 2018. TikTok offers eligible GMV Max campaigns ad credits when ROI falls below 90% of the daily target. The first is third-party capital paid in cash. The second is the seller’s own count, paid in credits, behind a long list of exclusions, including the day a campaign is paused.
Now hold everything else constant. Two sellers pitch the same outcome deal at the same CPA, counted by the same agreed method. One adds a funded, capped cash remedy for a defined shortfall at a stated fee. The test allowance does not move; the buyer needs the same evidence either way. What moves is the buyer’s retained loss if the miss happens.
Coverage does not validate the strategy. It changes who eats the miss. And the fee for that miss is the price of the specified risk being transferred.
A vendor that is paid a share of spend earns less every time its agent pauses, which requires it to prize restraint while paying it for volume. The agentic intermediary can earn a spread for standing between two parties when the thing they agreed on does not happen. Software can price the risk. A funded counterparty still has to bear the loss.
My prediction: By the end of 2027, at least one media deal between independent firms will settle from escrow. Its terms will name the source used to count results and provide a cash remedy for a counting error. And the test will be an executed deal with terms that can be verified.
Closing the loop
As the use of AI agents expands, we have to keep in mind that, although AI can interpret data, it cannot vouch for it. But neither can capital.
A test estimates what the spend did. A contract says who answers if it didn’t. A balance sheet says whether that answer is worth anything.
The standard of proof belongs in the test. The promise to pay belongs in the contract. The cash behind the promise belongs on someone’s balance sheet.
The right to spend more should be earned. Any promise to cover the miss should be priced, bounded and backed by an underwriter.
“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.
Follow Evgeny Popov and AdExchanger on LinkedIn.
For more articles featuring Evgeny Popov, click here.
