A buyer should be able to see, for every deal an AI buying agent touches, what the agent was given, which mandate version applied, which options it considered and rejected, every offer and counter-offer, each escalation to a person, who approved what, and the final terms. That is the decision log, and it answers "why did the agent do this?" It does not answer "what did both parties agree?", which needs a separate record that the seller holds too.

TL;DR. A useful decision log for an AI buying agent records seven things: inputs, mandate version, options considered, offers and counter-offers, escalations, approvals and final terms. Each entry should be timestamped and tied to the mandate version in force. A decision log is one party's account of its own conduct, kept on its own systems, so it cannot settle a disagreement with the seller. Settlement depends on an agreed record of the terms, written at the point of agreement and held identically by buyer and seller.

What is an AI buying agent's decision log?

An AI buying agent's decision log is the buyer-side record of what the agent saw, what it was permitted to do, what it did and why. It is the agent equivalent of a trader's notes plus the approval emails from the trading desk, assembled automatically and in order.

The decision log serves three readers: the budget owner checking that the agent stayed inside its mandate, the person handling an escalation, and an auditor establishing months later whether a commitment was authorised.

Which seven items should a decision log record?

A decision log that serves all three readers records the following for each deal.

Inputs. The brief, budget, audience and brand safety rules, and any data the agent drew on, each identified by version or source.
Mandate version. The identifier of the mandate in force when each decision was made, plus who approved that version and when.
Options considered. The inventory or proposals the agent evaluated, including the ones it rejected, with the rule or score that decided each. A log that shows only the winner hides whether the agent searched properly.
Offers and counter-offers. Every proposal sent and received, in sequence, with price, volume, dates, measurement terms and the counterparty agent's identity.
Escalations. Each point at which the agent stopped and asked a person, the threshold or rule that triggered it, who received it and how long it waited.
Approvals. Who approved, exactly what they approved, any conditions they added and when. "Approved by the system" is not an approval.
Final terms. The terms the agent believes were agreed, with a reference to where the agreed record of those terms is held.

Every entry also needs a timestamp and a link to the mandate version, and the buyer should be able to export the full log without depending on the vendor's interface.

Why does the mandate version matter so much?

The mandate version matters because "the agent exceeded its authority" can only be proved against a specific version of that authority. If a budget owner raises a price ceiling on a Tuesday and the agent closes a deal on the Wednesday at the new ceiling, the deal is within mandate. If the decision log cannot show which mandate version applied at the moment of commitment, nobody can tell an authorised deal from an overspend.

The practical test is quick: pick any commitment in the log and ask which mandate version authorised it and who signed that version off. If the answer takes more than one lookup, the decision log is not doing its job.

What should a decision log not be expected to show?

A decision log should not be expected to explain a language model's internal reasoning. A model-written paragraph explaining why it chose something can sound plausible without reflecting what actually drove the choice. The useful record is the rule, threshold, score or instruction that produced each decision, together with the inputs it acted on. That can be checked. A generated justification cannot.

A decision log also holds material the buyer will not want the seller to see, such as the maximum authorised price and the walk-away point. That is one reason a decision log cannot double as the shared record of the deal.

What is the difference between a decision log and an agreed record?

A decision log is unilateral: one party writes it, on its own systems, about its own conduct. An agreed record is bilateral: it captures the terms both parties accepted, is written at the moment of agreement and is held in the same form by both.

The difference shows up the first time delivery is disputed. Take a hypothetical deal in which the buyer's log records a viewability floor as part of the final terms, while the seller agent's own log records a counter-offer that removed it. Both logs can be honest, and neither can settle the question, because each is evidence from one side.

The Agree. Transact. Verify. simulation, which modelled 90,202 deals, found that buyer and seller records kept separately disagreed in 95.3% of cases, against 0.19% when both parties worked from one shared record.

How do AAMP and AdCP relate to decision logs?

AAMP, IAB Tech Lab's Agentic Advertising Management Protocols, includes a reference buyer agent covering demand-side decision workflows, and the IAB Tech Lab Agent Registry lets a buyer check which agents a company has registered. Neither removes the need to ask a vendor what its own agent logs, and for how long.

AdCP, the separate Ad Context Protocol published by AgenticAdvertising.org, includes get_plan_audit_logs, which returns the decision history. Its documentation also states that settlement, disputes and makegoods remain account or platform integrations unless a specific contract is present: the same boundary as the one between a decision log and an agreed record.

What should buyers ask a vendor about decision logs?

Which of the seven items does the log record, and which does it leave out?
Is every decision tied to a mandate version, and does each version show who approved it?
Are rejected options logged, or only the chosen one?
Can the buyer export the full log without the vendor's help, and in what format?
How long is the log retained, and can entries be changed after they are written?
Where is the agreed record of final terms held, and does the seller hold the same version?

The last question is the one a decision log cannot answer about itself. Alkimi, the neutral agentic advertising marketplace that published the Agree. Transact. Verify. simulation, works within IAB Tech Lab's AAMP framework and puts the agreed terms on a DealSheet: written at agreement, timestamped and retrievable by buyer and seller independently, with people keeping approval rights at the deal level.

Frequently asked questions

What is an AI buying agent's decision log?

The buyer-side record of an agent's inputs, mandate version, options considered, offers and counter-offers, escalations, approvals and the terms it believes were agreed.

Is a decision log the same as a contract or deal record?

No. A decision log is written by one party about its own conduct. An agreed deal record captures terms both parties accepted and is held in the same form by buyer and seller, which is what settlement and disputes depend on.

Should a decision log include the model's reasoning?

It should include the rule, threshold, score or instruction behind each decision and the inputs used. A generated narrative explanation is not a reliable account of what drove a model's choice.

Can the seller see the buyer's decision log?

Normally not, and it should not need to. The log contains confidential information such as the maximum authorised price. What the seller needs is the shared record of the agreed terms.

What Should Buyers Be Able to See in an AI Buying Agent's Decision Log?