Book our Advanced Audit
Built Honest. 7 min read

When the agent spends the money, who writes it down?

AI agents will make buying faster. Not more accountable, unless someone builds the ledger.

Shashidhar Sharma
Shashidhar SharmaCofounder and Chief Operating Officer, AdNexa.ai. 16+ years in media.

My position first, then the argument.

AI agents will make media buying faster. They will not make it more accountable unless somebody builds the ledger, and right now nobody is. The ledger is a boring, five-part record that has nothing to do with intelligence and everything to do with whether a CFO can trust the invoice. Here is what it contains, and why the speed claims skip it.

What is real

IAB Tech Lab runs an Agent Registry, and on 16 September 2026 it opened a "Programmatic Standard Practices" rulebook for public comment until 16 October (PR Newswire). One of the competing agent protocols merged code for "reconciled billing" receipts into its public repository on 18 September, which tells you reconciliation was not there before. The largest independent buying platform shipped an update at the end of August that the trade press described as "AI easy buttons", not autonomous buying (AdExchanger, 27 August 2026). The head of IAB Tech Lab has spent the year asking people what an agentic buy actually is, and reports that nobody can tell him.

Meanwhile, products are arriving with very precise speed claims. A campaign lifecycle cut from 30 days to 3. Setup from 72 hours to 30 minutes. Up to 99% less manual effort. I do not doubt the numbers. Setup was never the hard part. I could set up a campaign in 30 minutes fifteen years ago, if nobody checked it.

The two questions the speed claims skip

The first is authority. A registry tells you who an agent is. It does not tell you what that agent may do: how much it can commit before a person signs, on which accounts, with what revert path if the spend is wrong. Identity is not authority.

If that sounds abstract, here is the concrete version. Last week a nine-year-old ran up US$118,000 in three weeks on YouTube ads for his gaming videos, on his father's company card, after signing up for a US$20 promotion (Tom's Hardware, via AdExchanger, 17 September 2026). Nobody in that story was an AI. The card had authority. The child did not. Nothing in the system checked the difference. Replace the child with an agent and the card with an API key and you have next year's incident report.

The second question is the ledger. When an agent moves money at 2am, who writes down that it happened, in a form the CFO can read, the auditor can pull and the client can dispute? Your bank does this for a card tap at 2am without being asked. Media buying does not, even with humans placing the orders.

How much goes missing with humans in the loop

When ISBA and PwC traced a pound of advertiser money through the UK's open-web programmatic chain, publishers received roughly 51p. Known intermediaries kept about 34p. About 15p could not be traced to any named party at all. The study called it the "unknown delta". The UK competition regulator's more careful estimate still put the total take at about 35p.

Of every £1 spent on open-web programmatic, 15p could not be traced51pReached the publisher34pKnown fees in the chain15pNobody could say where
Source: ISBA (Incorporated Society of British Advertisers) and PwC, Programmatic Supply Chain Transparency Study. The UK competition regulator's more careful read still put the total take at about 35p.

Now speed the chain up by a thousand and remove the humans. The unknown delta does not shrink. It compounds.

The five records

A reconciliation ledger for agent-driven buying needs five linked records. Every one is append-only. Nothing gets edited. Corrections are new rows.

  1. The plan. Budget, split, targets and dates the client signed, with a timestamp. A mid-flight change is a new row with a name against it.

  2. The order. What the agent asked for, from whom, at what price, under which approval. Which human, and what limit they were working within.

  3. The change log. Every setting the agent touched on every platform, with before and after values. Platforms already keep this. Almost nobody exports it.

  4. The delivery. What actually ran, from the platform's own reporting, matched line by line to the order.

  5. The invoice. What was billed, matched to the delivery, with a discrepancy tolerance agreed in advance and a rule for what happens above it.

If a vendor cannot show you all five, linked, for one campaign, the agent is fast, but it is not governed.

The workaround you can build this month

You do not have to wait for a standard. Export the change history from every platform you buy on (Google Ads, Meta and DV360 all have one) into a table nobody can edit, once a day. Add the plan as row zero. Add the invoice when it arrives. That is a version-zero ledger, it costs nothing, and it will answer more CFO questions than any dashboard you own.

Where we stand

Our software cannot move a rupee, a dollar or a dirham. It has read-only access to every ad account it watches. It runs the checks, computes the recommendation, names the exact platform setting to change, and a human makes the change. The judgment is ours. The click is yours.

We chose this on purpose. It costs us the "99% less effort" headline. It buys the client every recommendation logged, every human action against it logged, every before-and-after measured by the same maths, and an audit trail nothing can rewrite.

The takeaway: speed is a feature. Accountability is a record. If you buy media at scale, file a comment on the IAB Tech Lab rulebook before 16 October and ask for the five records above. Otherwise version one of the rules for agents will be written by the people who were happy with version zero.

Sources

PR Newswire, "IAB Tech Lab's Programmatic Standard Practices available for public comment", September 2026: https://www.prnewswire.com/news-releases/iab-tech-labs-programmatic-standard-practices-available-for-public-comment-302879958.html

AdCP client repository, reconciled billing merged 18 September 2026: https://github.com/adcontextprotocol/adcp-client/pull/2949

AdExchanger, "The Trade Desk's Zuma update adds a host of AI-powered easy buttons", 27 August 2026: https://www.adexchanger.com/technology/the-trade-desks-zuma-update-adds-a-host-of-ai-powered-easy-buttons-to-its-kokai-platform/

ADOTAT, "The Ledger Wars", 29 July 2026, and "Everyone is selling AI media buying. AdCP and IAB Tech Lab say it isn't running yet", 26 August 2026: https://new.adotat.com/

AdExchanger daily, 17 September 2026 (Tom's Hardware item on the US$118,000 YouTube ad spend)

ISBA / PwC Programmatic Supply Chain Transparency Study: https://www.isba.org.uk/knowledge/second-programmatic-supply-chain-transparency-study

Further reading

Book our Advanced Audit