On 30 September, AdExchanger published a headline that belongs on an agency owner’s desk rather than in the ad-tech trade press: insurers are building ad tech’s AI accountability layer. Not a vendor, not a regulator. The insurance industry — the party that eventually prices your mistakes — has started building the oversight machinery for AI in advertising.
The figures behind it come from a Gartner survey conducted in April. Approximately 60% of companies using AI will run into unexpected charges or exceed their budgets due to a lack of human oversight. More than half — 56% — implemented AI tools without clear policies in place.
Put those two numbers together and you have the shape of the problem. The market adopted the writing before it built the record of it. Now a stakeholder with money on the line is starting to ask for the record.
The stakeholder nobody planned for
Betty Louie, partner and general counsel at The Brandtech Group, said it plainly on stage at AdExchanger’s Programmatic IO conference in New York: insurance underwriters are quickly becoming “a new player in the AI discussion.” She was joined by Nicholas Godlove, legal director of AI and global privacy data at Yum! Brands, and Alan Parker, chief innovation officer at BBDO North America.
Louie’s framing is worth reading twice. An insurance policy can add a layer of protection around AI use — and it also adds a layer of fragmentation, because every insurer brings its own standard. “Insurance companies will have their own points of view [about] policies and processes in place,” she said, including how active that human in the loop really needs to be.
That last clause is the whole ballgame. A human in the loop is now a vendor default line, and an underwriter is a party that gets to define what counts as one. When the definition arrives, it will arrive as a question: show me.
What “fully auditable” means, in their words
Louie’s argument is that AI products and workflows must be fully “auditable” — and then she defines it. This is the closest thing the market has to a specification, and it comes from a lawyer rather than a vendor.
“If an ad campaign misfires impressions or misses its revenue targets, an agency or tech company should be able to look back and easily find answers, such as which large-language model was activated when a mistake occurred, what data was generated and whether the mistake in question was a human or machine error.”
Strip the sentence down and it is a field list. Which model acted. What was generated or changed. Whether a person or a machine was responsible. And the whole thing has to be findable after the fact, by someone who is not reconstructing it from memory.
| The question | What it demands |
|---|---|
| Which model acted? | An identity for the AI system, attached to the change it produced. |
| What was generated? | The actual output — the value before and after, not just that something changed. |
| Human or machine? | An attribution that separates the person who decided from the system that executed. |
| Findable later | A record you can look back into months after the campaign, without the tool that made it. |
Notice what kind of question this is. None of them are performance questions. Not ROAS, not cost per acquisition, not why the CPM spiked. Every one of them is a provenance question, and provenance is not a metric a dashboard reports. It is a record a system keeps — or does not.
Even the brands are building it out of people
The organisations closest to the problem have not waited for a product to solve it. Alan Parker, chief innovation officer at BBDO North America, described a “delivery and maintenance team” staffed with specialists for particular products and use cases, and said the agency runs regular performance tracking for agentic workflows. Yum! Brands keeps a dedicated AI team overseeing the legal and security side of deploying AI tools, tying its brand portfolios to a single AI use policy.
Nicholas Godlove, Yum!’s legal director of AI and global privacy data, was blunt about why the headcount exists: it is “extremely difficult” to implement policies at the company level given the “myriad types of AI technology and use cases infiltrating the ad industry,” from AI-generated creative through to agentic media buying. “The rules we’ve built over the last two to three years have been increasingly restrictive,” he said.
Read that as a signal, not a case study. A global brand answers uncertainty about AI oversight by building a function and staffing it. An agency with twelve people does not have that option — so it reaches for whatever the tool vendors put in front of it. This quarter, that was a gate.
The market’s answer was a gate. A gate is not a record.
Give the vendor layer its due. This quarter the MCP tooling market got the human-in-the-loop half right, and it did it fast. Supermetrics switches write access on per ad account and creates new campaigns paused. Windsor’s writes “require the end user’s explicit approval.” Toffu’s DV360 MCP leads with “Every write waits for your approval.” Synter advertises a “Human Spend Gate on Every Mutation,” and Markifact, a newer hosted ad MCP vendor, sells “approval on every write.”
| Vendor | The gate, in their own words |
|---|---|
| Supermetrics | “Decide when a person has to sign off: never, before a campaign goes live, before budget changes and launches, or before any change to a live campaign.” |
| Windsor | “requires the end user’s explicit approval” |
| Toffu | “Every write waits for your approval.” |
| Synter | “Human Spend Gate on Every Mutation” |
| Markifact | “Hosted, one OAuth login, approval on every write.” |
Five vendors, one phrase, and the gate itself has stopped being the differentiator. So ask the second-order question: when the gate opens, what is left behind?
For the one vendor that publicly keeps a change history, the answer is its own sentence: “Every change is logged in Supermetrics against who made it.” Read it carefully. The log lives inside Supermetrics — the platform that made the change. It names who made it, the credential that fired the write, which is not the same as the person who approved it. And it names neither the jurisdiction nor the client engagement, nor offers a way to hand the record to someone without your login.
The five questions a gate does not answer
These are not hypothetical. They are the questions an underwriter, a supervisor or a client’s compliance officer will ask months after the campaign, when nobody remembers the thread.
- Who approved it? A named person who accepted the decision — not the service credential that executed it.
- What was the value before and after? “A budget changed” is not the same as the number it changed from and to.
- For which client engagement? One agency runs many clients; the record has to say whose account it was.
- Under which jurisdiction? Promotion rules are national. The same change for a UK-facing brand is a different question than for a German one.
- Can you export it without logging in? If the only copy lives inside the platform that made the change, that platform is also the arbiter of it.
The law already knows how to write a record duty — for banks
One detail in the regulatory landscape sharpens the point. The EU AI Act’s Article 26(6) requires deployers of high-risk AI systems to keep logs “for a period appropriate to the intended purpose of the high-risk AI system, of at least six months.” Then it spends a second paragraph on one industry in particular: deployers that are financial institutions “shall maintain the logs as part of the documentation kept pursuant to the relevant Union financial service law.”
That gap is the same one the insurance market is now pricing. The regulator wrote the duty for the sector it could see. The insurer is writing the question for the layer it can’t.
Where Ott fits
Ott’s Activity Logging records the four fields a provenance question resolves to: the actor, the value before and after, the timestamp, and the linked object. It is stored independently of the ad platform, so the answer survives a platform change, an account ban or a reporting outage, and it stays readable without a login to the tool that made the change. An MCP connection means an AI assistant working on your client’s accounts reads through the same structure rather than beside it, and workspaces carry the client and brand hierarchy, so a change is filed against the engagement it was made for.
Three limits, stated plainly, because we hold vendors to this standard and should hold ourselves to it. Ott is read-only — it records what happens in the accounts and in your operations; it does not write to ad platforms. On coverage, Meta is connected today with Google Ads on the roadmap, so an agency running four platforms gets the record for the surface it uses most. And we publish no retention number, which is the same gap we criticise above; it is a limit, not a hidden detail.
Scope matters more than volume for a record, because the wrong meter is the wrong incentive. Ott prices by brands and seats, with no per ad account, per data source or per Telegram join fee, so adding accounts to a client does not change the cost of keeping the record. See how that is metered, or read the side-by-side comparison.
The same question, whoever asks it
Here is the useful thing about this quarter’s news. An underwriter asking for the AI accountability layer and a supervisor asking for the change history on one regulated campaign are asking the same question. Not did you have oversight — everyone has a gate now. They are asking can you produce the record: by a named approver, for a named client, under a named jurisdiction, from a system that is not the one that made the change.
A gate closes and vanishes. What remains is the thing your insurer, your regulator and your client will all eventually ask for. It is cheaper to build it before they do.