# Everyone's safety answer is 'a human'. Nobody says which one.

Three agencies launched agentic media systems in a single Advertising Week. Asked how clients are protected, every executive said a human. None named the human, the threshold or the log.

Published by Fiction Tribe, 2026-10-09. Canonical: https://velocity.fictiontribe.com/field-notes/a-human-is-not-a-safety-plan

Key figure: <7% of enterprise work is performed autonomously by AI, per ISG.

## The short version

- Three agentic media systems launched in one Advertising Week. Asked how clients are protected, every executive gave the same answer: a human.
- None published a threshold, an escalation path, a named accountable role or an audit.
- ISG puts autonomous work at under 7% of the enterprise total. IAB Europe says 78% have named an AI governance owner. Both sources have an interest, set out below.
- Write the oversight down on one page: who approves what, above which amount, and who answers when it goes wrong.

"Human in the loop" is the most reassuring phrase in advertising right now. Ask which human, in which loop, with what authority, and the answer gets vague quickly.

In a single Advertising Week, three agentic media systems went to market. Dentsu's Carat launched a planning, buying and optimisation agent with NBCUniversal, Freewheel and the AI firm Newton. Tinuiti opened its Bliss Point measurement layer to agents as an MCP server. Stagwell relaunched Assembly as Stagwell Media around an agentic system called Machine OS.

Digiday reported how the executives answered the obvious question: how are clients protected? Every one of them answered with a human.

## How are agencies keeping AI media agents safe?

With people, by their own account, and with very little on paper. Freewheel's chief product officer David Dworin was the most candid about the agents: "they're new, and we don't trust them, and they make mistakes." WPP Media's Lauren Wetzel described explicit controls, approval thresholds and a clear line between probabilistic AI recommendations and deterministic execution. Dentsu's Rebekah Shalit placed every task on a sliding scale from low to high human oversight.

All sensible. All adjectives. Three products shipped in five days with no published threshold, no disclosed escalation path, no named accountable role and no audit.

> "Human in the loop" sounds like a control and functions as an indemnity.

The phrase is doing the work "brand safe" did in 2018. One limit on the evidence: this is conference-stage testimony from the firms selling the systems, with exactly one client referenced and that one unnamed.

## How much work do AI agents actually do on their own?

Very little so far, which is why the human answer comes so easily. ISG's 2026 State of Enterprise AI report, released on 23 September, found less than 7% of work is performed autonomously by AI. Fifty-five percent of AI-enabled work is human-led. Nearly a quarter is human-reviewed. Almost 14% involves people only for exception handling.

> **&lt;7%** of enterprise work is performed autonomously by AI today. ISG, State of Enterprise AI, 2026

Enterprises told ISG they expect that to nearly double, to 13%, by the end of 2027. Read the source with care. ISG sells AI transformation advisory, so it has an interest in a gap it can help close. The 13% is a self-reported expectation, not a measurement. And the report blends a survey with observations from ISG's own client engagements, so it is not a clean probability sample.

Still, the direction is clear. Today a human is in the loop because the agent cannot do the job alone. That is circumstance, and circumstance is a poor substitute for design.

The money is already ahead of the operating model. Digiday's earlier reporting cites Gartner: CMOs are allocating 15.3% of marketing budgets to AI, while only 30% claim the capability to scale it. One agency executive told Digiday's Programmatic Marketing Summit that advertisers, agencies and partners are each building their own agentic workflows, with nothing governing the different approaches.

## Who is accountable when an AI agent gets a media buy wrong?

Usually someone with a title, and rarely anyone with a rulebook. IAB Europe's 2026 Impact of AI on Digital Advertising report, published on 24 September, says 78% of respondents have identified a formal owner for AI governance. It also says 58% expect agentic ad buying to reach operational use or scale within a year, that advertising-specific guidance is less widespread than general governance, and that lack of expertise or training is the most-cited barrier.

> **78%** have named a formal owner for AI governance. Marketing-specific guidance is less widespread. IAB Europe, via ExchangeWire, 2026

Discount these figures as well. They reach us through an ExchangeWire PressBox release, not independent reporting. The survey covers IAB Europe members and its own AI Working Group, a self-selected group with a stake in looking prepared. No sample size is disclosed.

Even so, naming an owner is the cheapest governance act there is. The expensive part is telling a media buyer what an agent may do with a budget. IAB Europe's Dimitris Beis asked the right question: who is accountable when something goes wrong?

## What should a written oversight plan contain?

One page. We would rather read that page than hear another panel on trust. Four entries.

**Approval thresholds.** Who approves what, above which amount. A name and a number, per decision type.

**Unattended scope.** What the agent may spend or change without a person. What it may never touch.

**The log.** What gets recorded, and who reviews it, on what schedule. An unread log is storage.

**The answerable person.** Who picks up the phone when it misfires, and what happens next.

An escalation map like this is cheap to produce and nearly impossible to fake. Ask your agency for theirs. Ask your own team for yours. Right now it is the only part of an agentic system a client can inspect.

A human is a job description. A safety plan has a name on it and a number beside it.

## Questions people ask

### How do agencies keep agentic media buying systems safe?

By their own account, with human oversight. When Dentsu's Carat, Tinuiti and Stagwell launched agentic media systems in one Advertising Week, every executive Digiday quoted answered the safety question with a human, but none published a threshold, an escalation path, a named accountable role or an audit.

### How much work do AI agents do autonomously today?

Under 7% of enterprise work, according to ISG's 2026 State of Enterprise AI report, with enterprises expecting 13% by the end of 2027. ISG sells AI transformation advisory and the 13% is a self-reported expectation, so treat both figures as indicative.

### What should an oversight plan for AI media agents include?

One page naming who approves what above which spend threshold, what the agent may do unattended, what it may never touch, what gets logged, who reviews the log, and who is answerable when it misfires. IAB Europe reports 78% of its respondents have named an AI governance owner, but says marketing-specific guidance is less widespread.

## Sources

1. [As agencies push agentic media future, execs work to ward off client rogue fears](https://digiday.com/media-buying/as-agencies-push-agentic-media-future-execs-work-to-ward-off-client-rogue-fears/), Digiday. Conference-stage testimony from the firms selling the systems. One client referenced, unnamed.
2. [Who owns agentic workflows? Agencies struggle to govern new tools as marketing budgets surge](https://digiday.com/marketing/who-owns-agentic-workflows-agencies-struggle-to-govern-new-tools-as-marketing-budgets-surge/), Digiday. Earlier reporting, including the Gartner budget figures.
3. [2026 State of Enterprise AI: The AI Value Gap](https://www.financialcontent.com/article/bizwire-2026-9-23-ai-is-changing-how-work-gets-done-but-business-value-still-lags-isg-study), ISG, via Business Wire, 2026-09-23. ISG sells AI transformation advisory. Blends a survey with observations from its own client engagements.
4. [State of Enterprise AI report 2026](https://isg-one.com/advisory/artificial-intelligence-advisory/state-of-enterprise-ai-report-2026), ISG, 2026-09-23. The report itself.
5. [Agentic Ad Buying Gains Ground in IAB Europe's 2026 AI Report](https://www.exchangewire.com/blog/2026/09/24/agentic-ad-buying-gains-ground-in-iab-europes-2026-ai-report/), ExchangeWire, 2026-09-24. A PressBox release, not independent reporting. No sample size disclosed.
6. [Impact of AI on Digital Advertising report 2026](https://iabeurope.eu/news-impact-of-ai-report-2026/), IAB Europe, 2026-09-24. Survey of IAB Europe members and its own AI Working Group, a self-selected population.
