Adam Torkildson
Agent-run marketing campaigns tend to fail without making a sound, and the damage almost never shows up in a dashboard until the quarter is already closed. That's the uncomfortable part. The bids kept adjusting, the emails kept sending, the reports kept generating a green check, and somewhere in the middle a body of work drifted off its goal and nobody caught it.
This isn't a story about a bad model or a rogue prompt. It's about what happens when execution stops being something a human hand-assembles and becomes something a system decides on its own, and how marketing teams are learning, sometimes expensively, that the old ways of watching a campaign don't watch this kind of work at all.
The Failure Wears a Green Check
The classic software failure is loud. Something breaks, an error fires, a pager goes off, someone fixes it. Agent failures behave almost nothing like that. An agent will complete its workflow, return a plausible output, log a success, and be wrong in a way that only becomes visible when the money doesn't show up.
Dataiku's engineering team has written about this bluntly: once agents start acting on their own instead of assisting people, wrong decisions accumulate across thousands of successful transactions and tend to surface through customers, auditors, or regulators rather than through monitoring. That's the core diagnostic problem. The green check is answering the wrong question — did the task run? — when what you actually need to know is whether the task did the right thing.
Underneath, the mechanics are ugly. Goals drift, context gets truncated, and an agent told to optimize for qualified pipeline starts optimizing for the nearest cheap proxy: form fills, low-cost clicks, a segment that converts to nothing.
For a longer treatment of how teams are rewiring their week around this — which roles change, which meetings disappear, and where humans still have to decide — TechGape on what happens when the agent runs the campaign is a useful read.
Why the Obvious Fix Falls Short
The instinct, when leadership finally sees the miss, is to bolt more oversight onto the front of the process. Tighter prompts. Stricter guardrails. A weekly review of the agent's outputs.
It rarely holds up, for three reasons worth naming plainly.
- Prompt tuning treats a system problem as a language problem. The failure is usually not in what the agent was told; it's in what happened five steps into the loop, when the plan met a real bidding auction or a real inbox. Rewriting the brief doesn't change the drift.
- Sampling misses the pattern. Reviewing a handful of the week's outputs feels responsible and catches almost nothing. The subtle errors look right on inspection, and they surface as variability across many runs rather than as any single obvious mistake.
- The dashboard is the wrong instrument. Traditional performance reports were built to watch human-paced work. They summarize what shipped, not whether the reasoning that produced it held together, and they don't flag when a campaign is drifting toward the wrong outcome.
There's a compounding factor here too. A lot of what agents do sits close to legal exposure — claims in ad copy, targeting decisions, disclosures — and reviewers who used to see every asset before it went live are now signing off on process rather than pieces. A recent practitioner guide from Davis+Gilbert lays out how easily autonomous execution can push claims into market without a human ever reading them.
What Actually Works Looks Like Instrumentation, Not Supervision
The teams getting this right have stopped trying to watch the agent's outputs and started watching the agent's reasoning. You instrument the loop — the goal it was given, the data it pulled, the action it chose, the result it observed, the adjustment it made — and you make that trace legible to a human on a schedule that isn't quarterly.
In practice, four moves separate teams that catch the misfire early from teams that catch it in the QBR:
- Define what "good" means in the loop, not in the report. Write down the intermediate signals — audience quality, creative fatigue, funnel-stage conversion — the agent has to protect on the way to the headline metric. If the only definition of success is revenue, drift stays invisible until revenue moves.
- Log every decision, not every output. You want the trace: what the agent saw, what it chose, what it changed. That's what makes a subtle failure inspectable weeks later, when someone finally asks why paid social started chasing a segment nobody would have targeted on purpose.
- Name the human who owns the sign-off. Not "the team." A person, by role, who is accountable for the class of decision the agent is making. Autonomous execution needs a named brake, or the brake is nobody's job.
- Review the pattern, not the sample. Look at distributions of decisions over hundreds of runs. That's where the drift lives. A weekly spot-check of ten ads will keep missing it.
The teams that will spend the next few quarters cleaning up silent misfires are the ones treating agents like faster humans. The teams that won't are the ones treating them like systems: instrumented, inspectable, and owned by a person with a name.
