← Research library

Seven patterns we keep returning to in agentic marketing pilots

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Agentic marketing pilots are easy to start and hard to keep. Halfway through the year, the question we keep returning to is why some pilots grow into programs while others produce a good demo and then quietly stop.

This post collects seven patterns from our research and prototypes. They are themes we track and expectations we are testing, not measured findings, and we expect to revise them. We group them by what they concern: scope, inputs, people, and evidence.

Scope: start boring and stay narrow

Pattern 1: the pilots we would bet on start boring. We expect lasting pilots to begin with unglamorous work: research briefs, reporting, data cleanup, and first drafts. These tasks have clear inputs, clear outputs, and a human who already knows what good looks like. That makes them easy to review and easy to measure.

Ambitious starts carry a different risk. A pilot that begins with an autonomous campaign can produce an impressive demo. Then the hard questions arrive: who reviews the output, what happens when the agent is wrong, and who signs off on spend. If those questions have no answers, the pilot stalls while they are debated.

Pattern 2: narrow agents outlast general ones. An agent with one clear job is easier to test, improve, and trust. An agent meant to "help with marketing" is hard to evaluate, because no single rubric fits everything it does. We expect general agents to be used inconsistently, with each person finding a different use and no one able to say whether it works.

Consider a hypothetical team choosing between two pilots. One drafts weekly performance summaries from a fixed report. The other answers any question the team asks about its campaigns. The first can be checked against the report in minutes.

The second needs a new judgment for every answer, so its quality stays a matter of opinion. Our impact and risk matrix is one way to make this choice deliberately.

Inputs: context beats cleverness

Pattern 3: better context fixes more than better instructions. When agent output is weak, the tempting fix is a more elaborate prompt. Our working view is that the cause is more often missing or poor context: inaccurate product facts, a vague brief, or a voice guide that describes the brand without showing it.

The mechanism is simple. An agent can only be as accurate as the facts it reads. If the pricing page in its knowledge base is two versions out of date, no instruction will make its copy correct. Rewriting the prompt treats the symptom, and the error returns the next time the agent reaches for the same fact.

A pattern worth testing is to log, for each rejected output, whether the cause was the instruction or the input. If inputs dominate, the next hour of work belongs in the knowledge base. We describe how to build one in building a marketing knowledge base agents can trust.

People: ownership and review capacity

Pattern 4: someone has to own it. We expect pilots without a named owner to fade. A pilot with an owner who reviews outputs weekly, fixes problems, and reports results has a reason to keep improving. Our view is that ownership matters more than the choice of model or tool.

Ownership has a specific shape. The owner decides what the agent is for, notices when it drifts, and makes the call to expand or stop. Without that person, problems get noticed by whoever happens to be affected, and no one connects them into a pattern.

Pattern 5: review capacity is the hidden limit. Agents produce output faster than people can review it with care. A pilot that plans for review, with tiers by risk, batching of similar items, and sampling of repeat formats, can grow without breaking. A pilot that does not plan for review tends to end in one of two places: a bottleneck where work waits for days, or a rubber stamp where approval means nothing.

Consider a hypothetical pilot where one reviewer has four hours a week and each draft takes 15 minutes to review. That reviewer can handle 16 drafts a week. If the agent produces 40, the gap has to be closed by design, not by asking the reviewer to read faster. We cover the design options in approval queues that don't become bottlenecks.

Evidence: measurement and earned trust

Pattern 6: measurement has to be designed in. A team that decides what to measure before launch can answer whether the pilot is working. A team that tries to measure afterward usually cannot, because the baseline is gone. Once the old process changes, there is nothing clean to compare against.

The measures do not need to be elaborate. Time per task before and after, the share of outputs approved without major edits, and the number of errors caught in review cover a lot of ground. What matters is recording them from the first week, while the old way of working is still fresh.

Pattern 7: trust is earned in steps. We expect teams that move agents gradually up the autonomy ladder, with evidence at each step, to have fewer serious incidents. A team that jumps straight to high autonomy is more likely to step back after a visible mistake. That costs time, and it also costs internal confidence, which is slower to rebuild than the agent.

The step that matters most is the first move away from full review. We recommend tying it to a specific threshold, such as a run of approved outputs with no major edits, set in advance by the owner. A threshold agreed before the data arrives is harder to argue away after.

A quick check for your own pilots

The seven patterns translate into a short set of checks. We suggest running them on each active pilot before adding scope.

  • Scope: the pilot does one job with clear inputs and outputs.
  • Context: the facts the agent reads are current, and someone owns keeping them that way.
  • Owner: one named person reviews the pilot weekly and can stop it.
  • Review plan: expected output volume fits within reviewer hours, or there is a plan for the gap.
  • Baseline: you recorded how the work was done before the agent arrived.
  • Autonomy rule: the evidence needed to reduce review was written down before launch.

A pilot that fails one check is not doomed. It has a specific gap to close, and closing it is usually cheaper than adding features.

What this means for the second half

None of these patterns are surprising in hindsight. They mirror how good teams have long adopted new tools and processes. The difference we expect with agents is speed: both gains and mistakes compound faster, because an agent repeats whatever it does at volume.

That makes the discipline behind these patterns more important as programs grow. A weak fact in one draft is a small problem. The same fact repeated across hundreds of outputs is a brand problem.

The open question we are studying next is how these patterns change when pilots become programs with several agents. Ownership, review, and measurement all get harder when work passes between agents, and we expect new patterns to appear at the handoffs.

How to apply this: check each current pilot against the seven patterns. For any pilot missing an owner, a measurement plan, or a review design, fix that before adding scope.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.