What we have learned running AI agents.

1. 01

   AI output is cheap. Checking it is the job.

   AI makes drafts, ratings, and replies almost free to produce. When nobody checks them, the cost lands on the reader: the client, the candidate, the manager who has to redo it. As output gets cheaper, a short answer someone can defend is worth more than a long one nobody read.

   What we do about it

   • Every agent workflow gets a written standard for good output, with real pass and fail examples, before it goes live.
   • A person checks a sample on a set schedule. A drop in the pass rate is logged as an incident, with a cause and a fix.
   • Client reports are short, and a person can defend every number in them.

2. 02

   Finding the right workflow is the hard part.

   The tools are rarely the slow part. The slow part is finding which workflows pay, ranking them, and mapping how the work actually happens. The people closest to the work are too busy to map it, and how they describe it is often not how it is done.

   What we do about it

   • We map the work from its records and by watching it, not only from interviews.
   • We rank workflows by payback and readiness, then ship one before proposing a program.
   • Each workflow gets a named owner on the client side and a review rhythm, because behavior change takes longer than the build.

3. 03

   You can build it. Running it is the work.

   Teams with engineers can build an agent quickly, and often should. What stalls is everything after launch: checking output, working the escalations, changing instructions when the model or the data changes, and deciding when the agent may act alone. Nobody owns that by default.

   What we do about it

   • We sell the running and the accountability, not the build. Our fee pays for judgment and a monthly record.
   • Every agent has a level: it reads, it drafts for a person to approve, or it acts inside agreed limits. It earns more room only when the numbers show it.
   • If a client already runs their agents well, we say so and step back.

4. 04

   Adoption follows the leader's own use.

   Many leaders ask their teams to adopt AI without using it day to day themselves. Without hands on use, it is hard to judge what is possible, what is hype, and what a fair result looks like. Teams notice.

   What we do about it

   • In discovery we ask the sponsor what they used AI for this week, and plan around the answer.
   • The sponsor gets one AI task of their own in the first week, tied to their real work.
   • The sponsor's name goes on the monthly review, not only the workflow owner's.

5. 05

   Trust comes from reports that show their misses.

   It is hard to tell what in AI is real, and most buyers do not follow it full time. They need someone to filter it for them, and they will only trust that filter if the numbers hold up. A monthly report with no misses in it is the one to doubt.

   What we do about it

   • We never claim a result we have not measured. Unmeasured stays "not measured", never zero.
   • Every monthly record lists the incidents and how each was fixed.
   • We say no to tools and agents that do not fit, even when a client is excited about them.

Some of these notes began as reading and conversations across the AI consulting field. The practices and the examples are our own.