Our AI pilot worked and nothing happened. How do we get it into operations?

A pilot reaches operations when someone answers for the result, the routine is written down, and the team knows what the agent does alone, what a person reviews, and which old step gets switched off. If one of those three is missing, the pilot stays a demo, however good the model is.

How do you govern AI agents in a company? Permissions, evals and owners

What should a mid-sized company's first AI project be?

Why do AI projects fail in companies?

How do you measure AI adoption in a company?

Why it stays a demo

The demo goes well. Everyone claps. Two months later it is still sitting in a folder, used by the person who built it.

It usually comes down to the same things:

  • Nobody takes it over. An innovation team or a vendor built the pilot, and the area that was supposed to use it never became responsible for it.
  • There is no routine. The agent answers well in the demo, but nobody defined what triggers it, which source it works from, what it can do, and when it has to stop.
  • Exceptions show up in production. The demo was the happy path. Operations bring the product that doesn't exist, the customer on another price list, the incomplete voice note.
  • Nobody decided what gets switched off. The old process stays the same and the pilot becomes one more step.

The test before you scale

Before taking a pilot into operations, answer five questions about the workflow:

  1. Source. Which information rules? One official source, no competing versions.
  2. Owner. Who answers for the result? If nobody does, what you have is a conversation.
  3. Criteria. What counts as a good result? Good and bad examples, written down.
  4. Jurisdiction. What can the agent do alone, what needs review, and what is off limits?
  5. Evidence. What gets recorded from each cycle so you can correct and learn?

If you can't answer them, the next step is to map the routine before you scale.

What a routine that actually runs looks like

Take an order that arrives on WhatsApp, sometimes as text, sometimes as a voice note, sometimes incomplete. "Let's have an agent read it and create the order" sounds fine, but first you need six pieces:

  • Trigger: the incoming message.
  • Source: the catalog and that customer's price list.
  • Action: preparing the order.
  • Check: required fields, products that exist, the customer's terms.
  • Memory: the recorded order and the corrections.
  • Brake: which price, credit or shipping exceptions go to a person.

With those pieces the agent has a job. Without them it has a long conversation.

What to do, in order

  1. Pick the workflow where the pilot already showed something. One.
  2. Name the owner inside the area that will use it, not in the team that built it.
  3. Gather real cases, exceptions included. About ten is enough to start.
  4. Write down the six pieces of the routine and the agent's limits.
  5. Build a test set from those cases and run it before every change.
  6. Agree which old step gets retired, and who decides on that day whether to retire it, fix it or roll back.
  7. Run it with the owner, learn, and only then move to the next workflow.

An example

A technical expert carried knowledge that sales, support and delivery needed more often than one person could handle. We installed a curated source base, an agent workflow for support and sales context, and an escalation path for when confidence or authority fell short. Technical commitments without review and answers outside the curated sources were left out on purpose. The expert could serve more people without giving up responsibility, and the company could see which documentation was missing. More in the Technical Knowledge Agent case.

When Business Cyborgs helps and when it doesn't

We do this in an Agent Sprint: one problem, one workflow, 2 to 4 weeks, projects from USD 5,000. We deliver the working agent or workflow, a short operating guide, the test set, lessons on sources and data, the backlog for the next iteration, and recommendations on permissions and owners.

It isn't for you if there is no internal owner, no real examples, no source material, or the goal is autonomy without limits. If the information isn't in shape, we start with preparation. If there is also no owner, workflow and cadence, we'd rather not start.

See the Agent Sprint · Talk to a Business Cyborg

FAQ

Why do AI pilots fail to reach production?

Because nobody in the area that will use it answers for the result, the routine isn't written down, and nobody decided which old step gets switched off. The model is rarely the problem.

How long does it take to move a pilot into operations?

For a bounded workflow, an Agent Sprint takes 2 to 4 weeks if there is an owner, real examples and sources.

Do I need more data or a better model?

Usually you need one official source, examples of a good result, and clear limits for the agent. That matters more than switching models.

Who should own it?

Someone from the area that lives the workflow and can decide to fix, continue or roll back. Not the team that built the demo.

Further reading (in Spanish): BC 44 and BC 48.