First BellEric Bakken · AI consulting

← Training · By Eric Bakken · Updated Nov 2026

AI training

Structure by task: AI tools that carry the standard

Short answer: For repeated work that leaves the building, put the standard inside the AI tool: the role, the sources, the template and the limits. People can ship without waiting on the owner, and the review pile stays smaller. This is my experience in my own companies, not a study.

Fresh as of 12 Nov 2026. This may be outdated in 3 to 4 months, or sooner.

The review queue problem

I found this in my own companies, not in any study. Give people an open-ended AI tool and you do not control the work product. If the work goes out to customers or the public, you do not control your design standards.

The tool makes output cheap, so people make more of it, faster. You end up with a review queue you have to keep watching. Someone has to decide what ships and what goes back.

When open-ended is fine, and when it is not

Open-ended AI is fine for thinking out loud: a first draft for yourself, a plan, learning something new, untangling a problem. That work is private, and if it is wrong, you own the standard.

Add structure when the work repeats, when it leaves the building, when it touches money, customers, legal or brand, or when many people do it. It is the task that decides, not anyone’s rank.

Open-ended or structured? It depends on the task
Open-ended is fine when… Add structure when…
The work is private thinking or exploring The work is repeated
It is a first draft for yourself It leaves the building
You are learning something new It touches money, customers, legal or brand
You own the standard for the output Many people do the same work

What a structured tool is

People in a company have roles and expectations, and the tool should too. Build the guidance and limits into it, and write them down: the job it does, where its facts come from, the shape of the finished piece, what it must never do, and where it stops and hands off to a person.

Then people can ship their work without waiting on the owner. Review gets smaller and sharper, but it does not disappear. A person still gives the okay. That is the trust ladder, and a tool climbs it as it earns it.

The role card

A role card is one page. This is the shape.

A role card for a structured AI tool
Field What it says
1. Role and owner The job this tool does, in one sentence, and the named person who owns the card.
2. Who uses it Which roles, and for which recurring tasks.
3. Scope What is in, and what is plainly out.
4. Approved sources The only places it may take facts from. Everything else is “I don’t know.”
5. Never A short list: prices, promises, legal or medical statements, personal data, anything not in the sources.
6. Output template The shape of a finished piece: length, sections, required fields.
7. Tone and design standards Voice, banned words and brand rules, with one good example and one bad one.
8. Stop and hand off The triggers for stopping, and the person it goes to.
9. Review Who checks, how often, and which outputs always get a human okay.
10. How we know it is right A few test cases with known good answers, a date last checked, and what triggers a rewrite.

Who writes it, and who keeps it current

Each card has a named owner, someone who answers for it. In training we write the first one together, from your own work, with real examples, not a template.

Getting that card into working software is separate build work, and it gets its own estimate. After that, upkeep is yours to arrange: do it yourself, hire someone, or contract with First Bell to maintain it. Every card needs a review date, because the tools change and a card goes stale.

What structure does not fix

A structured tool can still get facts wrong, so keep a review pass. It cannot fix an unclear goal. If you do not know what good looks like, it just guesses fast.

It needs test cases with known good answers, so you can see where it trips, and a date for the next check. It shrinks the review pile. It does not remove it. I can’t promise results.

Where this comes from

This comes from running two businesses and building systems for both. There is no study behind it and no published numbers. It is my experience, plain as that. If you try it and disagree, tell me. I would like to know.

Quick answers

Does this mean I don’t trust my staff?

No. This is delegation. The structure carries the standard in the tool, so people can ship without waiting on me. It is about the task, not about anyone’s rank.

Do we still need to review the work?

Yes. Review gets smaller and sharper as trust climbs the ladder, but a person still gives the okay at every rung.

Can I just give everyone an open-ended chatbot?

For private thinking, yes. For repeated work that leaves the building, add structure first. A bot’s guess sent to a customer reads as your answer.

Keep reading