How Do You Set Practical AI Rules for Your Team?

One person on your team pastes a customer's file into a chatbot. Another writes a reply and sends it without reading it twice. A third refuses to touch AI at all. Shared rules end that split, so the same task produces the same standard of work no matter who runs it. You set practical AI rules by agreeing, for one repeated task, who can use AI on it, which information they may put in, who checks the draft against that information, and what happens when the draft is wrong.

We help your team agree on the source material, instructions and review process for a repeated task. You check the output against the information your business supplied.

Key takeaways

• Without shared rules, each person makes their own call on data, quality, privacy and review standards. 1 2
• Keep the shared rules few and strong, and open them with three or four sentences describing what the business does. 3 4
• Separate the rules that apply everywhere from the rules for one task. The VirtusLab guidance reports this separation prevents contradictions. 5 6
• AI output will be wrong sometimes, so name the person who checks each draft against your source material. 7 In our sessions, participants test the workflow on a safe example and check the result against their source material.
• Agree the rules with the people who do the work. We scope training around your team's tasks, location and experience, then send a fixed quote.

What do practical AI rules for a team need to cover?

They cover four things: who uses AI, what information goes in, what the output must contain, and who signs it off. Rules of this kind sit between general standards, hard limits and the prompt for a single task. 8

Written rules keep the team inside ethical boundaries and deal with bias, privacy and the effect on people's jobs. 2 Skip them and every person makes an individual judgment call about data, quality, privacy and review standards. 1 That is where inconsistency starts, because ten people invent ten answers.

Start with one repeated task and build a workflow your team can run together. Quotes, enquiry replies, job reports, listing copy. The task sets the boundaries of the rules, so you are not trying to govern every possible use of AI at once.

Who on your team can use AI, and for which tasks?

Name the people and name the task. Permission attaches to the work, not to the person's curiosity.

Your business task determines the training example we build. Pick the job your team repeats most, decide which roles run it, and leave everything else out of the first version. A quote is written by the person who prices work. A reply to a complaint is written by the person who owns the account. AI drafts inside that boundary, and the boundary stays where it already sits in your business.

Rules also shape more than the wording of a draft. 9 They shape how the assistant behaves with the tools it can reach. 10 If the assistant can search files, send drafts or pull records, say which of those it may do for this task.

Which information can your team put into an AI tool?

Only the material your business has approved for that job. Source supplies the notes, documents or approved facts the assistant works from.

We build training around your business tasks, source material and the output you need. For a quote, that is site notes, measurements and your price list. For a reply, it is the customer's message and your policy. Your own notes, templates and rules do the work. Anything outside that list is off the table until the team agrees to add it.

Data choices carry consequences. AI systems can carry forward biases that sit in their training data. 11 Diversity and inclusion measures in how data is collected reduce bias in what the system produces. 12 The same logic applies to the material you feed a draft: a thin or one-sided source produces a thin or one-sided answer.

How do you write rules an AI assistant can follow?

Write them as five inputs the assistant reads every time: Context, Task, Source, Rules and Output. Rules work as a standing contract the model follows across a long session, not a single instruction typed once. 13 Most modern tools hold them as project rules, agent instructions or system prompts. 14 Cursor documents its version, and in Claude Code the equivalents are agents or skills. 15

Context describes the business, the customer and the situation. Task names the job the assistant needs to do. Source is the material it may use. Rules set the limits. Output describes the format and information the finished draft needs to contain.

A quote workflow reads like this:

We're a service business preparing a quote for a homeowner. Turn the supplied site notes into a draft quote. Use only these notes, measurements and the attached price list. Itemise labour and materials. Keep the supplied prices unchanged and flag missing information. Return a scope summary, itemised table, total including GST and questions for me to check.

Open your shared rules with three or four sentences describing what the business actually is. 4 That gives the model your domain vocabulary before it writes a word. 16 17 Name the tools you use, so the assistant does not reach for the wrong one. 18 Say where finished work goes. 19 Then keep the general rules limited in number and strong, covering only what truly applies everywhere. 3 Models handle too many instructions at once badly. 20

We help you build and test a workflow you can use again, so the wording stops being a guess.

How do you stop rules for different tasks from contradicting each other?

Split them by responsibility and let each task read only its own set. This separation prevents contradictions. 5

The VirtusLab guidance describes a software team's split:

The reason for the split shows up the moment two jobs need opposite behaviour. Integration tests generally avoid mocks because they focus on real interactions between components, while unit tests commonly use mocks to keep the tests isolated. 24 One combined rule file would force a choice that breaks one of them.

Tools give you ways to route the rules. Glob expressions in Cursor attach a rule to matching files, such as .kt for Kotlin files or /Test.kt for test rules. 25 26 In Claude Code, CLAUDE.md holds the core project context plus simple routing rules that send each action to the right agent, whether that is business logic, unit tests, integration tests or endpoints. 27 Each agent runs inside its own isolated context while the main agent delegates to the right specialist. 28 Every agent then reads only the rules for its own responsibility, which avoids contradictions and keeps responses clean and deterministic. 29

Your version is smaller. One rule sheet for the business, one per task, and no task sheet repeating what the business sheet already says.

Who checks AI output before it reaches a customer?

The person responsible for the work. AI outputs will be wrong sometimes. 7 A named reviewer turns that from a surprise into a step.

Give the reviewer a short checklist. Compare names, dates, figures and claims with your source. Recalculate totals. Check that the draft answers the request and stays within your scope. The person responsible for the work approves the final version.

In our sessions, participants test the workflow on a safe example and check the result against their source material, so the review habit forms on the day rather than on a live customer.

What should your team do when the AI gets something wrong?

Fix the input, then re-run. A draft can sound confident while missing context, misreading the instruction or inventing detail. 30 Treat a wrong answer as a gap in the five inputs.

Point to the unsupported claim. Supply the missing fact or ask the assistant to leave it as a question for you. Check the next draft for the same problem.

Then tighten the Rules line so the same error cannot repeat. Rules set the limits and identify missing information to flag. Keep supplied prices unchanged and flag missing information does more for a quote than any amount of correcting after the fact.

How do you agree the rules as a team and keep them current?

Agree them in a room with the people who do the work, using a real task and a real example. Rules written by one person for everyone else get ignored.

We bring private AI training to your workplace, so the team builds the workflow and the review rules on the job they already repeat. You can also travel to the Sunshine Coast for our public workshop. Sunshine Coast onsite training covers Maroochydore, Caloundra, Mooloolaba, Buderim, Birtinya, Nambour and Coolum.

After your enquiry, we discuss your team's tasks, location and experience. We agree on the training scope and prepare a fixed quote. Team size, session length and travel determine the price.

We do this work in our own business too. At Searchmaxxed we built AI workflows for creating blog posts, doing SEO and preparing custom proposals. We helped automate Radiant Roof Repairs' internal processes using AI, and we used AI to automate SEO, enquiries and admin for Glen House Homoeopathy. Each one needed its own source material and its own review step, which is why we start with your industry and your task rather than a template.

Frequently asked questions