How to scope your first small-business AI automation project

A useful first AI automation project starts with a repeated job that someone can explain. It should have identifiable inputs, a clear output and a person who can judge whether the result is right. Buying software before defining those things makes it harder to know what success means.

This guide takes you from a task to a supplier-ready brief. For a broader list of possible processes, see small-business automation examples.

1. Describe today's work

Follow a real task from its trigger to completion. Record who handles it, which systems they use and what happens when information is missing. Include the awkward cases rather than documenting only the clean version described in a sales demonstration.

Measure the current workload over a representative period. Count preparation, checking, interruptions and rework. Note variations such as busy days or different document formats. This baseline will make the later comparison more useful.

2. Decide whether the job needs AI

A rule can move a form submission into a spreadsheet, check that a field is present or send an approved reminder at a defined time. An AI component may help when text or documents vary and need interpretation, such as suggesting an enquiry category or drafting a summary.

Use deterministic rules for calculations and clear conditions where possible. Do not add a model to a reliable workflow simply because the project is called an AI project. A better form, shared inbox or clearer process may solve the problem.

3. Write a narrow first brief

Fictional example: a workshop receives maintenance enquiries in a shared inbox. The trial will suggest a category, extract the requested date and list missing details. It will create an internal draft for the coordinator. It will not quote prices, promise availability, email customers or book an engineer.

That brief separates the useful preparation from consequential actions. It also exposes the sources needed: approved categories, the incoming message and a record of what the coordinator considers complete.

4. Check the information before integration

Identify the authoritative source for each field. Resolve duplicate records, unclear ownership and conflicting instructions. Agree which information can be sent to an external provider, how it is retained and who may access it.

Begin with fictional or appropriately authorised test material. Do not copy a customer database into a new service just to see whether the tool works. The ICO AI and data protection guidance is a starting point for reviewing personal-data use.

5. Ask for a complete cost

Separate discovery, building, integration, testing, training and handover. Ask what is included in each, which assumptions the quote relies on and what triggers a change request. A low build price may exclude work the business still needs.

Also record recurring licences, model usage, hosting, monitoring and maintenance. Estimate internal review time and ownership effort. Compare quotes against the same brief rather than treating the headline fee as the total cost.

For Blue Canvas's published offers, use the current pricing page. Project-specific scope, delivery and any tax treatment should be clear in the written proposal.

6. Agree acceptance before the build

Write examples of normal work and expected failures. Include missing inputs, duplicates, conflicting sources, unavailable connections and a request to do something outside the agreed boundary. State when the correct response is to stop and ask a person.

Check the whole journey. If the screen says an item was saved, verify the record exists once in the intended system. If a message is drafted, check the recipient and attachments as well as the wording. A convincing demonstration is only one piece of evidence.

7. Train the person who will own it

The owner should know how to run the workflow, recognise an unsuitable result, correct a source and recover from failure. Give them a short operating guide and practice with the fallback. Decide who approves changes after launch.

Set access according to the job. Keep a record of integrations and credentials without putting secret values into ordinary documentation. Make sure the business retains access to the accounts and data it needs if the supplier relationship ends.

8. Decide from the evidence

Compare the assisted workflow with the baseline using the same task types. Include checking, corrections and failed runs. Report changes in time, quality and completion separately. Faster generation does not establish lower overall cost.

Expand only the parts that performed acceptably and have an owner. A trial that shows a simpler process is better has still answered a useful business question.

For help with the brief, see AI consulting for small businesses or book a free conversation.