How to Scope Phase One AI Automation

A phase one AI build should be small enough to approve and specific enough to test, but not so narrow that it becomes a dead end. The boundary is the strategy.

This week's client questions made that tension visible. The buyer asked whether the build phase meant one or two action items or something all-encompassing. She also wanted to start small while building for scale from the very start.

That is the right concern. A good phase one does not automate everything. It proves one repeatable loop and leaves the foundations clean enough for the next loop.

What makes a good phase one AI scope?

A good phase one AI scope has one workflow, one owner and one measure of success. If any of those are missing, the build becomes a collection of promising ideas rather than a controlled business change.

• One workflow: the exact task run the system will reduce or improve.
• One owner: the person who decides whether the output is useful enough to keep.
• One measure: the evidence that the workflow is saving time, improving follow-up or reducing dropped context.

How do you start small without blocking scale?

Start small by limiting the workflow, not by ignoring the future architecture. The first phase should use the same design principles the scaled system will need: clear inputs, controlled access, review points, logging and a named place where useful context returns.

For example, a real estate team might begin with a daily briefing that reads appointments, relevant CRM notes and internal tasks. That first loop does not need to solve every sales workflow. It does need to prove that context can be captured, interpreted and returned to the team in a way they will actually use.

That is why AI adoption fails when the workflow is too hard. A small build still fails if it asks busy people to maintain one more disconnected system.

What should phase one exclude?

A disciplined first phase should exclude anything that adds risk before the review model is trusted. That usually means no unsupervised external communication, no changes to sensitive financial or identity data and no broad access to systems that are not needed for the first workflow.

It should also exclude vague expansion promises. The phrase “fully integrated system” is useful as a destination, not as a phase one scope. The first build should translate that destination into the smallest useful proof.

• Bad first phase: connect every system and see what AI can do.
• Better first phase: create a daily internal briefing from calendar, CRM and task context, reviewed by the team lead.
• Best first phase: define the trigger, source systems, output, review step, owner and scale condition before build starts.

How should phase one be priced and approved?

The commercial model should match the learning model. If the first phase is a controlled loop, approval is easier because the buyer can see what is being bought: a working workflow, not an undefined AI programme.

That does not mean every future build must be quoted before the first one starts. It means the first proposal should separate the foundation from the build-outs. The buyer should understand what is included now, what decisions unlock the next phase and what evidence will justify further work.

What evidence earns phase two?

Phase two should be earned by evidence, not enthusiasm. The system should show that users came back, the output saved a real step, sensitive cases were reviewed and the workflow improved after exceptions.

Use three gates:

• Useful: the workflow removes or reduces a repeated task the team already feels.
• Trusted: people rely on the output without needing to rebuild the context from scratch.
• Controlled: mistakes, sensitive cases and exceptions have somewhere visible to go.

This is also the practical difference between a generic tool and an organisational brain: the work improves because context survives and action appears in the right place.

A simple phase one charter

Before build starts, write the phase one charter in plain language. It should fit on one page.

• Business burden: what pain are we removing?
• Workflow boundary: where does the task start and end?
• Source systems: what context can the assistant read?
• Output: what does it create, route or update?
• Review rule: what must a human approve?
• Scale condition: what evidence earns the next phase?

How Catalyst scopes the first build

Catalyst Systems starts with the work people are already carrying. We look for the repeated conversation, decision, promise or follow-up that keeps being rebuilt from memory, then turn one narrow loop into a system the business can trust.

If your AI project needs to start small without becoming throwaway work, book a Sprint conversation. We will help you define the first workflow, the review path and the evidence that should unlock phase two.