How to choose your first AI automation

Your first AI automation should solve a task you already understand. Choose work that repeats, has a clear finish, and can be checked before it changes something important. For a small consultancy or agency, preparing a client-enquiry summary is often an easier starting point than handing an entire inbox to an autonomous agent.

The aim is a dependable improvement to existing work. This guide offers a way to choose that first task and an illustrative workflow you can adapt. It is a planning exercise, not a customer case study or a claim of measured time savings.

Start with the recurring task

Write down a few tasks that interrupt your week. Examples might include sorting enquiries, turning meeting notes into draft actions, collecting onboarding details, or assembling a weekly project update.

Describe each task as an action with a finish: “Turn an incoming enquiry into a summary and a proposed next step.” That is easier to build and evaluate than “Use AI in sales.”

For each candidate, answer five questions:

1. Does it repeat? A recurring task gives you enough examples to understand the normal cases and exceptions.
2. Are the inputs available? Identify the message, document, form, or record the workflow needs, and who is allowed to use it.
3. Can you describe a good result? Decide what must be present, what must remain unchanged, and what counts as a mistake.
4. Can someone check it quickly? A draft summary is useful only if reviewing it takes less effort than doing the task yourself.
5. Can you stop and recover? Start with an output you can discard or correct while you learn how the workflow behaves.

If you cannot explain the current process, first clarify it. Connecting tools will carry that uncertainty into the next step.

Decide whether the task needs AI

Some automation is straightforward: copy a confirmed form field into a record, send a reminder on a known date, or move a file when a rule matches. Those steps may work well without a language model.

AI can be useful when an input needs interpretation, such as summarising a free-form enquiry or extracting proposed actions from notes. It can also misread details or supply information that was never provided. Give it a bounded job and check the output.

In Building effective agents, published in December 2024, Anthropic distinguishes workflows with predefined paths from agents that choose their next steps. Its general design advice is to start simply and add complexity when it improves the outcome. The article's tooling discussion is dated; the distinction is useful for thinking about the work.

For a first project, define the steps yourself. Our guide to AI agents versus automations explores that distinction further.

Map one complete workflow

Consider a fictional consultancy receiving an enquiry about a team workshop. The message mentions the team's goals but leaves the date and budget unspecified.

A bounded enquiry workflow could follow these steps:

1. Receive: take a copy of the enquiry from an agreed inbox or form.
2. Interpret: extract the request, stated constraints, and missing details. Keep a link to the original message.
3. Prepare: draft a summary and questions for the missing information. Do not invent availability, prices, or promises.
4. Review: a person checks the details and decides the next step.
5. Act: after approval, create a follow-up task in the agreed system. Sending a customer response is a separate action that needs its own clear approval rule.
6. Record: retain the result, its source, and whether the follow-up was actually created.

AI handles the interpretation and draft. The surrounding workflow coordinates the tools, checks, and hand-off. A generated summary does not establish that a task was created or a message delivered.

Write the outcome before you build

Use a short brief like this:

When a new enquiry arrives, prepare a summary of the request, list the information still needed, and suggest one next step. Cite the original message. Mark unknown details as unknown. Keep the result as a draft until I approve it. After approval, create one follow-up task and record its identifier. If a required step fails, stop and show me the failure.

Adapt that brief to your actual tools and permissions. “Draft only” in a prompt is not an access control: configure the workflow so the draft stage cannot send messages or change customer records.

You can use Give AI a better brief to make the input and review criteria clearer.

Test normal cases and awkward ones

Begin with fictional examples or copies of data you have permission to use. Check an ordinary enquiry, a request with missing details, an ambiguous message, a duplicate submission, and a failure in the destination tool.

For each example, compare the result with the source. Did the system preserve the facts? Did it flag uncertainty? Did it wait for approval? If the destination was unavailable, did it show the failure rather than claim success?

Run the same input more than once. Check how the system prevents duplicate follow-up tasks when an event is retried. Record failures and revise the workflow before relying on it for routine work.

Measure the whole job

Before changing the process, note how long the current task takes and what commonly goes wrong. When testing the proposed workflow, include preparation, review, corrections, and recovery in the comparison.

A quick AI response can still create more work if it requires careful rewriting. Look for a better completed outcome, not just a faster generation step. Do not project a test result across the business until you have evidence from the actual work.

Our earlier article on automating a task or doing it manually offers another way to think about the trade-off. Read it in the context of its original publication date.

Give the workflow an owner

Decide who reviews the output, where a failure appears, and how to pause the workflow. Document which accounts it uses, what it can change, and what happens when an input is incomplete.

Your first automation is ready for a limited pilot when you can explain the complete path, show its failure states, and inspect the real result. Expand its scope only as the evidence supports it.

If you have a recurring piece of client administration in mind, tell us about the workflow. York Studio can explore whether learning, a consultation, a workshop, or a scoped implementation is a useful next step.