All insights

How to choose your first AI workflow

Start with one repeatable task, a clear owner, and a result your team can check. A practical way to find an AI project worth doing.

A designer selects a small violet component from a workbench.

The practical takeawayChoose one repeatable task and define how a person will check the result before you connect more tools.

  1. Find repeatable work
  2. Run a small pilot
  3. Check the output

A useful first AI project usually starts with a familiar frustration: the same information gets copied between tools, routine questions keep interrupting the team, or a straightforward document takes too long to prepare. The opportunity is specific enough that you can describe what happens today and what would improve.

That is a better starting brief than “we need an AI agent.” Before choosing a tool, choose the work. You want a task with a real owner, a clear input, and an output that someone can evaluate. This guide gives you a practical way to narrow the options and run a manageable first pilot.

Start with work someone already understands

Ask your team which recurring task feels needlessly slow. Watch the task happen once. Write down the steps, the systems involved, the judgment calls, and the exceptions. You may discover that part of the problem is simply a missing form field or a broken integration. Fixing that can be more valuable than adding another application.

For an illustrative nonprofit example, imagine a coordinator turning volunteer inquiry forms into a weekly planning sheet. AI might help summarize open-ended availability notes. A normal integration can move the name and email into the correct columns. The coordinator still decides where each person belongs.

This distinction keeps the proposal grounded. Anthropic’s engineering guidance on agent design recommends starting with the simplest solution that meets the need and adding complexity when it earns its place. A small, supervised step can be a perfectly good first implementation.

Choose a result your team can check

Some tasks are easier to evaluate than others. Extracting a requested appointment date from a message has a source to compare against. Deciding whether a customer deserves a special commercial arrangement involves context and responsibility that may sit outside the message.

For your first pilot, favor tasks where the source is available, mistakes are reversible, and a reviewer can spot a problem quickly. Keep the proposed output close to the evidence. A summary should link back to the original material. A drafted response should remain a draft until the responsible person approves it.

Also define the handoff for uncertainty. “I cannot determine this from the information provided” can be a useful result. Requiring an answer to every question encourages your team to overlook missing context. The process should make it easy to ask a person instead.

Write a one-page pilot brief

A short brief is enough to expose the important decisions before implementation. Give the pilot a narrow boundary. “Prepare an internal summary of new fleet service inquiries” is easier to test than “automate sales.” Include these items:

  • Input: exactly which form, document, or message starts the task.
  • Output: what gets produced, where it goes, and the required format.
  • Reviewer: who checks accuracy and approves the next action.
  • Data boundary: which information the approved tool may receive.
  • Fallback: how the work continues when the tool is unavailable or uncertain.
  • Success test: what would make the pilot useful enough to keep.

For example, success could mean preparing complete summaries with fewer corrections and less total staff time. Write your own threshold before the pilot. It should reflect your workflow and the consequence of a mistake, rather than an impressive-looking benchmark from another organization.

Measure the whole task, including review

Time a few examples of the existing process, then compare similar examples using the proposed workflow. Count the time spent checking, correcting, and transferring the result. A fast first draft that creates a long editing job may not improve the working day.

Keep a small evaluation log. Record whether required information was captured, whether unsupported statements appeared, how much editing was needed, and whether the final output was useful. Include unusual but realistic cases: an incomplete inquiry, conflicting dates, or a message covering two services.

These are practical acceptance checks, not a claim that a small pilot predicts every future result. NIST’s AI Risk Management Framework treats evaluation and risk management as ongoing activities. In practice, review the workflow again when its inputs, tools, or business purpose change.

Give the workflow an owner after launch

A successful pilot still needs someone to look after it. Decide who receives failure alerts, who can change the instructions, and who checks a sample of outputs. Keep a plain-language note explaining what the workflow does and how to turn it off.

Start with a limited rollout. Let the people doing the work point out friction, then adjust the process before expanding access or allowing the system to take additional actions. The goal is dependable assistance that your team understands.

For your next step, ask one team member to bring a recurring task and three sanitized examples. Use them to write the pilot brief together. TripleSkinny can help turn that brief into a working implementation, connect it to your existing tools, and stay involved as it becomes part of everyday operations.

Sources & further reading