← Back to question navigation

Question Navigation / Decision Brief

Which scenario should an enterprise AI Agent start with? Criteria and first steps

Conclusion first:First find a piece of work that recurs, has clear inputs and outputs, has someone who can accept it, and can be rolled back if something goes wrong. Get that piece running smoothly first, then decide whether you need a knowledge base, tool invocation, automation, or greater autonomy.

Suitable for business leaders, business managers, and professional services teams that haven't yet determined where to start with AI projects.

This page first makes the judgment clear, then explains when it applies, how to start, and where human professional responsibility is still required.

Don't start with "roll out Agents across the whole company"

What companies really need to solve is usually not "do we have an AI" but why a certain piece of work is repeatedly prepared, repeatedly re-explained, and repeatedly waiting for confirmation. Only by narrowing the problem down to one real task can you judge whether the investment is worthwhile.

Models, platforms and interfaces can all be swapped; work objectives, responsibility boundaries and acceptance evidence must be clarified first.

Check the four conditions first

  1. Repeatedly happens:Not a one-off flash of inspiration, but a task that comes up every week or every month.
  2. Clear inputs and outputs:Know what it needs to read and what format of result it must ultimately deliver.
  3. Someone is responsible for acceptance:Business owners can judge facts, applicable definitions and criteria, and quality—not just whether it "reads right."
  4. Risk can be controlled:Start with steps that are reviewable, reversible and not directly committed externally.

The first batch of work suitable for a trial run

  • Turn meeting, email, or chat records into to-dos, risks, and next steps.
  • Customer background, needs gaps, meeting preparation and follow-up drafts for sales or pre-sales.
  • Find evidence from specified materials and produce an internal preparation draft with sources and items to confirm.
  • Coordinate materials, time and tasks across teams, and write results back to business records.

These tasks all allow the Agent to prepare first, while humans retain fact verification, pricing, resources, commitments, and final sign-off.

Things not to rush into for now

  • No clear business objective—just chasing a trending model.
  • Expects the Agent to directly replace key professional judgment or make irreversible commitments to customers.
  • No de-identified materials, no acceptance owner, and no manual takeover path in case of failure.

Write the first step as a one-page work instruction

Before starting to build, write down six things: trigger conditions, input materials, expected output, permitted sources, who must confirm, and how to fall back to a human on failure. One page like this says more about whether the project is worth continuing than buying a complex platform first.

Start from a real problem

Bring a piece of work that's currently happening to the conversation.

You don't need to pick a model first, or write your requirements as a technical spec. Just explain what you're delivering, where your materials are, and which step takes the most time — we'll first determine what category of problem it is.

Contact Andy