How to choose the first workflow to hand to an AI agent

The best first workflow for an AI agent is a boring one. It runs every day or every week, it is full of emails, documents or tickets, one person owns it, you can tell when it went well, and a mistake is caught before it leaves the building. The agent prepares the work and a person approves it.

The worst first workflow is the one leadership is most excited about: the customer facing flagship, the process that decides about people, or the process nobody can quite describe. Those fail slowly and publicly, and they teach the organization the wrong lesson about agents.

What makes a good first use case for an AI agent?

Two of the companies building the models say much the same thing. OpenAI's practical guide to building agents tells teams to prioritise "workflows that have previously resisted automation": those with nuanced judgement and exceptions, rules that have become too complex to maintain, or a heavy reliance on unstructured data such as documents and natural language. Anthropic, in Building effective agents, says agents add the most value on tasks that have "clear success criteria, enable feedback loops, and integrate meaningful human oversight".

We turn that into six questions about each candidate workflow:

Criterion	What good looks like	Warning sign
Volume	It happens daily or weekly	It happens twice a year
Input	Emails, PDFs, tickets, free text	Clean fields that a simple rule could handle
Owner	One named person who feels the pain	"Everyone" or "the department"
Outcome	You can tell a good result from a bad one	Success is a matter of taste
Cost of a mistake	Caught internally, easy to undo	Reaches a customer, a regulator or a bank
Access	The agent can reach the inbox, files and tools it needs	Getting access is a project of its own

The owner matters more than it looks. A workflow without one person who wants it fixed has nobody to review the agent's drafts, nobody to say what "good" means, and nobody to keep using it once the project ends.

Which processes should you not automate first?

Four kinds of process make poor first projects, even when they look attractive.

• Anything that decides about people. Screening job applications or assessing someone's creditworthiness are listed as high risk uses in Annex III of the EU AI Act, and carry obligations that an internal back office task does not. They may be worth doing one day. They are not where you learn.
• Your most visible customer process. Every error is public, so the pressure to make the agent fully autonomous arrives before you have the evidence to justify it.
• A process nobody can describe. If three people give three different accounts of how it works, an agent will faithfully automate the confusion. Fix or document the process first.
• A process that does not need an agent. If the input always looks the same and the steps never change, plain automation is cheaper and more reliable. OpenAI's guide says it directly: if a use case does not clearly meet the criteria for an agent, "a deterministic solution may suffice". We wrote about that choice in AI agent, chatbot or automation.

The UK government's AI Playbook, written for the public sector but useful well beyond it, makes the same point from the other side: choosing a use case "must be led by business and user needs, pain points and inefficiencies, not what the technology can do".

How do you compare candidate workflows?

Keep it simple enough to do in one afternoon.

1. Ask the people who do the work. Not only managers. Ask each team what takes too much of their week and what they would happily never do again. You want a list of five to ten candidates.
2. Score each candidate from 0 to 2 on the six criteria in the table above.
3. Drop anything with a 0 on owner or cost of a mistake. No owner means no adoption. An expensive, irreversible mistake means the first project carries a risk it does not need.
4. Sort by total and look at the top three. Pick the one where the owner is most motivated, not necessarily the highest score.
5. Write one sentence of success. For example: "By the end of the second month, the agent prepares most of the weekly supplier invoices and the finance lead approves them in the tool they already use." If you cannot write that sentence, you have not picked a workflow yet.

Why is the back office usually the best place to start?

Because it scores well on almost every criterion at once. Invoices, contracts, supplier emails and internal reports arrive in volume, they are mostly text, they have an obvious owner, and a mistake is caught internally before anything is sent or paid.

We run our own group this way. The administration and finance of our own group of companies used to run through one overloaded inbox. Today agents sort the mail, prepare the reports, draft the paperwork and line up the payments, and a person approves anything that moves money or carries a signature. It was set up in weeks and then tuned in production, and it runs every day. You can read how it works in the back office case.

The pattern is not limited to finance. For a brand, the same shape appears in supplier correspondence, product data, retailer reporting or the weekly competitor summary nobody has time to write.

What should the first weeks look like?

The agent prepares and a person approves, from the first day. OpenAI's guide recommends that actions which are "sensitive, irreversible, or have high stakes" trigger human oversight "until confidence in the agent's reliability grows", and closes with the advice to "start small, validate with real users, and grow capabilities over time".

In practice that means three things.

• Build into existing tools. The agent works in the inbox, chat, task board or shared drive the owner already uses. Approval happens there too.
• Measure from week one. Track what the agent did, what a person corrected and what it cost. Corrections are the most useful signal you have: they show exactly where the agent needs better instructions, better data or a narrower scope.
• Widen only on evidence. When corrections on a narrow task fall close to zero, that task can run with less supervision. Money, signatures and anything a customer reads in your name stay with a person. We explain why in why our AI agents never move money on their own.

Where to start

List five to ten candidate workflows with the people who do the work. Score them on volume, input, owner, outcome, cost of a mistake and access. Drop the ones without an owner or with an expensive mistake, pick one, and write the sentence that says what success looks like. Then run a small, real test before you write a strategy around it. That is how we approach AI strategy: one grounded experiment before a long plan.

If you have a shortlist and are not sure which workflow to start with, we are happy to go through it with you.