Choosing the first workflow for an AI agent

Choose a workflow your team does by hand every day, that takes real time, and where a mistake is cheap to catch. Map it as people run it today before you rank anything, then rank each step by the time it takes now and the risk of getting it wrong. The first workflow for an AI agent comes from the steps that take the most time and carry the least risk.

Your team decides from the first workflow whether to trust an agent, and that decides whether it gets a second one. This is how we choose it in the first two weeks of a forward-deployed engagement, and how you can choose it yourself.

Signs a workflow is ready for an agent

Agents do well with the parts of a job that are repetitive and written down. A workflow is a good candidate when most of these are true:

People do it by hand every day, or many times a week.
The work is checking, matching, chasing or re-keying: moving information from one place to another and comparing it with something else.
The inputs are already written down, in emails, PDFs, forms, spreadsheets or a system of record. An agent can read what a person reads.
Someone can tell a right result from a wrong one quickly, by looking at it.
The tools involved have a way in: an API, an export, or a sandbox copy that can be connected.
One person owns the workflow and wants it to change.

The last one is easy to skip. An agent changes how people spend their day, and someone has to decide what changes, answer questions during the week, and say when the result is good enough.

In practice the candidates look ordinary: supplier invoices in finance, or checking delivery paperwork against orders in operations. Nobody would put that work on a slide, and it can fill hours of somebody’s week.

Signs it should wait

Some work is better left until later, or left with people:

Every case is different, or the work comes up a few times a month. There is too little of it to learn from or to save time on.
The rules live in one person’s head and shift from week to week. Mapping can bring them out, but an agent should not be the first thing to try to follow them.
A wrong answer reaches a customer, moves money or breaks a regulation before anyone can catch it.
The process or the system is about to change, for example halfway through a move to a new ERP.
Nobody can get access to the systems, even a sandbox, and the data rules are not yet settled.

A workflow on this list can come back once the problem is fixed. It should not go first, because the first one is where your team forms its view of agents.

Map before you rank

The usual way to choose is to collect ideas in a meeting and vote. That ranks the ideas, and ideas tend to be the work people talk about rather than the work that fills their week. The quiet, repetitive steps rarely come up, because nobody thinks of them as a project.

So we start by sitting with the people who do the work and writing down every step of one workflow as it runs today: what the step is, who does it, in which tool, and what kind of work it is. Ask someone how long a task takes and you tend to hear how long it should take. Sit beside them and you see the copying between windows and the second look at a smudged PDF.

Example: supplier invoices, as a team runs them todayMapped, not ranked yet

Download the invoice from the shared inbox

Who
Accounts

Where
Email

Kind of work
Collecting
Re-key it into the ERP

Who
Accounts

Where
ERP

Kind of work
Re-keying
Match it to the purchase order and delivery note

Who
Accounts

Where
ERP, shared drive

Kind of work
Matching
Chase the missing delivery note

Who
Accounts, warehouse

Where
Email, phone

Kind of work
Chasing
Approve the payment

Who
Finance lead

Where
ERP

Kind of work
Deciding

Four of the five steps are collecting, re-keying, matching or chasing.

An example, not a client’s workflow. Filled squares mark steps an agent could take on; the open square is a decision that stays with a person. A real map also records the measured time per case for each step.

Mapping gives you two things besides the map. The first is the numbers to measure against later: how many cases come through, how long each one takes, and how often a person has to step in. Without them, nobody can say afterwards whether the agent helped. The second is the kind of work each step is, which is where the ranking starts. In the example, four of the five steps are collecting, re-keying, matching or chasing, and only one is a decision.

Rank by time today and the risk of getting it wrong

With the map in front of you, two questions sort the steps.

How much time does it take today? Count how often the step happens as well as how long it takes each time. A short task done all day adds up to more than a long one done once a month.

What happens when it is done wrong? Ask how far a mistake travels before someone sees it, whether it can be undone, and who sees it first. A re-keyed field that is checked again at approval is a low risk. A payment released to the wrong supplier is not.

The same five steps, rankedStart top left

MoreTime it takes todayLess

Start here

Re-keying into the ERP
Chasing delivery notes

Next, with a person checking

Matching to the purchase order

Fold into the first

Downloading from the inbox

Keep with people

Approving the payment

LowerRisk of getting it wrongHigher

The first workflow takes real time today and is cheap to get wrong.

Placed with the team, by judgement, using the measured times. The axes are a sorting aid, not a score.

Each corner tells you what to do with the steps in it:

Start here: more time, lower risk. The first workflow comes from this corner. It is worth automating, and a mistake is caught before it costs much.
Next, with a person checking: more time, higher risk. Worth doing once the first workflow is running, with a person reviewing every case until the numbers show the agent can be trusted with more.
Fold into the first: less time, lower risk. Small on their own, but cheap to add when they sit next to the first workflow. Downloading the invoice is a natural part of re-keying it.
Keep with people: less time, higher risk. Decisions like approving a payment. An agent can lay out everything a person needs to decide quickly, and the decision stays with the person.

The team places the steps together. The people who do the work usually know best where a mistake would hurt, and a ranking they helped make is one they will hold the agent to.

If you have several candidate workflows, map and rank each of them, then take one into the pilot. One workflow, done properly and measured, tells you more about agents in your business than three started at once. The rest stay on the shortlist for later.

What the agent does, and what a person checks

For a first workflow, aim for an agent that does the repetitive steps and passes everything it cannot settle to a person, with the evidence in front of them. In the invoice example, that looks like this:

Example: the first workflow, with a person in itReady for approval

Invoice arrives The agent picks it up from the shared inbox
Re-keyed into the ERP Supplier, invoice number, lines and amounts Goes to a person: If a field cannot be read
Delivery note found Or the agent chases the warehouse for it Goes to a person: If the warehouse has no record of the delivery
Ready for the finance lead Approving the payment stays with a person

Needs a person

Invoices the agent cannot settle go into one queue, with the invoice, what the agent did and why it stopped.

The agent prepares the invoice; a person approves the payment.

An example, not a client’s workflow. Each step either passes the invoice on or sends it to a person.

The queue is part of the design from the start. Every case the agent stops on goes to one place, where a person sees what came in, what the agent did and what it could not resolve. A good queue makes that person faster, because they start from the evidence instead of from the beginning.

We built Theia, our own plate recognition product for logistics yards, around the same split. The camera reads the plate, and the reading is checked against Vahan, the national vehicle registry, and against the GST e-way bill for the load. Trucks that do not resolve go to a queue, where a person sees the detected truck, the plate as read and a field to verify it. The gate keeps moving while a person looks at the few trucks that need it. There is more on that in what building Theia taught us.

While the agent is new, a person checks every case it handles, not only the ones in the queue. In an engagement this happens in weeks 7 and 8, when the agent runs beside your people and is measured against the numbers from the mapping. The checking narrows only when those numbers say it can.

What you need to have ready

Choosing well takes some of your team’s time. Before an engagement starts, we ask for:

One person who owns the outcome and can make decisions.
About 2–3 hours a week with the people who do the work, so the map reflects how they work.
Access to the tools involved, or a sandbox copy of them.
Real examples of the work, under your data rules: real invoices, emails or records, handled the way your policies say.
Someone from IT or security, for access and to review how the agent connects.

The examples matter most. An agent tested only on tidy, invented cases meets its first smudged scan soon after it goes live. Real examples, including the awkward ones, become the set the agent is checked against for as long as it runs.

If your data has to stay on your own systems, the agent can run on your own servers; it can also run in your cloud or in ours. It is worth settling this with IT or security early. Our terms for your data are set out in writing on the forward-deployed engineering page.

What you keep after the first two weeks

By the end of the first two weeks you have four things on paper:

A written map of the workflow as your team runs it today, step by step.
A ranked shortlist of what to automate first, by the time each step takes today and the risk of getting it wrong.
The numbers to measure against: volume, time per case, and how often a person has to step in.
A fixed plan and a fixed price for the pilot.

They are yours to keep whether or not you go ahead. If you do, weeks 3 to 6 build the first agent on your tools, with frequent demos and a progress update every week, and weeks 7 and 8 run it beside your people. After that your team runs it, or our engineers move to the next workflow on the list. How we work covers the rest of what working with us is like.

If you are weighing outside engineers for this, hiring forward-deployed engineers sets out what to ask any provider. Forward-deployed engineering is one part of our AI engineering work. When you have a workflow in mind, tell us about it, and a partner will tell you honestly whether it makes a good first one.