Pick the workflow that runs often, crosses at least three people, has a named owner, and can be measured end to end within a few weeks. Those four filters remove most of the bad candidates before you spend anything. The most common reason a first AI project shows no return is not that the redesign was done badly. It is that the wrong workflow was chosen, and no amount of good execution saves a bad pick.
Deloitte found that 37 percent of organizations making real changes use the same pattern: fully own one workflow, test it, then scale. This is how you choose that one.
Step 1: List every workflow that crosses three or more people
Open a document and write down workflows, not departments.
A workflow has a trigger and a finish line. "Onboarding a new customer" is a workflow. "Customer success" is not. If you cannot say what starts it and what ends it, it is a function, and you cannot measure a function.
Only write down the ones that pass through at least three different people.
That number is not arbitrary. One person doing a task alone is a job for a better tool, and you will get a modest speed gain. The compounding delay in a business lives in handoffs, so a workflow with no handoffs has very little for you to win back. Three or more people means there are at least two handoffs, which is where the time actually goes.
Aim for a list of ten to fifteen. More than that and you are describing tasks.
Step 2: Score each one out of three on four things
Give every workflow on your list a 1, 2, or 3 for each of these. Do it fast, by instinct. You are sorting, not auditing.
Volume. How many times does this run per week? Once a quarter scores a 1. Several times a day scores a 3. Rare workflows are hard to learn from because you get one data point a month.
Pain. Does anyone complain about this unprompted? If people have built private workarounds, spreadsheets, or side chats to survive it, that is a 3. Nobody mentions it, that is a 1.
Stability. Does it run roughly the same way each time? A workflow with forty exception paths scores a 1. One that follows the same shape almost always scores a 3. You cannot redesign something that has no reliable shape.
Reversibility. If this goes wrong, how bad is it? Something you can undo the same afternoon is a 3. Something that touches a regulator, a payment, or a patient is a 1.
Add up the four numbers. Anything scoring 9 or above is a real candidate.
Step 3: Apply the disqualifiers
Now delete candidates, even high scorers. A workflow is disqualified if any of these are true.
Nobody owns it. If you cannot name one person who is accountable for the outcome, stop. You will spend the whole project negotiating with three directors who each own a piece. Deloitte's finding that only 4 percent of organizations report AI value at board level is partly an ownership problem. Unowned work produces unowned results.
The data lives in someone's head. If the reason it works is that Yasmin has done it for nine years and knows who to call, there is nothing for a system to read. That is a documentation project first.
It is mid-reorganization. If the team, the tooling, or the process is already changing for other reasons, your before and after measurement is worthless. Wait.
It is the most regulated thing you do. Do not start on the workflow with the highest audit exposure. Start somewhere you are allowed to be wrong once.
Step 4: Time it properly, including the dead air
Take your top candidate and measure how long it actually takes today.
Not how long the work takes. How long the workflow takes. Those are very different numbers and the gap between them is your entire opportunity.
Do it like this. Find ten recent times this workflow ran. For each one, write down the timestamp when it started and the timestamp when it finished. That gives you elapsed time.
Then separately, add up how many minutes of actual human effort went into it.
Compare the two. In most businesses the elapsed time is many times larger than the effort time. A request that takes forty minutes of real work takes six days to come out the other end. The difference is waiting: sitting in an inbox, waiting for an approval, waiting for someone back from leave, waiting for a meeting that happens on Tuesdays.
That difference is what you are going after. If elapsed time and effort time are close, this workflow has nothing to give you and you should pick the next candidate.
Step 5: Write the baseline down before you touch anything
This is the step people skip, and skipping it is why so many projects cannot prove anything afterwards.
Write down four numbers, today, before any tool is bought:
- Median elapsed time from trigger to finish
- Number of handoffs between people
- How often it gets sent back or redone
- Cost, using rough salary math on the effort time
Put them in a file with today's date and tell the workflow owner they exist.
You will not get a second chance to measure the before. Once you start changing things, the original state is gone, and every claim you make later becomes an argument instead of a fact.
Step 6: Decide who owns each step, person or machine or nobody
Now walk the workflow step by step. For each one, pick exactly one of three answers.
A person does this. Judgment, relationships, accountability, anything where being wrong is expensive and context is thin.
A machine does this. Fetching, formatting, summarizing, chasing, routing, checking something against a rule, moving information from one place to another.
This should not happen. Be serious about this one. A step that exists only to tell someone that something happened is usually a notification, not a step. A review that has never once changed an outcome is a habit, not a control.
Metrigy studied 759 companies and found that 53.7 percent use AI to automate existing workflows instead of changing how the work gets done. This step is the difference. If you finish this walkthrough and every step still exists in the same order with the same owners, you have not redesigned anything. You have just bought a faster version of what you had.
One rule that matters more than it sounds: whatever machine you put on a step needs the same access and context the person on that step had. If it has less, it will keep handing work back to a human to fetch things, and you have replaced one handoff with two.
Step 7: Measure the one number end to end
When it runs, measure elapsed time from trigger to finish. That is the number.
Not tokens. Not messages handled. Not how many people say it feels faster. Step level improvements are where false wins come from, because you can make every individual step quicker and change the total by almost nothing if the waiting between steps is untouched.
McKinsey's 2026 survey found that of 25 organizational attributes tested, workflow redesign had the biggest effect on whether AI produced any EBIT impact at all. The organizations that saw it are the ones who could point at a whole workflow and say it used to take nine days and now takes two. That is a sentence a board understands and a per-seat licence count is not.
What a good first pick looks like
Frequent enough that you learn something every week. Enough handoffs that there is real waiting to remove. Boring enough that nobody panics if it breaks once. Owned by one person who will answer your messages. Measurable in weeks, not quarters.
Most companies pick the opposite: the most visible, most complicated, most regulated workflow, because that is the one leadership talks about. Then it takes eight months, proves nothing, and the conclusion becomes that AI does not work here.
Pick the boring one. Prove the number. Then go take the hard one with evidence in your hand.
Frequently asked questions
How do I choose which workflow to automate with AI first? Score your candidates on volume, pain, stability, and reversibility out of three each, then disqualify any workflow with no named owner, undocumented knowledge, an active reorganization, or the highest regulatory exposure. The best first pick runs frequently, crosses at least three people, and can be measured end to end within weeks.
Why does a workflow need to cross three or more people? Because the delay in most business processes lives in the handoffs between people, not inside the tasks themselves. A workflow with three people has at least two handoffs. A single person workflow has none, so the only available gain is making one person slightly faster.
What is the difference between elapsed time and effort time? Effort time is how many minutes of actual human work a workflow consumes. Elapsed time is how long it takes from trigger to finish in real days. In most businesses elapsed time is many times larger, and that gap is waiting. The gap is the opportunity.
What should I measure before starting an AI workflow project? Record four numbers before making any change: median elapsed time from trigger to finish, number of handoffs, rework rate, and rough cost from salary math on effort time. Once you start changing the workflow, the original state is unrecoverable and later claims become arguments rather than facts.
Should I start with my most important workflow? No. Start with one that is frequent, stable, reversible, and owned by one person. Starting with the most visible or most regulated workflow usually means a long project with nothing provable at the end, which is a common reason companies conclude AI does not work for them.
