Start with workflows that repeat many times a week, take hours, move between people, and involve copying, checking or chasing information by hand. Score each candidate on three things: hours per week, how easy it is to build, and what happens if the AI gets it wrong. Start where the hours are high, the build is simple and the risk is low. Keep the steps that need judgement with people.
Why the first choice matters so much
The first workflow you change with AI is more than a project. It is the proof everyone else in the company will judge AI by.
If it works, people ask for the next one. The team that used it tells other teams. The board sees numbers and approves the next cycle. If it fails, or if it works but nobody can show it, AI becomes "that thing we tried". Recovering from a visible failure takes far longer than starting carefully.
So choose the first workflow for proof and for learning, not for ambition. The biggest opportunity in the company is rarely the right first project.
Where AI helps today, and where it does not
AI is strong at work that follows rules and repeats. It is weaker where the rules are unclear or the stakes are high.
| AI does this well today | Example |
|---|---|
| Reading and sorting incoming work | Labelling customer emails and routing them to the right person |
| Checking documents against clear rules | Checking a client's documents against an onboarding checklist |
| Moving information between systems | Filling CRM fields from an email instead of retyping |
| Drafting from a pattern | The email that asks a client for a missing payslip |
| Assembling reports from several sources | The monthly management report from CRM and accounting data |
| Answering standard questions | "Where is my application?" with the status from the system |
It does less well, or should not act alone, where a wrong answer has legal or financial consequences, where the rules live only in someone's head, where the cases are rare and all different, or where a message goes straight to a client without anyone checking. That does not rule those workflows out. It means the AI prepares and a person decides.
Four signs of a good candidate
It repeats. Work that happens forty times a week gives AI forty chances a week to help, and gives you enough cases to measure. Work that happens once a quarter does not.
It costs hours. Frequency times minutes per case. A five-minute task done 200 times a week is worth more attention than a two-hour task done once a month.
It moves between people. Every handover adds waiting time, and waiting time is often where clients feel the delay. A workflow with three handovers usually has more to gain than one that stays with one person.
People copy, check or chase by hand. These are the steps AI takes over most reliably today, and the steps people are most willing to give away.
Two more conditions make the first build much easier: the data is in systems you can connect, and a mistake gets caught before it reaches a client.
How to score your candidates
Once you have a baseline of your main workflows, with steps, volumes and minutes, scoring takes an afternoon. Use three questions per workflow.
How many hours per week does it take? Take this from the baseline, not from a guess. It is the size of the prize.
How easy is it to build? Is the information in systems you can connect, such as Microsoft 365, Google Workspace or your CRM? Are the rules clear enough to write down? Can you test it on past cases?
What happens if the AI is wrong? If a person checks every output before anything leaves the company, the risk is low. If a mistake would reach a client, a regulator or the books unchecked, the risk is high.
Put hours on one axis and ease on the other. Start in the corner with many hours and an easy build. Then filter on risk: high-risk workflows can still start early, but only with a person checking every step that carries judgement.
| Easy to build | Hard to build | |
|---|---|---|
| Many hours a week | Start here | Plan carefully, start with the preparation steps |
| Few hours a week | Quick wins, good for learning | Leave for later |
A worked example: six workflows at one company
This example is illustrative: a mortgage and lending adviser with 170 people. After a two-week inventory, the team drew six workflows and measured them.
| Workflow | Hours a week | Ease to build | Risk if wrong | Decision |
|---|---|---|---|---|
| Mortgage file preparation | 53.5 | Medium | High | Next cycle, starting with document collection |
| Customer email triage | 42.7 | High | Low | First cycle |
| Client onboarding and document check | 33.7 | High | Medium, with a check | First cycle |
| Supplier invoice processing | 11.4 | High | Low | Later |
| Monthly management report | 6.3 | High | Low | Quick win in the first cycle |
| Complaint registration | 5.5 | Medium | High | Later, with Compliance |
The biggest workflow is not the first one. Mortgage file preparation takes the most hours, but much of it is writing the advice rationale and the four-eyes check: regulated judgement that the company wants people to own. The team plans to start there next cycle, with the preparation steps only, collecting documents and calculating income, while advisers keep the rationale.
Customer email triage and client onboarding go first. Both take many hours, both run on email and the CRM, and in both a person can check the AI's work before anything reaches a client. The monthly report joins as a quick win: few hours, but visible to the board every month, which makes it good proof.
Look at handovers, not only steps
When you look at a workflow step by step, it is tempting to pick the slowest step and make it faster. Often a better target is the handover between two steps.
In client onboarding, the slowest moment is not a task at all. It is the two days a file waits for a missing document. AI that spots the gap the moment the file arrives and drafts the request immediately can cut those two days to hours, even if no single step gets much faster. A step that is 30% faster inside one team is a local improvement. A handover that disappears changes how fast the whole company serves a client.
Ask the board question
When two candidates score about the same, ask one more question: which process, if it ran twice as fast, would your board actually notice?
The answer is usually a workflow that touches revenue or clients: the time from signed deal to first invoice, the time from application to advice, the time to answer a complaint. Choose that one. The first workflow has to change a number people care about, not only a number on a team dashboard.
Five mistakes when choosing
Choosing what is most visible, not what costs most. A chatbot on the website is visible. Checking documents in the back office costs more hours.
Choosing the favourite project. Every leadership team has one. Score it like every other candidate.
Choosing something rare. A process that happens a few times a year cannot be measured within weeks and will not prove anything quickly.
Choosing work that nobody owns. If no one in the team will own the new workflow, it will not survive the first problem.
Choosing work with locked data. If the information sits in a system you cannot connect, the build turns into an integration project. Pick something else first.
After you choose
Take a baseline of the chosen workflows if you do not have one yet, then design the new version next to the current one. Mark each step as automated, AI with a human check, or a tool or instruction. Read how to measure AI ROI for the measurement, and why AI pilots fail for the traps between design and daily use.
Frequently asked questions
How many workflows should we start with?
Two or three. Fewer gives you too little proof; more will not go live within the first sprint.
What if our biggest workflow is also the riskiest?
Split it. Let AI take the preparation steps, such as collecting, reading and checking, and keep the decision with people. You free most of the hours without handing over the judgement.
Do we need clean data first?
No. You need the data that one workflow uses, in a system you can connect. Clean what that workflow needs; leave the rest for later.
Which departments usually have the best first candidates?
The ones with high volumes and clear rules: client service, operations and finance. In regulated businesses, the back office often has more hours to free than the front office.
Want help choosing your first workflows? Book a 30-minute call. For the full approach, read where to start with AI in a mid-size company.
All examples in this article are illustrative.