Picking your first three workflows

The fastest way to stall an AI programme is to try to help everyone with everything. You end up running generic sessions, answering scattered questions, and reporting "usage is up" without being able to say what changed. The fix is boring and it works: pick three specific workflows, make them measurably better, and use them as the story for everything that follows.

A workflow, for this purpose, is a piece of work that a named group of people does repeatedly, with a recognisable start, a recognisable end, and an output someone else uses. "Writing" is not a workflow. "Drafting the first version of the monthly client report from the account notes" is.

This playbook takes about two weeks alongside everything else you're doing.

1. 1

   Step 1: collect candidates

   from the people who do the work.

2. 2

   Step 2: score each one

   on frequency, pain, fit and visibility.

3. 3

   Step 3: choose three

   , and make them different.

4. 4

   Step 4: baseline

   before you touch anything.

5. 5

   Step 5: write a one-paragraph brief

   for each.

Step 1: collect candidates from the people who do the work

Don't start from a vendor's list of use cases or a leadership wish list. Both are useful later, as a sense check. Start from what people actually spend their week on.

Good sources, roughly in order of usefulness:

• Short conversations. Thirty minutes with people across teams. Ask: "What's a task you do every week that you'd happily never do again?" and "Walk me through the last time you used AI for work." The second question shows you what's already happening informally.
• Team leads. Ask each one: "What does your team do repeatedly that eats time but doesn't need your best people's judgement?"
• Existing complaints. Backlogs, slow turnaround times, the reports that always arrive late. Pain that's already visible is easier to fix in public.
• Your enthusiasts. They've probably already found two or three things that work. Don't assume those are the best candidates, but they're tested.

Write every candidate down in the same format: who does it, how often, what goes in, what comes out, and who uses the output. If you can't fill in all five, it isn't specific enough yet. Go back and ask.

Aim for twenty to forty candidates. Fewer and you haven't looked widely enough; more and you're probably splitting hairs.

Step 2: score each one on four things

Score every candidate from 1 (low) to 3 (high) on:

• Frequency: how often it happens multiplied by how many people do it. A weekly task done by forty people beats a daily task done by one.
• Pain: how slow, tedious or error-prone it is now. Ask the people doing it, not their manager.
• Fit: whether current AI tools are genuinely good at the core of it. Drafting, summarising, restructuring, extracting from messy text, first-pass analysis and rewriting for a different audience usually score well. Exact calculations, facts nobody will check, anything needing live access to systems the tool can't reach, and decisions with legal weight usually score badly.
• Visibility: whether someone senior would notice if it improved. This isn't vanity. Your first three are also your evidence, and evidence nobody sees doesn't buy you the next quarter.

Then apply two filters before you add up scores:

• Data filter. Does doing this with AI require pasting in anything your policy (or the absence of one) makes risky: customer personal data, unreleased financials, HR cases? If so, park it until that's resolved.
• Owner filter. Is there a manager who wants this to work? A great workflow in a team whose manager is indifferent will underperform a decent one in a team whose manager is keen.

The workflow prioritiser does this scoring for you and sorts the list, if you'd rather not build the spreadsheet.

Step 3: choose three, and make them different

Take the top scorers, but don't just take the top three numbers. Look for a spread:

• One quick win. High fit, low risk, done by a team that's ready. You want something showing results within a fortnight.
• One high-frequency workflow. Lots of people, done often. This is where the aggregate time adds up and where the measurement story is strongest.
• One visible workflow. Something leadership already cares about, like a report they read or a turnaround time they complain about.

Sometimes one workflow covers two of these. That's fine. What you're avoiding is three variations of the same thing in the same team, which teaches you very little about the rest of the organisation.

Resist a fourth. You'll be tempted, especially when someone senior suggests one. Write it on the "next" list and say so:

That's a strong candidate for the next round. We're focused on three until we can show results.

Step 4: baseline before you touch anything

For each of the three, go back to the people who do it and agree a baseline. You need:

• Time per instance, as an honest estimate from the people doing it. "About three hours" is fine. Ask two or three people and use the middle answer.
• Volume: how many times a week or month, across how many people.
• Quality signal: whatever already tells you if it's good. Rework rates, number of review rounds, complaints, turnaround time. Use something that exists rather than inventing a new metric.

Write these down with the date and who told you.

Step 5: write a one-paragraph brief for each

Before you start, write a short brief for each workflow and share it with the team's manager. Use this template:

Workflow: [specific description]
Who does it: [team, number of people]
Volume: [times per week or month]
Baseline: [time per instance, quality signal, source and date]
What AI does in it: [the specific step, e.g. first draft from notes]
What a person still does: [checks, judgement, sign-off]
What not to paste in: [anything off limits]
Success looks like: [the change you expect, stated plainly]
Owner in the team: [named person]

The "what a person still does" line matters more than it looks. It's the answer to the quality worry before anyone raises it, and it stops the workflow sliding into "the AI does it now" when nobody's checking.

Mistakes to skip

• Picking the most impressive demo. The workflow that looks best in a meeting is often a one-off. Pick what happens every week.
• Asking managers instead of doers. Managers know what they wish took less time. The people doing the work know what actually does.
• Ignoring fit because the pain is high. If the tool is bad at the core of the task, a painful workflow becomes a painful workflow with an unreliable assistant, and you've taught a team that AI doesn't work.
• Skipping the baseline because it feels obvious. It won't feel obvious in two months when someone asks what changed.
• Letting someone else's priority jump the queue without a trade. If a leader insists on a workflow, swap it in for one of yours rather than adding it.

At the end of two weeks you should have

• A list of twenty to forty candidate workflows, each written in the same five-part format
• Scores for each, with the data and owner filters applied
• Three chosen workflows with a deliberate spread: quick, frequent, visible
• A baseline for each, with sources and dates
• A one-paragraph brief for each, agreed with the team's manager
• A "next" list you can point to when new requests arrive

Those three workflows become your first sprints, your first guides and your first report. If you want to score your list without building a spreadsheet, the workflow prioritiser is free. The Role Book picks up from here, with the sprint plan, the workflow deep-dive interview and the guide template for turning each of the three into something the next team can pick up without you.