Skip to content
AI explainer series

The AI isn’t the project.
The workflow is.

When an AI rollout stalls, the build is rarely the problem. Nobody mapped the process first, and nobody planned the month after launch. Workflow first fixes both, before any AI goes in.

  1. Before
    Before the build

    Map it with the people who run it.

    Forty steps came down to eighteen before anything was built.

  2. The build
    What most people call the project

    Build the four steps that repeat.

    The steps that happen the same way every time. Everything that needs a judgement call stays with people.

  3. The month after
    The half that often gets skipped

    Stay with them while they use it.

    A demo, a pilot, a first version and a week in the room. This is what decides whether it gets used.

One whiteboard, three times: mapped and cut, built, then in use.
A project drawn as one bar in three segments, with the same whiteboard above each. Before the build: the board holds ten notes, five of them cut to empty dashed outlines and one marked where it stalls, and forty steps come down to eighteen. The build, which is what most people call the project: the same board with the cut notes gone and the AI mark on two of the five left, the four steps that repeat. The month after, the half that often gets skipped: the same board in use, shaded green, with a tick and two people beside it, for a demo, a pilot, a first version and a week in the room.
How it works

The process exists.
Nobody ever wrote it down.

Most processes live in people’s heads, because writing them down was never anyone’s job. Each person knows their own part well. Almost nobody knows how the parts fit together.

  • SalesCloses the deal
    1. Agree the deal
    2. Get it signed off
    3. Send it to legal
    4. After that, not mine

    Three steps. It ends the moment it leaves my hands.

  • LegalWrites the agreement
    1. Read the deal
    2. Ask for the company nameback to sales
    3. Ask for the lengthback to sales
    4. Ask for the priceback to sales
    5. Ask who signsback to sales
    6. Write the agreement

    Six steps, and four of them are me asking for something.

  • OpsKeeps the record straight
    1. The intake formstarts here
    2. Chase what is missingtwice

    It starts at the intake form. Which nobody fills in.

  • FinanceInvoices it
    1. Before thissomebody else’s
    2. Check it’s signed
    3. Invoice it

    Two steps. I see it when it’s already done.

Only one of the four maps includes the intake form, and that’s exactly where the whole thing stalls.

One handoff, drawn by four people. Each map is right about its own part, and none of them match.
The same handoff as four people draw it, each as a short column of steps. Sales draws three: agree the deal, get it signed off, send it to legal, and then a dashed question mark, after that, not mine. Legal draws six: read the deal, four steps each asking sales for a fact that was missing, the company name, the contract length, the price and who signs, and then write the agreement. Ops draws two: the intake form, marked as where it starts, and chasing what is missing, twice. Finance draws two, checking it is signed and invoicing it, under a dashed block reading before this, somebody else’s. A note underneath says the intake form appears on one of the four maps and is also where the whole thing stalls.

Ask four people how a process works and you’ll get four different answers, each one right about its own part.
Now put AI on top of that before anyone writes it down.

The problem

The pain showed up in legal.
The cause was one step earlier.

A company we worked with asked for AI in its legal team. The real problem was intake, one step earlier, where thirty salespeople handed over deals thirty different ways. Look one step upstream of wherever the pain shows up.

What arrives from sales
  • An email thread“Closed, see below”
    • Company name
    • Length
    • Price
    • Who signs
  • A voice noteFour minutes long
    • Length
    • Price
  • A message“Can we send the contract today”
and twenty-seven more, no two alike
One legal desk A week of chasing The same four questions, thirty times
  • What’s the company’s legal name?×30
  • How long does the contract run?×30
  • What did we agree on price?×30
  • Who signs?×30
Three of the thirty handovers. Each is missing something, so the same four questions go back every time.
On one side, what arrives from sales: an email thread carrying only the company name, a four-minute voice note carrying the contract length and the pricing, a message carrying none of the four facts, and a note saying twenty-seven more, no two alike. The first note names all four facts, the one it has and the three it lacks; on the other two only the facts they carry are written, and the missing ones are grey bars. On the other, one legal desk with a week of chasing on it, sending the same four questions back every time: the company’s legal name, how long the contract runs, what was agreed on price, and who signs.

AI would have helped legal chase faster. The real fix was making sure there was nothing to chase.
That fix starts on a whiteboard.

The shift

Map the process. Cut it.
Then decide where AI goes.

Draw every step on a whiteboard with the people who do them. In this example, more than half existed for a system switched off years ago or a form nobody removed. Forty steps became eighteen, with no AI yet.

The whiteboard, sales to legal 40steps drawn18left, and no AI yet
  • 3Where it stallsThe intake form nobody fills in
  • 9Where it stallsContract edits going back and forth by email
  • 18Where it stallsChasing the right person to sign
Forty steps on the board, twenty-two cut, eighteen left, and three places where the work stalls.
The whole handoff from sales to legal drawn as forty numbered notes on a whiteboard, ten to a row. Twenty-two of them are cut and left as empty dashed outlines, because they exist for a system that was switched off or a form nobody removed, which leaves eighteen. Each note that is left carries a grey line standing for what was written on it. Three of the survivors are salmon notes with a pin, the places the work stalls: step three, the intake form nobody fills in, step nine, contract edits going back and forth by email, and step eighteen, chasing the right person to sign.

No AI yet, and the process is already less than half the size.
Now sort the eighteen that are left.

Inside the workflow

Eighteen steps left.
Four of them are the same every time.

Four of the eighteen happen the same way on every deal: reading the notes, filling the template, drafting the summary, preparing it for signing. Those four go to AI, because that’s what it’s good at. Anything that needs judgement or a signature stays with a person.

The same every timeReading, matching, drafting. Four steps.
  • 3
    Read the deal notes into the intake formWhatever shape they arrived in
    AIreads
  • 6
    Fill in the contract templateThe four facts from intake, in the same places every time
    AIfills
  • 10
    Draft the agreement summaryA first version, for a person to correct
    AIdrafts
  • 37
    Prepare the documents for signingThe right documents, in the right order
    AIprepares
Needs a decision or a signatureFourteen steps, and none of them go to AI.
  • 14
    Approve the termsSomebody is accountable for what was agreed
    legal
  • 9
    Negotiate the contract editsA conversation, not a document
    legal
  • 40
    Sign itA name on a page, and it has to be theirs
    signatory
  • And eleven moreChecks, calls and chasing, left exactly where they were
Four steps go to AI. Fourteen stay with people.
The eighteen surviving steps fall into two piles. In the first, the four that happen the same way every time and go to AI: reading the deal notes into the intake form, filling in the contract template, drafting the agreement summary, and preparing the documents for signing. In the second, the fourteen that need a decision or a signature and do not move: approving the terms, negotiating the contract edits, signing, and eleven more checks, calls and judgement calls. Every step carries the number of its note on the whiteboard: the AI pile is steps 3, 6, 10 and 37, one of them the intake stall it clears; the people pile keeps 9, 14 and 40 and eleven others, among them step 18, chasing the signer.

That’s the whole build. The harder part is getting people to use it.
Here’s what that takes.

The rollout

An hour of training isn’t a rollout.
A week of somebody in the room is.

The build works, and nobody has used it yet. In a 2024 BCG survey, about 70% of the problems companies hit with AI came from people and process, not technology. So the rollout starts before anyone gets a login.

The usual rollout
  • A licence eachand a one-hour training invite
  • Nothing scheduled after it
  • Day 30Back to the way it was done before
Workflow first
  • A demoFive people watch it run on a deal they know
  • A pilotOne region, two weeks, and the map corrected in three places
  • The first versionWith a note saying what changed and why
  • A week in the roomOne training, then every question becomes a change
  • Day 30The second region asks for it
The same build, rolled out two ways. One slides back to the old habits by day 30. The other gives the change a real chance to stick.
Thirty days, drawn twice from the same finished build. On the first track, AI on the mess: a licence each and a one-hour training invite on day one, nothing scheduled after it, and on day thirty everyone is back to the way it was done before. On the second, workflow first: a demo where five people watch it run on a deal they know, a two-week pilot in one region that corrects the map in three places, a first version shipped with a note saying what changed and why, one training and then a week of somebody sitting in the room where every question becomes a change, and on day thirty the second region asking when they get it.

That week in the room is easy to leave out of the budget, and then a one-hour training ends up being the whole rollout.
Plan the week, and the tool gets used.

Over time

The map doesn’t go on a shelf.
It changes with every update.

Every update ships with a short dated note on what changed and why, and the whiteboard changes with it. Fewer steps and fewer stalls usually mean faster work, and that’s what people notice first.

Sales to legal: change log40 drawn18 at launch17 today
WhenWhat changed, and whyStepsStalls
LaunchDay 1The intake takes the four facts and the signer in one form. AI reads the notes, fills in the template, drafts the summary and prepares it for signing.181
Pilot fixWeek 2The pilot region corrects the map in three places: two steps nobody had drawn, and one drawn twice.191
UpdateMonth 2Contract edits move into one shared draft, so the three email steps become one, and the last stall goes.170
Region twoMonth 4The second region goes live on the same map, with the same note and one training.170
Four updates in four months, each with a note. The step count goes 18, 19, 17, 17, and the last stall is gone by month two.
The change log for the sales to legal process, headed forty steps drawn, eighteen at launch and seventeen today. At launch, day one: the intake takes the four facts and the signer in one form, and AI reads the notes, fills in the template, drafts the summary and prepares it for signing; eighteen steps, one stall left. The pilot fix, week two: the pilot region corrects the map in three places, two steps nobody had drawn and one drawn twice; nineteen steps, one stall. An update in month two: contract edits move into one shared draft, so three email steps become one and the last stall goes; seventeen steps, no stalls. Region two, month four: the second region goes live on the same map, with the same note and one training; seventeen steps, no stalls.

The count keeps moving after launch, and every change has a note behind it.
So when the next team asks for it, the map is still accurate.

Common myths

Four myths about fixing the process first.

You might think

“Documenting the process first slows us down.”

Actually

Drawing the map takes a day or two with the right people in the room. It’s the shortest part of the project, and it decides how long everything after it takes.

You might think

“We just need better training.”

Actually

Training takes an hour. Getting people to use something takes weeks, and it starts with a demo before the tool is finished.

You might think

“AI will learn our process from the data.”

Actually

Your data shows what people do today, workarounds included. Learn from that and the AI copies the workarounds, just faster.

You might think

“Getting people to use it is HR’s job.”

Actually

It belongs to whoever leads the change, and it starts before the build, not after.

Keeping it honest

Four rules
for doing it in the right order.

Building the tool is rarely where things go wrong. Buying or building it before anyone has looked at the process is. These four rules are about doing things in the right order.

01

Map before you buy.

Buy a tool before the map and you’re betting on a process nobody has looked at. The bet is usually on the wrong step.

02

Cut before you automate.

Automate a step that shouldn’t exist and it becomes permanent, because now other things depend on it.

03

People keep the judgement calls and signatures.

If someone is accountable for a decision, it stays with them, however capable the software is.

04

Ship it in updates, each with a note.

There’s no final version, just the next update and a dated note saying what changed. Small, dated updates are easier to follow than one big rewrite nobody reads.

How to start

Start with the process
people complain about most.

You don’t need a workshop to pick a starting point. Ask what slows people down, and the same process comes up more than once. Give it three weeks to map it, cut it and try it with the team.

  1. Day one

    Pick the process everyone complains about.

    Ask whoever owns it for the documentation. What comes back, or doesn’t, tells you where to begin.

  2. Week one

    Map it on a whiteboard with the people who do it.

    Count the steps and mark where it stalls. It takes a day or two, and that count becomes the number you measure everything against.

  3. Weeks two and three

    Try the cut version on real work.

    Run it on deals the team recognises, show it to them, and write down what they say. If you’d like a hand with this part, it’s what an AI Jumpstart is for.

Next step

Stop guessing where AI belongs.
We’ll show you.

An AI Jumpstart starts where this page does: mapping the work with the people who do it. Then we build the trimmed-down version on your records and stay with the team through the first version. It takes two to four weeks.

See the AI JumpstartBook a free strategy call