AI for Business4 min read

An AI Roadmap Without a Stop-Doing List Is Just More Work

Pair every AI investment with a decision to retire, redesign, protect or stop work, so the roadmap changes the operating model instead of adding tasks.

Bokili EditorialPublished October 27, 2026· Verified October 27, 2026

AI initiatives split into four work decisions: retire, redesign, protect and stop

Most AI roadmaps are good at adding work. They list pilots, tools, training, governance reviews and launch dates. What they rarely say is which old task will disappear, which handoff will change, which human judgement must stay, or which use case should stop. That omission turns an AI programme into a second operating system laid on top of the first.

A useful roadmap therefore needs a stop-doing list. For every AI investment, make one explicit work decision: retire, redesign, protect or stop. This is not a cost-cutting shortcut. It is a way to make the operating model visible before enthusiasm hardens into permanent extra work.

Why adoption can increase the workload

A new AI tool changes more than the step where someone types a prompt. It can create new review queues, exception handling, access requests, policy checks and training needs. The UK Government’s People Factor guide describes AI implementation as both a technical transformation and a social transition. It also stresses that tools must be embedded in workflows and team processes, not simply made available.

That matters because a busy roadmap can hide duplicate work. A team may keep the old template, add an AI draft, add a human review and add a log for assurance. Nothing is removed. The output arrives faster, but total work may grow. Before claiming progress, trace the full workflow from request to decision.

	Addition-only roadmap	Work-decision roadmap
Success question	Did people adopt the tool?	What changed in the work?
Old process	Usually remains in place	Retired or redesigned deliberately
Human role	Added as a generic reviewer	Protected where judgement changes the outcome
Failure response	More guidance or prompts	Stop, narrow or redesign the use case

Add four decisions to every AI roadmap

The Retire–Redesign–Protect–Stop card

Retire

Name the manual step, duplicate record or approval that can end after the new method meets an agreed evidence threshold.

Redesign

Show the handoff that must change. Assign an owner, update the standard process and remove the old route.

Protect

Name the human judgement, relationship or control that remains essential. Give that person enough time, context and authority to act.

Stop

Write the condition that pauses or ends the use case: weak evidence, excessive exceptions, unacceptable risk or no meaningful benefit.

The four choices are intentionally different. Retire removes redundant work. Redesign changes how work moves. Protect preserves a valuable control. Stop prevents an unsuitable use case from consuming more attention. A single initiative may need all four at different points.

A worked example: monthly management reporting

Imagine a finance team pilots AI to draft the narrative for a monthly management report. The roadmap includes tool access, a prompt library, training and a three-month evaluation. That sounds complete, but it says nothing about the current process.

Reading is a start. Practice makes it stick.

Start learning

Turn the pilot into an operating decision

1
Map the current work
An analyst copies approved figures into a template, drafts commentary, checks unusual movements, sends the file to a manager and stores the final version.
2
Retire one step
If three consecutive cycles meet the agreed accuracy and review-time thresholds, retire manual first-draft commentary for standard sections.
3
Redesign the handoff
Feed only approved figures into the drafting step, then route exceptions and material movements directly to the analyst instead of reviewing every sentence equally.
4
Protect judgement
Keep the analyst responsible for explaining causes, uncertainty and decisions. The model may draft language; it does not own the interpretation.
5
Write the stop rule
Pause the workflow if the source figures cannot be traced, review time rises, or the exception rate stays above the agreed limit.

This example separates the tool from the change. The AI draft is only one component. The real design work is deciding what ends, what moves, what remains human and what evidence would make the team stop.

Use evidence before you remove work

A stop-doing list should not become an order to remove controls before the replacement works. NIST’s AI Risk Management Framework playbook recommends clear roles, responsibilities and ongoing review across the AI lifecycle. The UK Government AI Playbook likewise emphasises understanding risks, assuring systems and monitoring their use. In practice, retirement should follow evidence, not precede it.

Evidence to record before retiring a step

The task and decision boundary are clear.
Inputs can be traced to approved sources.
Output quality has been tested on normal and difficult cases.
The person reviewing the result has time and authority to intervene.
Exceptions have an owner and a route outside the AI workflow.
The old process has a named retirement date or a documented reason to remain.

If these conditions fail, the right decision may be redesign or stop rather than retire. That is useful information. A roadmap becomes stronger when it can reject an attractive tool that does not improve the system.

Make the stop-doing list part of governance

Review the work-decision card at the same forum that approves funding, risk treatment and scale. Do not leave it as a local team note. Ask the initiative owner to show the baseline workflow, the proposed change, the evidence threshold and the owner of each exception.

This also connects the roadmap to related decisions. Before scaling, define exit rules for the pilot. When time is saved, decide what that capacity becomes. If the pilot exposes a weak handoff, treat it as a workflow problem to fix, not as a prompt problem.

Write one stop-doing card

Choose one live AI initiative.
Write the current task from request to decision in five lines.
Name one step to retire, one handoff to redesign and one human judgement to protect.
Write one measurable stop condition.
Assign an owner and review date to each decision.

The test

If the roadmap cannot show what work changes, it is probably a technology list, not an operating plan.