Frameworks & Templates4 min read

The AI Workflow Friction Test: Remove, Preserve or Strengthen?

Do not remove every step from an AI workflow. Classify friction as waste, learning, safety or coordination before changing it.

Bokili EditorialPublished September 25, 2026· Verified September 25, 2026

An AI-assisted workflow moving from tangled waste through learning, safety and coordination checkpoints.

Speed is not the only design goal

The right AI workflow does not remove every pause. It removes steps that add no value, preserves steps that build judgement, strengthens steps that prevent harm and clarifies steps that transfer responsibility.

AI makes it tempting to treat every click, handoff and review as friction to eliminate. A team automates a first draft, then asks why the approval, source check or colleague review cannot disappear as well. The result may be faster on paper while becoming harder to trust, learn from or recover when something goes wrong.

Not all friction is the same. Some of it is waste: repeated copying, unnecessary formatting or waiting for information that already exists. Other friction is productive. It forces a person to inspect evidence, practise a judgement, protect a boundary or make ownership explicit. The design task is to tell those types apart before optimising the workflow.

Run the AI workflow friction test

NIST’s AI Risk Management Framework playbook recommends defining acceptable limits, testing whether systems are fit for purpose and planning course correction. Its management guidance also calls for monitored response, recovery, override and change-management mechanisms. The UK Government AI Playbook adds clear roles, escalation routes and evaluation throughout the life cycle. None of those controls requires keeping every old process step. They require keeping the steps that have a clear job.

Four types of workflow friction

Remove waste

Delete effort that neither changes the output nor protects a decision: duplicate entry, repeated formatting, avoidable searching and approvals with no defined criterion.

Preserve learning

Keep moments where a person must compare, explain or revise. These steps build the judgement that lets the team detect a weak result later.

Strengthen safety

Add or improve checks where an error could expose data, mislead a customer, affect a person or create a difficult-to-reverse commitment.

Clarify coordination

Make handoffs lighter but explicit. Name what moves, what evidence travels with it, who decides next and where an unresolved issue goes.

A fast workflow can still contain deliberate pauses

	Blanket automation	Friction-aware design
Question	Can we remove this step?	What job does this step perform?
Review	Treated as delay	Matched to evidence, risk or authority
Learning	Assumed to happen elsewhere	Built into comparison and revision
Failure	Handled after the fact	Includes limits, override and recovery

A deliberate pause should have a named trigger and a visible output. “Manager review” is weak design because nobody knows what the manager must inspect. “Confirm the three source-backed figures and approve the external recommendation” is useful friction. It tells the reviewer what to do, creates evidence and can be shortened when the risk is low.

Reading is a start. Practice makes it stick.

Start learning

Worked example: an AI-assisted policy update

A people team uses AI to turn a long travel policy into a one-page employee update. The old process contains six steps: copy the policy into a working document, rewrite it, ask a colleague to proofread, ask payroll to check allowance figures, obtain HR approval and publish the update.

Classify each step before changing it

1
Remove
Stop copying and manually reformatting the policy. Let the approved tool produce the first structured draft from the current source.
2
Preserve
Keep the colleague comparison between the draft and the employee question. This develops task judgement and catches a summary that is accurate but unhelpful.
3
Strengthen
Require payroll to verify every allowance and effective date against the controlled policy. The check produces a short evidence record.
4
Clarify
Replace a vague final approval with a named HR owner who accepts the wording, unresolved exceptions and publication date.

The redesigned workflow has fewer mechanical steps and stronger decision points. It is faster because it removes copying and generic proofreading, not because it deletes every human action. If an allowance is wrong, the team can see which check failed and who owns the correction.

Use three questions before removing a step

The friction decision card

What observable value does this step add?
What error, learning need or handoff would disappear with it?
Could the same purpose be met with a lighter trigger or smaller record?
Who owns the decision to remove, preserve or strengthen it?
How will the team know whether the change improved the outcome?

A ten-minute friction audit

Mark one workflow in four colours

Choose one low-risk workflow that already uses AI and list every human and system step.
Mark waste that can be removed without losing evidence, learning, safety or ownership.
Mark learning steps where a person compares, explains or revises.
Mark safety steps tied to data, accuracy, fairness, external impact or hard-to-reverse action.
Mark coordination steps and rewrite each handoff as an input, owner, decision and next route.
Change only one step, then compare the outcome before changing the rest.

Use the workflow change-log guide to record the reason for each change, and the AI handoff card to keep context with the work. The acceptance-test guide helps define the checks that should survive optimisation. Bokili’s learning features support short practice around the judgement steps that remain.

Good AI workflow design does not celebrate the fewest steps. It creates the fewest steps needed for a useful, learnable and recoverable outcome. Remove waste boldly. Preserve judgement deliberately. Strengthen the boundaries that matter. Make every handoff earn its place.