Quick Navigation
- Before we start — why adding AI to a bad process disappoints
- Why prompt-first thinking is too small
- The workflow lens
- Where AI belongs in a process
- Designing handoffs and review
- Exceptions, failures, and recovery
- The redesign sequence
- How to redesign without breaking trust
- What to automate last
- The workflow loop in one diagram
- Cheat sheet
Before we start — why adding AI to a bad process disappoints
There is a pattern that shows up in almost every early AI programme.
A team takes a messy process, adds an AI step somewhere in the middle, and expects the workflow to feel transformed. Instead, the same confusion remains. Ownership is still unclear. Inputs are still inconsistent. Exceptions still fall through gaps. The only difference is that one part of the mess now happens faster.
This is not an AI problem. It is a design problem.
AI workflow design begins with a hard truth: technology amplifies process quality. If the process is clear, AI can accelerate it. If the process is broken, AI can industrialise the confusion.
Why prompt-first thinking is too small
Teams often start with the wrong question: "What prompt should we use?"
That question matters, but it is too narrow. A prompt is only one step inside a larger workflow. If the surrounding process is unclear, the prompt will not save it. You still need to know who starts the task, what context is available, where review happens, how exceptions are handled, and what success looks like.
AI creates the most value when it is designed into the flow of work, not bolted onto one isolated activity.
The workflow lens
Every useful AI workflow has five parts:
| Part | Question |
|---|---|
| Trigger | What starts the work? |
| Context | What information does the system need? |
| Transformation | What does AI actually do? |
| Review | Who checks, approves, or edits? |
| Outcome | What changes as a result? |
This is the shift from toy use to operational use. Instead of admiring what AI can generate, you design how work moves from input to decision.
flowchart LR
A["Trigger"] --> B["Context"]
B --> C["AI Transformation"]
C --> D["Human Review"]
D --> E["Outcome"]
D --> F["Rework / Escalation"]
Where AI belongs in a process
AI is usually strongest in one of four workflow roles:
- Drafting: first-pass creation of text, summaries, or options
- Extraction: pulling structure from messy inputs
- Classification: sorting, tagging, routing, prioritising
- Assistance: recommending next actions for human judgment
It is usually weakest where the work depends on final accountability, ambiguous exceptions, or consequences the system cannot fully understand.
The design question is not "Can AI do this step?" It is "Should AI own this step, support it, or stay out of it?"
Designing handoffs and review
The handoff is where trust is won or lost.
If AI output arrives as a black box, people either distrust it or over-trust it. If it arrives with enough context to evaluate quickly, it becomes genuinely useful.
Strong handoffs usually include:
- the source material or evidence
- the AI's output
- a reason or rationale where helpful
- a clear next action for the human
- a visible place to edit, approve, or reject
Good workflow design reduces the effort of review without removing review.
Try it yourself — The handoff friction test
Take one workflow and identify where AI output would land.
| Question | Your answer |
|---|---|
| What would the reviewer receive? | |
| Would the source material be visible? | |
| Could the reviewer understand why the output was produced? | |
| Is there a clear approve, edit, reject, or escalate action? |
If those answers are fuzzy, the workflow is not ready for trustworthy AI handoffs yet.
Exceptions, failures, and recovery
Most workflow design failures come from designing only for the happy path.
Real work has missing data, contradictory instructions, unclear ownership, low-confidence outputs, urgent edge cases, and inputs that do not fit the template. If you do not design for exceptions, humans end up cleaning up chaos after the fact.
Every AI workflow should answer three questions in advance:
- What should happen if the model is uncertain?
- What should happen if the output is obviously wrong?
- Who owns recovery when the workflow fails?
Operational maturity is not eliminating exceptions. It is handling them deliberately.
The redesign sequence
One reason redesign efforts stall is that teams try to redesign everything at once.
A more reliable sequence is:
- Map the current workflow as it actually happens.
- Identify friction, delay, rework, and repetitive judgment.
- Decide which steps AI should support, not just where it technically could.
- Design review, exception handling, and ownership.
- Pilot the redesigned flow with a narrow scope.
This sequence matters because process redesign is not only about inserting AI capability. It is about rebalancing the workflow around new strengths and new risks.
How to redesign without breaking trust
The fastest way to create resistance is to redesign a process around AI without involving the people who actually do the work.
Trust grows when people can see how the workflow reduces friction, preserves accountability, and leaves room for human judgment where it matters. That is why the best redesign efforts are collaborative. The people closest to the work usually know exactly where AI can help and exactly where it should not be allowed to take over.
The redesign principle is simple: remove effort, not responsibility.
What to automate last
Teams often want to automate the final action because that is where the visible labour sits. Usually that should be the last thing you automate, not the first.
Why? Because the final action is where accountability, exception handling, and downstream consequence converge.
In many workflows, the best early automation targets are:
- preparation
- summarisation
- extraction
- classification
- drafting
- recommendation
These reduce effort while still preserving a human checkpoint. Once the workflow is stable, the evidence is strong, and the failure modes are understood, then deeper automation can be considered where appropriate.
The maturity sequence matters. Safe automation is usually earned in stages.
The workflow loop in one diagram
flowchart TD
A["Input / Trigger"] --> B["Context Assembly"]
B --> C["AI Step"]
C --> D{"Review needed?"}
D -->|"Yes"| E["Human Review / Edit / Approve"]
D -->|"No, low risk"| F["Automated Completion"]
E --> G["Outcome"]
F --> G
C --> H["Exception Handling"]
AI workflow design is the discipline of deciding where automation helps, where humans stay in control, and how work keeps moving when reality gets messy.
Cheat sheet
| Question | Good default |
|---|---|
| What should I redesign first? | High-volume, repetitive, review-heavy workflows |
| What makes an AI workflow usable? | Clear context, visible handoff, human review, exception routing |
| What causes most failures? | Designing only for the happy path |
| What is the core principle? | Remove effort, not responsibility |