Most internal AI projects start as a backlog item and end as a slide. A sponsor asks for a copilot. A squad estimates a quarter. The quarter becomes two, and the workflow that was meant to change is still the old one, with a prototype in a side folder.
A design sprint is a way to refuse that shape. You take one workflow, you give it one week, and you end with a decision to keep it, change it or stop it. AI sits inside the week as a way to research, draft and prototype faster. It is not the thing the room is there to admire. The AI design sprint we run with teams is this shape with a facilitator. You can run a thinner version inside the company if you are strict about the boundaries.
Before Monday
The week fails in the fortnight before it, when the brief is still a theme. Write five lines and get a sponsor to sign them.
- The workflow, in the operator's words, and how often it happens.
- The person who feels the pain, and who will try the new version with you during the week.
- The decision the week must produce, written as a question with three allowed answers: adopt, adapt or stop.
- The data you are allowed to touch, agreed with whoever owns the rule, before anyone pastes a record into a tool.
- The stop rule: the evidence that would make you end the week early rather than force a demo.
If you cannot name the operator, you are about to design for a persona. If you cannot name the data rule, you are about to create a governance incident in the name of speed. Fix both before you book the room.
Who is in the room
Keep it small. The operator. Someone who understands the customer or the downstream handoff. Someone who can build a thin prototype, which might be a script, a form or a well made prompt in the approved tool, not a new platform. A decider who can say yes or no on Friday without taking it to a committee that was not there. A facilitator who is not also the person in love with the idea. Five to seven people. More than that and you have a workshop about a sprint.
The week
You do not need theatre. You need a sequence that ends in a changed workflow, even if the change is to leave the old one alone.
- Monday is the map. Walk the current workflow step by step with the operator. Mark time, handoffs and the step where quality usually fails. Choose one step to change, not the whole chain.
- Tuesday is the options. Use a model to widen the ways that step could be done, then throw out anything that breaks the data rule or the stop rule. Leave with two options, not twelve.
- Wednesday is the thin version. Build the smallest thing the operator can run on a real case: a template, a checklist, a draft that a human still sends. No new vendor.
- Thursday is the trial. Run the thin version on several real cases beside the old way. Keep both outputs. Note where the new one was faster, where it was wrong and where a person still had to intervene.
- Friday is the decision. Adopt, adapt or stop. Name the owner of next week. If the answer is adopt, write the new steps so someone who missed the sprint can follow them. If the answer is stop, write why, so the idea does not return as a fresh pilot next month.
AI is a tool on Tuesday and Wednesday, and a suspect on Thursday. The trial is there to catch the fluent mistake. If Thursday is a demo to stakeholders instead of a trial on cases, you have built a pitch, not a workflow.
What ships means
Shipping is not a slide and it is not a prototype that only the squad can run. Shipping is the operator using the new step the following week, with the check you discovered on Thursday written down. If the decider needs another month of business casing before anyone may try it, you were not in a sprint. You were in a pre meeting.
A stopped idea is a result if the stop is written down. Organisations get stuck when every AI concept stays alive because nobody wants to be the person who ended it. The useful part of the week is a kill you agreed to in the brief, while the room was still willing to hear bad news.
After the week
Book the review when you book the sprint. Ten working days later, thirty minutes, same decider. Did the operator run the new step. What broke. Does the stop rule still hold. Then either fold the step into the ordinary process or take it out. Do not let it live as a special project.
Run the next one on a different workflow, with a different operator, using the same five line brief. The method is what you can repeat. A single impressive Thursday is a demo. One week is enough when someone in the room can actually choose. It is not enough when the week is a performance of speed around a choice nobody is allowed to make.