AI Opportunity Sprint
Walter maps one costly workflow with your team, compares a process change, plain automation and AI, and puts a working prototype in front of the people who would use it. You end with a go or no-go decision, and it is useful even if nothing gets built.
- Duration
- 3 weeks
- Price model
- Fixed fee, quoted after the fit conversation
- Your sponsor's time
- About 4 hours
- Your team's time
- About 75 minutes each, 3 to 5 people
What it needs from you
- A sponsor who can make the decision at the end.
- Three to five people who do the work, for an interview and a prototype test.
- A way to see the systems and documents involved, with your people.
Need to get a sponsor on board? Download the overview (PDF)
How the three weeks run
- Week 1
Map the work
Walter: Kickoff, short interviews and a walkthrough where the work happens.
- Week 2
Decide what to test
Walter: Scores every option, including a process change, plain automation and no change.
- Week 3
Prototype and test
Walter: Builds a realistic prototype on sample or synthetic data and tests it with your people.
- Then
Go or no-go
Walter: The readout and a recommendation.
What you leave with
The AI Opportunity Map
Where time is lost, what to change, and whether AI, plain automation or a process change is the answer.
A scored shortlist
Every candidate, with the reason it won or lost.
A tested prototype
What your people did with it, and what they said.
A starter evaluation set
Examples with the right answers that any future build must pass.
A go or no-go decision
If it's a go, the pilot scope and the criteria it must meet to go live.
One workflow, three weeks, one decision
Pinecrest Supply is a fictional 40-person distributor whose orders arrive by phone, email and paper. These are the work products its Sprint produced, and the call made at each step.
- Organization
- Pinecrest Supply (fictional): a 40-person regional distributor of facility and janitorial supplies
- Sponsor
- Operations manager
- People interviewed
- Three customer service reps, the warehouse lead, one outside sales rep
- AI tools during the Sprint
- De-identified notes only. No customer or order data in any AI tool.
- The long-term goal
- An order is in the system within the hour it arrives, entered once, and the reps spend their day on customers instead of re-typing.
What each stage produced
- 45% Phone
- 35% Email
- 10% Paper
- 7% Texts from sales reps
- 3% Web portal
- 01Order arrives
Customer: Phones, emails, sends a sheet or texts a sales rep
Service rep: Takes phone orders on a notepad; lunch-hour calls wait (F05)
- 02Rep keys it inF01 · 21 h/wk
Service rep: Customer, ship-to, lines and quantities: 7 minutes each, 180 a week
- 03Pricing lookupF03 · 6 h/wk
Service rep: Looks up contract pricing and substitutions
Senior rep: Keeps the pricing rules in one spreadsheet
- 04Confirm with the customerF02 · 5.5 h/wk
Customer: Answers questions about items, quantities and pack sizes
Service rep: Calls or emails back about anything unclear
- 05Warehouse picksF04 · 3.75 h/wk
Warehouse: Picks from the keyed order; 9 wrong picks a week from keying errors
You see where the hours go, measured, before anyone picks a tool.
The reps were fast and careful. The hours were lost to the system around them, so that's what we redesign.
What we tested, and what we decided
| Task | Done | Done with help | Not done |
|---|---|---|---|
| Approve a clean email order | 4 | 0 | 0 |
| Catch a pack-size mix-up (12 vs 12-pack) | 3 | 1 | 0 |
| Handle a handwritten sheet with a crossed-out line | 2 | 2 | 0 |
All three reps wanted the flagged fields first. The senior rep asked that substitutions only ever be suggested. The warehouse lead asked for one queue, whatever the channel.
30 synthetic orders agreed by the senior rep: 14 typical emails, 8 hard ones (two orders in one PDF, handwritten sheets, cases and units mixed), 4 that must be held for a person, and 4 shaped like past credit memos.
- Recommendation
- Process change and simpler automation first (form, pricing tables, confirmation step, phone script and screen). AI capture with a rep's approval on email and printed orders only.
- Scope
- Email orders and typed or printed order sheets. Handwritten sheets stay with a rep while those customers move to the form.
- Controls
- Every order approved by a rep. Unknown customers, new ship-to addresses, unmatched products and unusual quantities are always held. Customer data stays in Pinecrest's systems.
- Exit criteria
- 96% line accuracy on a 100-order evaluation set. Median rep time under 2 minutes per email order for four weeks. Wrong picks from keying errors under 2 a week.
- Stop rule
- An unapproved order ever reaches the warehouse.
- Next decision
- Fund the pilot. Owner: operations manager. Revisit phone orders after the pilot, with the baseline the new screen will have produced.
If it's a go, Pilot to Production takes it into everyday use. Already have a pilot? It doesn't need a Sprint first.
How Pilot to Production worksCommon questions
Will the answer always be AI?
No, and we'll say so. Often the answer is a mix: a process change or plain automation first, and AI on the steps where the input varies too much for rules. The Map says which is which, and a person stays in the design.
Is the prototype production code?
No, deliberately. If the use case goes ahead, it is rebuilt with evaluation, review and monitoring.
Do we need clean data?
No. Knowing what state your data is in is part of what the Sprint tells you.
How do you handle our information?
By default, client information stays in the client's own systems and licensed tools. We agree the tools, access and safeguards with you before anything is built, and prototypes can run on sample or synthetic data. The specific terms are set out in the MSA and SOW.
How we approach systems and dataWant to find out if it's worth it?
A 30-minute conversation. No preparation required.