- Case study
- Mapping Studio
- Subject
- AI for a growing logistics company
- File
- 01 of 03
A 20-hour specialist job, drafted in 8 minutes.
A senior analyst read one of my LinkedIn posts and asked me a simple question: could AI take on his team’s slowest job? This is the story of how a doubtful team ended up with an AI tool they now rely on in production.
- Client
- A fast-growing logistics company
- Users
- The integration analysts
- My role
- Discovery to delivery; I directed the build
- Built in
- 9 days
- Today
- In production on the company’s own servers
- Time
- 20+ hours replaced by 8 minutes
- A full mapping draft, then checked by a specialist.
- Measuredthe tool’s own timing
- Cost
- Under $5 a draft
- The budget was $14.
- Quality
- Over 80% right
- The bar was 60%; this is the analyst team’s own check.
Find
The ask: It started with a LinkedIn post.
A senior analyst at a fast-growing logistics company read it and called me. His company wanted to bring on new customers faster, and every onboarding started on his team’s desk with a job called mapping. They had already tried ChatGPT and Gemini on it. The files were simply too big: the chats ran out of room and handed back half answers.
He wanted to know if AI could do better, done properly.
Discover: “I watched a mapping happen before I designed anything.”
A mapping is a translation guide. A new customer describes a shipment in their language, the company describes it in its own, and an analyst matches every field between the two, by hand.
Company A
Company B
- Shipping date
- Consignee
- Pack code CT
- Weight
- Receiver
- Gross weight
- Despatch day
- Net weight
- Cartons
I sat with an analyst and watched one. One guide has 332 fields, and many of them are matched by meaning, not by name. An analyst put the cost plainly in a voice note during my research: 20 to 24 hours for a senior analyst, 28 to 32 for someone less experienced.
The voice note
Read the transcript
- My question
- I have another question. Previously, you mentioned it takes quite some time to finish a mapping. So, can you specifically tell me how much time will it take a senior business analyst to complete one single mapping?
- The analyst’s answer
- So, well, 8 to 10 years of experience as a business analyst counts as a senior business analyst. So for a senior business analyst, it would take somewhere around 20 to 24 hours on average for a mapping, and for a mid-level business analyst, it would be somewhere around 28 to 32 hours, yeah. I hope that answers.
Everyone downstream was waiting.
Mapping isn’t just slow. It comes first. Meetings with the customer, development, testing and go-live all wait for it, and the people who can do it well are rare.

Hover over a box, or tap its number, to read my handwriting.
Hover over a box, or tap its number, to read my handwriting.
Red: the steps I circled as the major pain point. BA: business analyst.
Read the whole sketch
- Tendering. Tendering process. The arrow to step 2 reads “wins tender”.
- EDI integration team.
- BA, requirement gathering. Requirement gathering: guidelines, formats, samples.
- BA, design, manual. Create mapping rules manually.
- Developer, build, manual. Developer writes the logic code and tests locally.
- BA, manual. Performs end-to-end testing; creates various test scenarios.
- Customer, manual. Performs UAT (user acceptance testing) and signs off UAT.
- DevOps, automated. Schedules go-live and deploys in production. The only automated step.
- BA, hypercare, manual. Actively monitors business flows during hypercare.
- Sticky note. Mapping: first bottleneck. Business pain point. No dev bandwidth.
- Sticky note. Mapping points: steps 2, 3 and 4, circled as major.
- 01Mapping
- 02Customer meetings
- 03Development
- 04Testing
- 05Go-live
Opportunity: Five ways to fix it. One worth building.
Mapping went first because it’s the first domino. The other steps are next.
Shape
Define: “They asked for 60%. I aimed for 97%.”
The senior analyst’s brief was clear: a mapping in under 50 minutes, at least 60% right, no silent errors, under $14 each. I set my own bar higher.
Time per draftin minutes
- Client’s ask: Under 50 min
- My target: 15 min
- Delivered: About 8 min
Qualitypercent right
- Client’s bar: 60%
- Team’s check: Over 80%
- My target: 97%, still the goal
We started with the most common message type, inbound, and drew hard lines around the tool. It drafts, but it never runs anything in production, never approves its own work, and never learns without a person signing off.
Design: “The doubt became the design brief.”
On day one, the analyst didn’t believe AI could do this. The team expected it could only match fields that already had the same name.
So I designed the tool to prove every answer. Beside each proposed match, it shows the evidence it used:
- the guideline’s own note on that field;
- the value from a real sample message;
- how an analyst mapped it before.
Nothing is invented at review time. The analyst checks the source, not the AI’s word. That panel won the sceptics over, and it showed the AI matching by meaning, not just by name.

- The proposed match
- The guideline’s own note
- A real sample value
Doubt is designed in, not hidden:
- When the tool can’t find a match, it says why.
- When it never reached a field, it says that too.
- Fields that need a business decision are marked for the analyst and shaded red in the export.
- The tool: The tool reads the documents
- The tool: The AI proposes each match, with its evidence
- The tool: Automatic checks
- The analyst: The analyst approves, fixes or rejects, with a reason
- The analyst: The spreadsheet goes to the developers
Make
Build: From first brief to a working tool in 9 days.
The tool reads the analysts’ own guideline files, PDFs and real sample messages. It writes its draft into the spreadsheet layout the team already hands to developers, so nothing downstream had to change.

Approve, fix or reject: every verdict carries a reason. 
Every edit visible: the AI’s original is always kept. 
Cost shown before you run: with a hard cap. 
Stop and resume: no work lost. 
A library that suggests: the right company guideline, never added on its own. - CCorrectPPartially correctXIncorrectMove between fieldsDrawn from the tool’s own key hints
Review from the keyboard: built for speed.
Speed came from a question, not more hardware. I asked why only some jobs were slow, found the cause, and a full draft went from 38 minutes to 8.
Test: Making sure good really meant good.
An automatic score can’t tell a good mapping from one that only looks right. So the experts graded the drafts, field by field, and I made sure the grading stayed honest: only the AI’s untouched work counts toward quality, and every fix had to prove itself on real cases before it shipped.

Launch
Launch: Live where it mattered most, first.
We launched with inbound messages, the type the team sees most. Within the first week, analysts were testing it on real work. We demonstrated it to senior management on proven files only. Then came the moment that counted: the senior analyst tested a draft, made a few changes, signed it off and sent it to the developers. The first AI-drafted mapping was in production.
- First prototype
- Pilot with analysts
- Demo to senior management
- First mapping in production
Adopt: “The output was good. The room was quiet.”
When I first showed the prototype, the analysts said the information was good. Then they went quiet. They weren’t using it. Something about it confused them, even though the tool was doing its job.
So I went back and watched them work. All day, every day, they live in one piece of software, with its colours and its type. Mine looked nothing like it. I took inspiration from their tool, matching its colours and fonts, and rebuilt mine to feel like home.
It worked. The analysts felt it was their own software. They started reading every draft and using it actively. I then trained the first analysts myself: what each screen does, and what to do when something goes wrong.


Run and grow
Improve: Live use kept making it better.
Every week of real use taught us something. The tool changed in response:
- Errors in plain wordsThey name the file and the fix.
- Honest gradingReviews record exactly what was graded.
- A cleaner exportWritten to the analysts’ standard.
- A better editorReworked from live analyst use.
- A smarter libraryIt suggests, never adds.
Then one of the client’s own customers asked a fair question: how do we know our data stays safe? I decided to move the AI off the cloud, and the company’s IT team deployed it on their own servers with a locally hosted model. The answers stayed the same when we re-ran the test files. We traded a little speed for complete control of the data, and the company is upgrading its hardware to win the speed back.
Scale: Built to outlast me.
I handed it over in person to the company’s AI architect, walked him through every part of it, and stayed on call online and on site. He runs it now.
It was built to grow:
- Four readers cover every combination of message formats the company deals with: 15 in all.
- The company’s own guidelines are stored once, so every new customer starts half done.
| FromTo | X12 | EDIFACT | XML | JSON |
|---|---|---|---|---|
| X12 | Covered | Covered | Covered | Covered |
| EDIFACT | Covered | Covered | Covered | Covered |
| XML | Covered | Covered | Covered | Covered |
| JSON | Covered | Covered | Covered | Can’t occur |
Value: What changed.
| Before | With Mapping Studio | |
|---|---|---|
| A full mapping draft | 20 to 24 hours | About 8 minutes, then checked by a specialist |
| Cost | Budget: $14 a mapping | Under $5 a draft |
| Quality | Bar: 60% right | Over 80% right, by the team’s own check |
| Where it runs | Nothing | In production, on the company’s own servers |
Next to measure: how long a review takes, and how much sooner new customers go live. Mapping was the first bottleneck, so onboarding should move faster; that’s the number worth proving next.
- Runs on the company’s own servers.
- Cost capped before every run.
- No company names inside the tool.
- Every change can be rolled back.
What’s next: The developer’s desk.
After a mapping is signed off, a developer codes it into the system. With today’s coding AI, that’s the natural next step, and the same approach would carry straight over.
I’d also add evaluation agents, AI that checks each draft before an analyst sees it, now that the tool runs on the company’s own servers and the cost of an extra check is no longer the limit.
And the lesson I’d carry to any team: people trust an AI tool that feels like the one they already use, and that shows its evidence instead of asking to be believed.
- 01Mapping
- 02Customer meetings
- 03DevelopmentNext
- 04Testing
- 05Go-live
How these numbers were measured.
- 20 to 24 hours: an analyst’s estimate for a senior analyst, recorded during research.
- 8 minutes: the tool’s own timing on a full 332-field draft, on the original cloud setup.
- Under $5: the measured cost of a full run.
- $14, 50 minutes and 60%: the client’s brief.
- Over 80%: the analyst team’s own check.