I made upa company.
To show how AI handles a company's documents, I needed a company. No client lends theirs for a demo, so I invented one and simulated three years of its work. Now it has 140 files, and like every company's files, they don't all agree.
- files
- 140 files
- languages
- 4 languages
- right answers, first build → careful
- 77%→98% right answers, first build → careful
Northwind Studio
Design agency · Rotterdam · since 2017
A studio of sixteen that makes logos, websites and app designs for software companies. Clients pay monthly or per project, mostly remote. They don't write code.
- Clients
- 10, from legal tech to energy
- Plans
- Starter, Studio, Partner
- Files
- Help pages to Slack threads
- Languages
- English, Dutch, German, Spanish
Not a real company. The people, the clients and the money are made up; any resemblance is an accident.
- OKOleksij van KoeropatkinFounder, studio director
- VWVladislava WolinskaHead of PR
- MRMaksim RyzjikovOperations and finance
- VPVladimir PosjivalovSenior lawyer
- SMSofie MilkovaLawyer
- VBVladislav BoerdonosAccount lead
- STSandra TsapljoekAccount lead
- MDMichiel DembitskiDesign systems lead
- VHVarvara HandogaSenior designer, brand
- VKVlas KoerbatskiSenior designer, web
- TPTomás PereiraSenior designer, product
- KMKwame MensahSenior designer, motion
- NHNoor HassanDesigner
- LPLeon PavlovDesigner
- EGEva de GrootDesigner
- YHYara HaddadJunior designer
Every change left old paper behind.
Prices went up, plans changed, one client got a favour. Nobody cleaned up the old files, like in every real company. That's exactly what trips up an AI.
- 2024Studio retainer €4,800 a month. Three revision rounds, €400 for an extra one.↳ The old price list never got deleted.
- Jan 2025New prices: Studio €5,200. Two revision rounds, €450 for an extra.↳ Last year's prices still apply to some clients.
- May 2025Harbor, the client portal, launches. In February 2026 its “Waiting for client” status becomes “In review”.↳ Older files still call it by its old name.
- Oct 2025A draft proposes Studio €6,200, Partner €12,000. Rejected on 4 November.↳ A price proposal that was turned down is still in the files.
- Dec 2025An urgent line for Studio clients, agreed in October, is reversed.↳ One meeting says yes, a later one says no.
- Jan 2026Studio €5,800. New Partner plan at €11,200. Annual discount 15% → 10%.↳ Three different prices for the same plan, all in the files.
- Mar 2026Rush fee 25% → 40%. A rush confirmed in February keeps 25%.↳ Which price applies depends on the date of an email.
- Mar 2026One client gets two months of hour rollover, once. Policy stays one month.↳ A one-off favour, written down in team chat. Not a new rule.
- Apr 2026Harbor is down for 3 hours 20 minutes. No files lost.↳ The facts are only in internal reports.
This is RAG.
Before the AI answers, it looks things up in your company's files. Then it answers from what it found, and only from that. That's the whole idea.
- 1Look it up
A client asks “what's your rush fee?”. The system searches all 140 files and picks the five passages most likely to hold the answer.
- 2Hand it over
Those five passages and the question go to the AI, with strict rules: use only this, say where each fact came from.
- 3Answer with receipts
The AI writes a short answer with a source on every fact. If the passages don't say, it says “I don't know”.
Why not just ChatGPT?
ChatGPT knows the internet. It doesn't know your price list, your contract terms or what you promised a client last Tuesday. Ask it anyway and it will guess, confidently.
Why not train an AI on our files?
Training is slow, costs real money, and is out of date the day your prices change. With RAG the AI reads the current files every time. Change a file, and the next answer changes.
Where it pays off
A support bot over your help pages. A sales assistant that knows every proposal. Search over the handbook, contracts and old email threads, so new hires stop asking the same five people.
The catch
A demo takes an afternoon. Making it right takes the rest: finding the right page among similar ones, picking this year's price over last year's, saying “I don't know” instead of guessing, and proving all of it with tests. That's what the bot below and the chapters after it are about.
Ask it. Check its homework.
A real AI answers from the files on the left. The quick version is how most people build it; the careful one is how I would. On my 65 test questions they got 77% and 98% right. The right side shows what each one did.
Two versions answer side by side: a quick one and a careful one. Same AI. Click a number in an answer to open the page it came from.
One part at a time.
Part 1 opens up the bot above, one piece at a time. Change it, ask again, see what happens. Part 2 gives it real jobs: strange files, customer records, sending emails, and what it all costs. Same company all the way through.
Part 1Answering from the files
- 1.1OpenWhere to CutEqual pieces or sections, and what a price table looks like after each.
- 1.2OpenWords or MeaningKeyword, vector, both, and a second look. Switch the parts, ask again.
- 1.3OpenOld VersionsThree prices for one plan. Which safety net actually keeps the bot on the right one.
- 1.4OpenCite or DeclineEvery number checked against its source, and “I don't know” when it isn't there.
- 1.5OpenProve It65 questions with an answer key. Every change is a re-run, not a vibe.
Part 2Giving the AI real jobs
- 2.1OpenPoisoned FilesOne file hides an instruction. Does the bot follow it?
- 2.2OpenPulling Data OutClient emails into CRM fields, with the evidence attached.
- 2.3OpenAgents That ActSend the email, clean up the CRM. What the guards stop, and what they cost in approvals.
- 2.4OpenCost and SpeedWhere the milliseconds and the cents go, and how to cut both.