CASE 02 · METHOD TEARDOWN
One enquiry, read properly before anyone answers it.
The sender, the firm and the e-mail below are invented for this page. This is a teardown of a method, not a client story — no result figures are claimed anywhere on it.
Most enquiries do not fail at the answer. They fail earlier: a price is guessed from an e-mail that never said enough, and three weeks later the job turns out to be a different job. Triage is the step that refuses to skip that gap. Here is the same enquiry before and after it.
What arrives
Written at 23:41 by someone describing a real problem badly — which is how real enquiries arrive. Nothing here is wrong with the sender.
Subject: quick question
Hi, we got your name from a colleague. We're a small accounting firm (11 people, Lyon) and we're drowning in client documents. Every January we get maybe 400 folders of receipts, bank statements, some scanned, some photographed badly, and two people spend six weeks just sorting them before anything useful happens.
Someone told us AI can read them now. Is that something you do? What would it cost roughly? We'd need it to work with our current tool (we use a French accounting suite, I can find out the name). Also our clients' data obviously cannot leave France, our compliance person is strict about that.
Could we talk this week? I'm mostly free Thursday.
— P.
What triage turns it into
The same text, restated as decisions and gaps. Note what is absent: no price, no promise, no feasibility claim.
ASKED FOR: bulk intake of client accounting documents (~400 folders/year, mixed scans and phone photos) → sorted, machine-readable output into an existing French accounting suite.
MATCHES: document processing + private-AI deployment. Not a media or automation job.
MISSING BEFORE A QUOTE (3): name and API surface of the accounting suite · one representative folder (worst-quality photo included) · who signs off on the data-protection side.
CONSTRAINTS STATED BY SENDER: data must stay in France. Hard requirement, not a preference — this rules out cloud OCR and decides the deployment shape.
FLAGS: client financial documents = third-party personal data; the firm is the controller, we would be a processor → written agreement required before any real file is touched. No pilot on live client data without it.
SUGGESTED NEXT STEP: 20-minute call, ask the three missing items, offer a paid scoping pass on 20 anonymised folders instead of a guessed price.
NOT DECIDED HERE: price, timeline, feasibility on their worst scans. All three depend on the sample.
The sample stays in English on every language version — it is the exhibit; the analysis around it is what gets translated.
The four rules doing the work
None of them are clever. They are simply applied every time, which is the part a busy inbox cannot do.
rule 01 — Name what is missing before naming a price
The card ends with an explicit list of what would be needed to quote. If that list is empty, the quote is a guess wearing a number. Three items here — the tool name, one representative folder, and the person who signs off on data protection — are worth more than any estimate written without them.
rule 02 — Repeat the sender's constraint as a decision, not a note
"Data cannot leave France" is not a preference to acknowledge politely at the end of a reply. It removes an entire class of solutions before design starts. Triage moves it from the middle of a paragraph to a line that changes the architecture.
rule 03 — Mark third-party personal data on sight
The documents belong to the firm's own clients. That makes the firm the controller and us the processor, and it means a written agreement comes before the first real file, not after the pilot goes well. Spotting this on day one is cheaper than discovering it in week three.
rule 04 — The next step must be smaller than the project
A scoping pass on twenty anonymised folders is something both sides can afford to be wrong about. A full quote on four hundred folders sight-unseen is not. The suggested step is deliberately the smallest one that produces a real answer.
What changed between the two cards
- The request was named in our own vocabulary — document processing plus private deployment — instead of being left as "AI for receipts".
- Two services were explicitly ruled out, so nobody pitches what was not asked for.
- The missing information became a numbered list instead of a vague feeling that something was unclear.
- The sender's compliance constraint was promoted from a passing sentence to an architectural decision.
- The legal role — controller and processor — was written down while the enquiry was still an e-mail.
- The reply now has one job: ask three questions and offer a small paid step.
What this actually improves
Not speed of typing. The gain is that the same enquiry produces the same reading on a Tuesday night as it would on a calm Friday morning, and that a quote is never sent while three unknowns are still open. That is a smaller claim than "saves you hours", and it is one we can stand behind.
How it runs in practice
The enquiry lands, the model produces the card against a fixed checklist, and a person reads both before anything is sent. The card is a draft for a human, never an outgoing message — nothing reaches the sender without being read by us first.
What this page does not claim
- No client agreed to this, because the client does not exist. Nothing here is a testimonial.
- No time or money saved is claimed — nothing was measured with a real firm.
- The card does not decide anything. It prepares a decision that a person makes.
- Nothing is sent automatically. There is no auto-reply anywhere in this method.
Send us a badly written enquiry on purpose
Describe your own situation as loosely as you like. You get the same card back — what we understood, what is missing, what we would refuse to quote until we see it.
Ask for a triage of my own enquiry
contact@castel.studio