Case studies
What we built, and how it is governed.
Each case is described with the client’s written consent. Where a client has asked not to be named, we describe the business and leave out the name. No saving is claimed until it has been measured against a baseline.
Case study 1 · October 2026
A UK freight forwarder: drafting air-export quotes
Every enquiry arrives as an email: an airport, a destination, the pieces, their weights and sizes. Someone on the team reads it, looks up the airline’s rate and the haulier’s card, works out the charged weight, adds handling and fuel, and writes the reply. Quotes take a large share of the team’s week, spent on steps that run the same way every time.
The steps are the same. The cases are not. Some cargo cannot be stacked. Some rates have run out. Some enquiries leave out the weight. An assistant has to know which case it is looking at.
What Candor built
The model reads. The code prices. A person sends.
An assistant that reads each air-export enquiry and drafts the quote, with its workings, for the team to check. Each part does only what it can be trusted with.
The language model reads the enquiry and writes the covering words. It never sets a price. A draft whose wording carries a price or a figure that is not in the quote is refused.
The price comes from code with its own suite of tests, working from the business’s own airline tariffs and haulier rate cards. Given the same details and the same rate file, it gives the same price every time. Asked why a quote came to its price, it shows the steps the code took, not the model’s account of them.
Every quote stays a draft until someone on the team checks and sends it. Where details are missing, it drafts a question asking for exactly those. Where the cargo needs judgement, such as cargo that cannot be stacked, it passes the enquiry to a person.
How it is governed
The six questions, applied
Candor asks six questions before anyone relies on an AI result. Here is how each was answered for this assistant.
- 1
Consequence
A wrong quote loses the job or loses money on it, and once sent it is hard to take back. So the assistant drafts and never sends.
- 2
Accountability
The business owns the machine it runs on and the workflow software’s licence. The person who sends a quote answers for it, as before.
- 3
Data and model provenance
Every price traces to a named tariff, a rate card or a rate the team typed in. A tariff past its end date is not used, and the workings name any airline with no current rate rather than estimate one. The gateway lets through only named models.
- 4
Operational integrity
It runs on a machine in the business’s own office. The text of each enquiry goes to Anthropic’s model to be read. The pricing runs on the office machine, and the records are kept there. The model is reached through one gateway, which holds the key and logs each call without its content. A daily report shows what was drafted, what was passed to a person and what was corrected.
- 5
Non-discrimination
The people it affects are the business’s customers. The price depends on the cargo, the route, the terms and the rates in force that day, not on who is asking. The customer’s name sets only the currency of the quote. A person checks every draft, and the daily report shows what was corrected.
- 6
Reperformance
The reading of each enquiry is saved once, tied to the text of the email. A correction is added on top, never written over it. A second person can take the saved reading, the rate file in use that day and the same code, and get the same price.
How it was built
From habit to a written procedure.
It is built on n8n, with Claude from Anthropic for the reading and writing, and Candor’s own code for the pricing. Before any of that, Candor used AI agents to read more than 250 of the business’s own documents: airline tariffs, haulier rate cards, the team’s quotation sheet, sample quotes, and recorded walkthroughs of a quote being done. Each tariff was sorted: current, expired, superseded, duplicate or to be checked.
How the team quoted had never been written down in full. Much of it lived in people’s heads and in a spreadsheet. From those documents and walkthroughs, Candor wrote it out as an explicit procedure: a flowchart for each step, and thirty points where two sources disagreed, each now a named rule that the business decides.
Why it matters
The work can be done again, and come out the same.
The main benefit is reperformance. Any quote can be taken by a second person, with the saved reading, the rate file in use that day and the same code, and it gives the same price, with every step shown.
The procedure no longer depends on who is in the office that day. When a rate changes or a rule is settled, the change is written down and dated, not remembered.
Send me the question you cannot answer yet.
One address. I read and reply to every email myself.