At a business law firm we work with, one task comes back with every contract: check whether a legal document complies with European law, and amend it where it does not. Two of the firm’s lawyers, a junior and a senior, told us what that task has cost them at each stage. Before AI, producing the analysis meant about four days of a lawyer’s time. With Claude, the way most lawyers use an AI assistant today, it came down to a day of back-and-forth and roughly €500 in tokens. With their method captured in Clarifeye, it takes about an hour and €44.
| Approach | Setup | Time | Token cost | Redo rounds |
|---|---|---|---|---|
| Before AI | 4 days | |||
| Claude alone | none | 1 day | ~€500 | all day |
| Claude + Clarifeye | 1h with Clara, paid once | ~1 hour | €44 · 11x cheaper | none |
With Claude alone, the back-and-forth eats the day
Their process is the one most law teams have settled into. The junior starts a session and writes a long prompt: the context of the matter, the applicable regulations, instructions on what to check and how to present the result. He connects Claude to the legal databases the firm relies on, so it checks the law rather than answering from memory, and uploads the firm’s model documents so the output looks like the firm’s work. Even then, every answer carries the same doubt: did Claude actually read all of it, or is it improvising around it? So he asks for rechecks, and waits while the assistant re-reads hundreds of pages.
When the junior is satisfied, the amended document goes to the senior. The senior spots an error straight away. Not an exotic one: the kind of thing he checks by reflex on every document of this type, and that no prompt had told Claude to check. So he opens his own session, loads the document plus the model documents he personally relies on, and runs his own round of checks.
Then the document goes back to the junior, and the loop continues. Each round repeats the same overhead: re-explain the context, re-upload the references, pay again for the assistant to read the same corpus it read the round before. The exchanges stretch from morning into the evening, wedged between the rest of both lawyers’ work. By the end, they told us, the task has eaten the whole day and about €500 worth of tokens.
The expensive part was the re-explaining
Nothing in that story is a model failure. Claude analysed what it was given, every time. The money and the hours went somewhere else: into telling it, over and over, what the task was and what to read.
The method lived in the senior’s head. The list of references lived in two people’s chat sessions. So every round started from zero against a large corpus of regulations, model documents and templates, and every round billed the full reading cost again. As our experience shows, structured knowledge cuts token use by an order of magnitude, for exactly this reason: what the system reads dominates the bill. Here, the same reading was bought five or six times.
That is the shape of the problem for any team using a raw assistant on recurring document work. The task repeats, the corpus barely changes, and the knowledge that makes the output acceptable is written down nowhere. You pay for that gap in hours and in tokens, and then a third time in redo loops.
With Clarifeye: one hour with Clara, then ten minutes to generate
Then the firm captured that method in Clarifeye.
Each lawyer spent thirty minutes in an interview with Clara, our AI interviewer. Between the two conversations, three things were captured into a knowledge store:
- the method the firm actually uses for this compliance analysis and the editing that follows;
- the legal databases to call when checking the law;
- the model documents to draw from, and what “looks like our work” means in practice.
Now they ask Clarifeye to perform the analysis. The amended document comes back in about ten minutes, for €44 of tokens. The junior verifies it within the hour; the senior’s review takes twenty minutes. No redo round follows.
The detail that matters more than the speed: the error the senior kept catching in the Claude rounds no longer makes it into the document. His reflex check, the one no prompt ever mentioned because it had never been written anywhere, came out in the interview and went into the store. It now runs inside the analysis instead of after it. The redo loop with the raw assistant existed precisely because that check lived only in his head; once captured, the first pass already meets his standard.
What the hour with Clara actually buys
Count the interview hour and the Clarifeye run still comes in at about two hours against a full day, and against the four days the same analysis took before AI. Most of that €500 went on rounds a captured method makes unnecessary.
The comparison also improves with every contract, because the hour with Clara is paid once. The knowledge store now holds the method, the sources, the standards. The next contract of this type starts at the ten-minute run, with no interview in front of it. Every future run draws on the same captured hour.
The junior can now run it
Something else changed hands during that hour. The senior’s checks used to be supervision: he applied them personally, after the junior’s work, on every document. Now they sit in the store and run whether or not he is in the room.
Which means the junior can take over contract preparation. He launches the analysis, the senior’s standards apply on the first pass, and the senior reviews a deliverable rather than redoing a process. The firm’s way of working stopped being something you learn by being corrected, and became something the next person can simply use.
The method outlives the model
One last property, easy to miss and decisive over a year: nothing captured in those interviews belongs to Claude, or to any model.
A knowledge store is served to whatever AI system the firm points at it (Clarifeye exposes it over MCP, an open protocol most AI tools now speak). Prompts die with the chat session that contains them; a captured method does not. When a better model arrives, or the firm changes vendors, nobody re-records the hour with Clara. The new model plugs into the same store and inherits the same method, the same sources and the same checks.
Where to start
If you want to find your own version of this story, look for the task your team repeats every week against the same pile of reference documents, the one where a senior still catches the same mistake at the same step. That check has never been written down. It costs an hour to capture, and right now it is costing you considerably more than that.