
Every reviewer we follow had a verdict on Claude Fable 5.2 within two days of its release, and almost all of them tested it on the same things: code, games, long essays. None of that tells a business owner what they need to know, which is whether it can be handed a boring, real office job and left alone. So on November 12 we took three jobs out of our own back office and handed each one over in a single paragraph, the way you'd brief a new hire on a busy morning.
How we ran it
Each job went to a fresh session of the model with no history and no coaching. Each had read-only access: it could look, and it could not send, submit or change anything. Each was told to report three things alongside the work: what it finished, where it had to guess, and which findings it trusted least. Afterwards we checked a sample of its claims against the source ourselves.
This is three jobs on one day, not a study. Treat it as a field note.
Job one: an operations audit from the outside
The brief: pick an independent Fort Worth remodeling company, read only its public website, and write what an operations consultant would want before a first call. How does a customer get from interested to booked, and where does that path lean on people doing things by hand?
It came back in about five and a half minutes with a ten-step map of the customer's path, the software the site visibly runs on, five ranked bottlenecks each tied to the page it came from, and nine questions for the owner. Every finding was labelled as observed or inferred, which we asked for and which it kept to. It noticed things a tired person skims past: an article inviting customers to send photos on a site with no email address or upload field anywhere, and a table of contents whose links all pointed at the wrong page.
We spot-checked about ten of its observed claims against the live site. All ten held. It did not submit the contact form to see what happens next, because the brief said read-only, and it said plainly that everything after the Submit button was therefore unknown. It closed by calling its own five bottlenecks hypotheses for a first call, not findings about how the company runs. That's the right amount of confidence.
Job two: a 30-page website check
The brief: run the check we do after moving a site, on thirty pages of our own. Status codes, titles, descriptions, headings, social images, every internal link, the sitemap.
It wrote its own checker, made about 125 requests at a polite pace, tested 400 internal links, and found none broken. It did find thirteen page descriptions outside the length we aim for and eleven pages falling back to a generic social image. Those are real, they're ours, and they're on our list now. We re-checked six of its findings by hand and all six were accurate.
The most useful line in its report was about what it couldn't do. We hadn't given it a list of the old site's addresses, so it couldn't test whether old links still land somewhere, and it called that the biggest gap in the whole check. It's right. That is the step that protects search traffic in a move, as we covered in our website migration checklist, and a less careful worker would have reported thirty green rows and let us assume.
The one thing we caught it getting plainly wrong was its own stopwatch. It reported roughly eight minutes. The log says under five.
Job three: a month of billing, reconciled
The brief: match October's logged time against the invoices that went out, and flag anything unbilled, double-billed or billed oddly.
This is the one it did not finish, and the reason matters. The read-only connection we gave it exposes summaries: uninvoiced time rolled up by client, open invoices, recurring profiles. It doesn't expose individual time entries or invoice line items. The model tried twenty other doors, found them shut, and then told us in the first lines of its report that what followed was a partial reconciliation, not a true match. It could see about six in ten of the month's invoices by number. It flagged eight possible oddities and labelled every one unproven.
It also stopped at a line we hadn't drawn very clearly. The full detail lives in our invoicing app behind a login. It could have gone looking for a way in. It reported that this was outside the access it had been given, and left it there. We haven't finished checking its billing flags ourselves, so we're not going to tell you how many were real.
The model didn't run out of ability. It ran out of access, and it told us exactly where.
What this changes about what you hand over
Two finished jobs, one partial, no invented results in anything we checked. A year ago the failure we planned around was the confident wrong answer. On these three jobs the failure was a partial answer, clearly marked. That is a much easier thing to manage.
It also moves the bottleneck. The model has changed less about our work than the plumbing around it has, which is the argument we made in the harness matters more than the model and again when we compared the frontier models in which one we use, and why it matters less. The billing job didn't stall on intelligence. It stalled because the data it needed sits in a place we never built a safe, read-only path to. Most small businesses have a dozen of those, and they're a large part of why AI projects stall at 80%.
So the practical read for this quarter: work that runs on things already readable, like a website, a shared folder or an exported report, can be handed over now, with a person checking a sample. Work that runs on data locked inside one application is waiting on access, not on a smarter model. And in every brief, ask it to say what it guessed. It will.
If you'd like to know which of your own back-office jobs fall on which side of that line, that's what the THINQ Diagnostic is for.


