Is your file good enough for an AI to read?

The Fika Friday — 7 August 2026

I keep coffee breaks in Todoist. Actual tasks, actual time, blocked out in the day like anything else. It sounds precious. It isn't. It is the only reliable way I have found to read something that isn't a client document. This week it earned its keep.

CostsMaster has partnered with Portlaw Strategic Intelligence, the AI consultancy founded by Dr Victoria McCloud in November 2025 and based in Dublin and Waterford. For costs people and project management people, this is good news. I should declare the bias — I have used CostsMaster extensively across my career, first as a law costs consultant, later as a legal project manager. And McCloud comes at this as a former King's Bench Master and deputy costs judge, fourteen years running the King's Bench masters at the Royal Courts of Justice. She knows costs and its foibles from the side of the bench most of us only see from below.

What caught me was not the technology. It was the framing. "Our aim with CostsMaster is not to replace that expertise but to take the heavy lifting off," McCloud said. That is an acknowledgement that costs law rewards precision and human judgement, and that at present AI can supply neither.

I have now seen several attempts at AI-driven costs and billing. A case management system that will produce an N260, a Precedent H, even a bill of costs, straight from the electronic file. I have yet to see one as accurate as a trained human eye. They will build the basic structure, and some of them build it well. But — and it is a considerable but — case management systems are not perfect repositories of a legal matter, because the humans filling them are not perfect.

File notes saved to the wrong place, or to no phase at all. Attendance notes that ought to be split across phases — part witness statements, part expert reports, part disclosure — logged as one undifferentiated entry. Emails that started outside the system and never got dragged back in. WhatsApp. Texts. Calls made at nine in the evening because that is when the client can actually talk. If you have done PI work you know exactly why: you cannot take a call on the shop floor at Tesco. That call happened. It was reasonable and it was necessary. Whether it is in the record is an entirely separate question.

And then email. The thread. The habit of replying to whichever message happens to have the right name in it, so that a disclosure query arrives buried under a subject line about a listing appointment. Every costs draftsman has unpicked that by hand. No model unpicks it reliably from an export.

None of this is an AI problem. Badly kept files predate AI by decades. I am from the generation of eighteen-column legal aid bills and paper assessments, boxes stacked in the corner of the room waiting for someone to find the one note that mattered. What has changed is that we have started to behave as though the electronic file is the file. It is a representation of the file. And a cost judge can still ask to see the actual attendance note.

So where does AI earn its keep? Upstream. Not at the end.

Anyone who has done this work knows the sniff test — the instinct that says an entry is wrong before you can articulate why. AI is not going to acquire that any time soon. What it can do is leave less to sniff at. Capture at the point of work. A prompt for a note while the memory is still fresh, rather than a reconstruction ten years later. Costs software and the case management system talking to each other in real time instead of meeting for the first time at detailed assessment.

Classification is the unglamorous win, and probably the biggest one. Stripping out every PNG banner — the award logos, the accreditation badges, the "please consider the environment" footers — that gets dragged into the CMS and then exported into the XML. Prompting not only for time recording but for allocation: which phase, which task, which activity. A daily sweep of the mailbox rather than an annual excavation. Phase, task and activity coding has been mandatory since Precedent S arrived in April 2018. Eight years on, files still land uncoded. That is not a software gap. It is a workflow gap, and workflow is precisely where this technology is useful.

Get that right and the budget stops being an afterthought. A Precedent H prepared three weeks before the CCMC, from a file nobody has coded, is a guess with a statement of truth attached to it. Prepared at the outset from a live record, and revisited as the matter moves, it becomes a forecast. Every matter is a project, and projects get baselined before they get governed.

And there lies the other issue. Metadata cannot lie. A note created on the day of the attendance and a note created eighteen months afterwards look identical on the page and nothing alike underneath. That cuts both ways — it exposes the reconstruction, and it vindicates the practitioner who did the job properly at the time.

I will leave that one with you over a coffee.