Contents
An assessor comes back from a block of flats with a hundred and forty photographs and a notebook. The report is due on Friday. Everything they did on site had a structure, floor by floor, door by door, escape route by escape route, and the moment the photographs went into a camera roll that structure was gone. What's left is a pile of evidence and a memory of where each piece belongs, and the write-up is the work of rebuilding the one from the other.
That's the part of the job that eats the week. It isn't the inspection, which is skilled and takes as long as the building takes. It's the transcription afterwards, and for fire risk assessors it got heavier in October 2023, when the Fire Safety Order was amended so that the full assessment has to be recorded rather than only its significant findings. Every observation, every time, with the assessor's name on it.
We build for people whose product is a report, and this is the problem we spend most of our time on: how to keep the structure of the visit all the way through to the issued document, so that nothing is typed twice and nothing is rebuilt from memory.
The form is the report
The first decision is the one that does most of the work, and it has nothing to do with AI.
The site form mirrors the report. If the report follows PAS 79, the form has the PAS 79 sections. If the practice has its own structure for a survey or an audit, the form has that. Every observation the assessor makes is recorded in the place it will appear in the report, and the photograph is attached to the observation it supports, not to the visit in general. A defective door closer is a finding, in the fire doors section, with its photograph, its location and its priority, from the moment it's recorded on the device.
That means there is no transcription step, because there is nothing to transcribe. The structure the assessor used on site is the structure of the case, and the report reads the case.
It also changes what a photograph is. A photo in a camera roll is evidence of nothing in particular. A photo taken through the case is timestamped, tied to the finding and the building, and will be the thing the contractor is sent and the thing the closure is compared to. Same picture, entirely different standing when someone asks about it in three years.
What the extraction does
The second layer is the one people ask about, so I'll be precise about what it is.
Plenty of what arrives with an instruction is already written down somewhere: the previous assessment, the fire strategy, floor plans, certificates for the alarm and the emergency lighting, a schedule of flats. On SwiftCase the extraction layer reads those into the fields of the case, so the assessor starts the visit with the building's history in front of them rather than in a folder of PDFs. On site, the same layer reads the form and the photographs into structured findings.
How it does that depends on the client. For most practices it's a language model configured to their document types and their fields, which means it can be set up for a new kind of document in an afternoon rather than waiting for a model to be trained. Where a firm handles enough of one document type to justify it, we build something more specific. Either way the assessor sees the same thing: fields filled in, with the source alongside, and the ability to correct anything it got wrong.
What it doesn't do is assess. It does the typing. The judgement about whether a compartment line is adequate or a door is fit for purpose is the assessor's, and the design keeps it that way by putting the extracted value next to the evidence rather than in place of it.
Generating the report
With the findings structured, the report is generated from the case. The significant findings and the action plan are drawn from the observations the assessor flagged, with the priority bands the practice has defined and the target dates that follow from them. The assessor edits and signs. They don't start from a blank page, and they don't start from last year's report with the dates changed, which is the habit that produces two assessors describing the same building two different ways.
The template lives in one place and has a version. When the Order changed in 2023 the practices running one template changed it once, and every report issued afterwards was complete. The practices running one template per assessor found out how many templates they really had.
Document generation is the plain name for this and it's deliberately unexciting. The interesting part is upstream, in keeping the structure. Once the case has it, producing a document is the easy bit.
QA is a stage, not a favour
The last thing before issue is a review by a second competent person, and we make it a stage with a name on it rather than something asked for in a corridor. The report can't move to issue without it. The reviewer is recorded. The practice decides how deeply to sample, more for new assessors and unusual buildings, less for experienced people on familiar stock, and writes that rule down so it can be explained.
This is where a lot of the operational visibility comes from. A partner can see what's on site, what's back and waiting to be written, what's in QA and with whom, what was issued this week. Reports stop waiting in a queue nobody can see, which in most practices is where the turnaround time actually goes.
This isn't only fire
I've used fire risk assessment throughout because it's a trade where the law has recently made the documentation burden explicit, and because the workflow is a clean example. But the shape is the same for anyone whose product is a report. A surveyor, an engineer inspecting plant, a vehicle assessor with a damaged car, a clinician writing up an examination: each of them captures evidence in a structure on the day, writes it up under pressure afterwards, has somebody check it, issues it, and will be asked about it later.
Keep the structure from the first minute and the rest follows. Capture against the case, read the paperwork into fields, generate the document from what was captured, review it as a stage, issue it with a record of who received it. Then start the next cycle, which is a subject of its own.
Where it's going
Everything above is live and in use. What we're building next is the narrative layer: the assistant drafting the prose of the findings from the structured observations, for the assessor to edit rather than write. The observations already carry everything the paragraph needs, and the assessor's time is better spent on the judgement than on the sentence. Beyond that is precedent, which is what the context graph is for: what did we find last time in buildings like this one, and what did it take to close. We'll describe both when customers are using them.
Further reading:
- Fire risk assessment workflow: instruction to issued report, step by step
- Document extraction: the paperwork reads itself, so your people don't have to
- Document generation: the report produced from the case, from one versioned template
- The review date is the product: what happens after issue