From voice note to site diary: how the automation actually works
The pipeline is short. A foreman sends a voice note or a photo to the project WhatsApp group. The audio is transcribed. The facts are extracted into a structured record: labour, plant, weather, progress, delays, visitors, each tied to a timestamp. A formatted diary entry is assembled and put in front of a named person, who corrects it and approves it. That approval is what turns machine output into a record the business will stand behind.
We have already made the case for why the site diary matters. That post is the why. This one is the how, because the how is where most attempts fail: not in the AI, but in asking busy people to change how they communicate. The trick is that nobody on site changes anything. The foreman was already sending that voice note at ten past five. The group was already the de facto project record. The automation just stops the record evaporating.
Here is each step, and what it exists to catch.
What happens between the voice note and the diary?
Five stages. Four belong to the machine, one is deliberately human.
| Stage | What happens | What goes wrong without it |
|---|---|---|
| Capture | Voice notes, photos and messages arrive in the group the site already uses | Nothing is captured at all; the day lives in someone's head |
| Transcription | Audio becomes text, with sender and time preserved | "Two joiners till Thursday" exists only as sound |
| Extraction | Diary facts are pulled out: labour, plant, weather, progress, delays, visitors, deliveries | Facts stay buried in chat scroll, unfindable in month six |
| Assembly | One dated entry per day, with gaps flagged rather than guessed | A tidy-looking diary with silent holes in it |
| Approval | A named person reads, corrects and signs off | Machine guesses acquire false authority |
The order matters. Extraction before assembly means every line in the finished diary traces back to a specific message from a specific person at a specific time. When someone asks in month fourteen how many operatives were on site the day the slab was poured, the answer is not "the AI said six". It is "the foreman said six, at 17:42, in this message, and the site manager approved the entry that evening".
What does the extraction actually pull out?
The same headings a paper diary always carried. Labour: trades and counts, ideally by subcontractor. Plant: what is on site and whether it stood idle. Weather: what was said, and whether it stopped work. Progress: what was done and where, in the site's own vocabulary of levels, blocks and gridlines. Delays: what stopped work, when, for how long, and who was told. Visitors and deliveries: who turned up, what arrived, what was rejected.
Two behaviours matter more than raw accuracy. First, the system quotes rather than embellishes. If the voice note says "drainage held up again, waiting on the utility company", the entry records that, not an invented paragraph about root causes. Second, silence is flagged, not filled. If nobody mentioned weather on a day the whole region was rained off, the entry says weather was not reported. A visible gap is honest. A gap papered over with a plausible guess is how a record loses credibility in front of an adjudicator.
Why does a human still approve every entry?
Because a diary is evidence, and evidence needs an owner.
The SCL Delay and Disruption Protocol puts contemporaneous records at the centre of how delay claims are made and resisted. A record generated by software and never read by anyone is contemporaneous, but it is fragile: the other side only needs one confident error to argue the whole diary is machine noise. An entry a site manager read, corrected and signed off on the day is a different class of document.
The approval takes two or three minutes, and the approver checks specific things. Labour counts against who was actually there. Delay wording, because the note usually describes the symptom and the diary needs the cause. Anything the extraction inferred rather than heard. And what is missing, because the classic tell is the quiet day: a one-line entry on the day everything went wrong usually means people were too busy firefighting to type. That is exactly the day the diary needs a full entry, and only the approver knows it.
Where does the finished diary actually go?
Into wherever your project information already lives, named so it can be found. If you work to ISO 19650 conventions, diary entries slot into the common data environment like any other information container, with consistent naming and status. If you run a SharePoint folder, the same discipline applies in miniature: one entry per day, never edited after approval, corrections added as new notes rather than silent rewrites.
The export format matters too. A PDF for the record, and structured data alongside it, so a QS can filter three months of labour returns in minutes when a small disagreement starts growing teeth.
What does it take to set up?
Less than the scepticism suggests: a WhatsApp group that already exists, an agreed list of what counts as a diary fact, and one named approver with a fifteen-minute daily habit. This is the pipeline Construction Metric runs as a product, and the same pattern our WhatsApp site assistant work builds for firms whose sites have their own quirks.
The honest summary: the machine does the typing, the timestamps and the memory. The human does the judgement, once a day, with a coffee. Neither replaces the other, and the diary that results is better than either would produce alone.