Multimodal large language model
What it means. An AI trained on a huge amount of writing that can also look at pictures. It reads a page the way a person does: words, tables and images together.
In SUBR. You upload the quote you sent the client, typed or scanned, with your site photos beside it. The model reads the letterhead, the priced table and the photos in one go.
Structured extraction
What it means. The AI fills in a fixed form, the same boxes in the same order every time, instead of writing a free answer.
In SUBR. Every quote comes back as the same record: client, site address, dates, every priced line as printed, totals and areas of work. A box the quote never mentions comes back empty, so "not on the quote" is never mistaken for "the AI forgot".
Confidence with a reason
What it means. The AI grades its own read and says, in one plain line, why it gave that grade.
In SUBR. Next to every read it says how sure it is and why: which key details it found, whether the priced lines add up to the total, and how many things it could not settle. The grade is the AI's own judgement, not a measured percentage.
Model escalation
What it means. A quicker model reads first. When it is unsure, a stronger model takes over.
In SUBR. When a read comes back with low confidence, or part of it fails, a stronger model reads the document again before your office sees it.
Uncertainty flagging
What it means. The AI writes down what it could not settle instead of guessing.
In SUBR. A figure that reads two ways, or an area with no size, goes on your office's list of things to check rather than into a box as a confident number.
Arithmetic cross-check
What it means. Adding up the priced lines and comparing them with the total, the way an estimator runs a finger down the column. Plain sums, no AI.
In SUBR. If the priced lines and the sale price disagree by more than a tenth, the quote cannot be approved until someone writes a reason.
Grounding check
What it means. The figures the AI reports are checked against the document itself, so a number it made up cannot pass as one it read.
In SUBR. The four totals that reach an invoice, net, VAT, gross and the CIS deduction, must appear in the quote itself. A total the AI worked out rather than read is left blank for your office to confirm, and any other figure it cannot find in the document is flagged. It is a safety net, not a proof that every price is right.
Human in the loop
What it means. A person checks the AI's work and approves it before anything is built from it.
In SUBR. The review has two ticks, "I have read this against the document" and "Approved for scope". No scope is built until both are on. Then your office names each trade's share and SUBR builds a scope per trade from the read.
Learning from corrections
What it means. The AI is given notes from your office's past corrections the next time it reads. The model itself is not retrained.
In SUBR. Your office fixes a wrong end date, and the next quote from the same sender is read with a note to look harder at that sender's end dates. The note never carries an old figure, date or name, so the AI still reads the new quote for itself.
Tenant isolation
What it means. Each firm's records are walled off from every other firm's, down to the database.
In SUBR. Your corrections are stored against your firm and read back only for your firm. Your lessons are never used to read another firm's quotes.
Email addresses held apart
What it means. Personal details are kept out of what the AI works from, so they can be removed on request.
In SUBR. Email addresses on a quote are taken out of the text the AI works from and stored separately, so they can be deleted on request without touching the rest of the record. This covers email addresses, not names or phone numbers.
Prompt injection defence
What it means. Text that reaches the AI from outside its instructions is treated as something to read, never as an order to follow.
In SUBR. What it learns from your corrections is kept as fixed sentences, never as text copied from a document, so a quote cannot slip instructions into what it learns.
Left out on purpose
OCR. There is no separate text recognition step. The model reads the PDF itself.
Splitting the quote into trades. Your office names each trade's share. The AI reads; it does not divide the work.