How AI reads a mechanical specification

A technical but accessible look at how document extraction works on real engineering documents: what AI does reliably, what it cannot do, and why the human estimator is still essential.

Reading means something different per document

A tender pack is not one kind of reading. An equipment schedule is a table, and the work is extracting rows faithfully: tags, quantities, duties, models. A specification is prose, and the work is interpretation: a metering strategy described across three paragraphs, an exclusions clause buried in section 7, a sentence that quietly declares the AHU comes with its own controls. A drawing is neither, and the work is visual: reading tags off a schematic, counting repeated symbols, noticing that a label convention means eight boxes, not one.

Tenderfly treats these as different problems. Schedules and specifications are read by a language model tuned for each; drawings go through a vision model plus a deterministic cross-check that reads the drawing's embedded text layer and verifies every tag the model claims to have seen. When the two disagree, the discrepancy is chased, not shrugged off.

The pipeline

1. Classification: every document is identified as a specification, schedule, schematic, floor plan or something else.
2. Extraction: equipment is pulled from each document with the method suited to its format.
3. Correlation: the results are merged into one equipment list. This stage is deliberately not AI: deterministic code matches tags across documents, merges duplicates and resolves conflicts by fixed rules, so the same documents produce the same list every time.
4. Points generation: calibrated IO templates are applied per equipment type to produce the schedule.
5. Pricing: your organisation's rates turn the schedule into a proposal.

Containing the model

Language models are probabilistic, and pretending otherwise is how AI tools lose estimators' trust. The containment is structural. Determinism lives in the stages after the model: cross-checks against the text layer, correlation by rule, templates that fix each equipment type's shape. And every extracted item keeps its receipt: the page and sentence, or the drawing and tag, that produced it. When the output is wrong, it is wrong inspectably.

Where it stops

Scope judgement stays human. When a specification says monitoring of the domestic hot water system is "to be agreed at detailed design stage", the right answer depends on the project, the consultant's habits and your commercial appetite. The system flags items like this rather than guessing, and it does the same when documents disagree with each other, presenting the conflict with both readings priced. A confident wrong assumption is worse than a flagged question.

The estimator's job, after

The first draft arrives complete: equipment, points, prices, evidence. The estimator's work shifts from producing the schedule to judging it, deciding the flagged items, adjusting for what they know that no document says, and approving what goes out. Their expertise still shapes every bid. It is just no longer spent on transcription.