AI · Language models
LLMs in construction procurement: where they help, where they fail, and how to ground them
A language model will happily name a fixture that does not exist. In procurement that is not a quirk; it is a wrong order. The fix is architecture, not prompting.
Short answer
Large language models help in construction procurement where the work is reading and writing: extracting attributes from specifications, summarizing requirements and drafting documents. They fail when asked to recall facts: they can confabulate products, catalogue numbers and figures with full confidence. They become usable when grounded: answers come only from verified data retrieved for the task, outputs follow a fixed schema, every value is validated, and a person approves anything that leaves the building.
Key takeaways
- Use LLMs to read and write, not to remember products.
- Ground every answer in retrieved, verified data (RAG).
- Ask for structured fields with evidence, then validate them.
- Keep a person in the loop for approvals.
- Measure errors by severity, not just by count.
Where language models help
| Task | Why LLMs fit |
|---|---|
| Reading specification sections | Long, language-heavy, inconsistent formatting |
| Extracting attributes from schedules | Turning free text into fields |
| Summarizing requirements | Substitution rules, submittal requirements, warranties |
| Drafting documents | Transmittals, RFI drafts, comparison notes |
| Explaining differences | Plain-language notes on why an alternate differs |
These are reading and writing tasks. The model works on text that is in front of it.
Where they fail
Ask a model to recall rather than read and it may produce a fluent, confident and false answer: a fixture family that does not exist, an option code from the wrong catalogue, a lumen figure that “sounds right”. The NIST Generative AI Profile lists this as confabulation, one of the risks specific to generative AI, and points to grounding, retrieval and human review as mitigations.
Other failure modes in procurement:
- Silent unit errors: lumens per foot read as per fixture.
- Hidden uncertainty: a guess presented with the same confidence as a fact.
- Context loss: a requirement in an addendum the model never saw.
Five controls that make LLMs usable
1. Ground in verified data
Products and attributes come from a verified catalogue retrieved for the task (retrieval-augmented generation). The model selects and explains; it does not invent.
2. Structured outputs
Ask for defined fields (type mark, voltage, dimming, CCT…) with the source text for each, not a free-form answer.
3. Validation
Check every value: units, ranges, option codes against the manufacturer’s ordering logic, consistency with plans and spec.
4. Confidence and routing
Low-confidence or conflicting lines go to a person instead of being guessed.
5. Human approval
People approve matches and anything that leaves the building. Their decisions become the evaluation set.
Measure by severity
Not all errors cost the same. A wrong finish is a resubmittal; a wrong voltage is a wrong order. Weight errors by consequence when you evaluate. See how to measure AI accuracy in spec matching.
Why construction is catching up
Construction has been among the slowest industries to adopt AI at scale, according to McKinsey research reported by BD+C. Language models change the economics because procurement is document-heavy. The firms that benefit are the ones that pair them with verified data and review. See our AI roadmap for lighting procurement.
Frequently asked questions
What is AI hallucination in construction?
A confident but false output: a product that does not exist, a catalogue number with the wrong options, a code clause that says something else. NIST’s Generative AI Profile calls it confabulation.
What is retrieval-augmented generation (RAG)?
A pattern where the system first retrieves relevant, trusted documents or records (for example verified product data) and the model answers only from them, citing what it used.
Do LLMs replace estimators or reps?
No. They take over reading, extraction and drafting. Judgement, approvals and relationships stay with people.
Which tasks should never be left to an LLM alone?
Approving equals, interpreting code requirements, certification statements and anything that commits money or a schedule.
References
Keep reading
- AI · Roadmap AI in lighting procurement: a technology roadmap from spec to site A technology roadmap for AI in commercial lighting procurement: reading specifications, matching every line to verified products, assembling quotes and submittals, and the next stages, learning from review, supply-chain intelligence and agentic workflows. Where AI helps today, where people stay in the loop, and what to measure. →
- AI · Document AI How AI reads a lighting specification: from PDF to a structured fixture schedule How AI turns a lighting specification and fixture schedule (PDF, scanned or Excel) into structured, checkable data: layout detection, table extraction, entity extraction with language models, unit normalization, cross-checks against the plans, and confidence-based human review. →
- AI · Evaluation How to measure AI accuracy in spec matching: the metrics that matter How to evaluate AI specification matching in construction and lighting procurement: what a 'match rate' should mean, precision and recall, top-k accuracy for alternates, human acceptance rate, severity-weighted errors, coverage, and business metrics such as time-to-quote and resubmittal rate. →
- Platform · Spec matching AI spec matching for commercial lighting SuppLux matches every line of a lighting specification to verified products and ranks equal or better alternates by photometrics, electrical data, mounting and lead time, with a specialist review before quoting. →