AI tender document extraction.

A worked workflow for commercial builders and civil contractors who read several 800-page tender packs a month to decide whether to bid and what to return.

Workflow5 min readUpdated 19 October 2026

Short answer

AI tender document extraction is a workflow that reads a tender pack (request for tender, conditions of contract and special conditions, specifications, schedules and addenda) and pulls out the dates, returnables, risk terms and scope items an estimator needs, each linked to its page. For New Zealand contractors it shortens the path to a go or no-go decision; people still price, judge risk and sign the tender.

  • Extract dates, returnables, risk terms and scope, each with a page reference.
  • Compare special conditions against your own notes on NZS 3910:2023 clause positions.
  • Every addendum triggers a rerun and a change log; the brief shows which addendum it reflects.
  • Drawings and quantities stay with estimators; spreadsheets are parsed by code, not the model.

What comes out of a tender pack

Tender extraction produces three working documents from the written pack. The first is a two-page brief: client, project, contract form, key dates (site visit, questions close, tender close with time zone, validity period), evaluation criteria and weightings, and the headline risks. The second is a returnables matrix listing every form, schedule, certificate, insurance and health and safety document the client asks for, with the page that asks for it. The third is a risk register of commercial terms: liquidated damages, retentions, bonds, payment periods, design responsibility and any amendment to the standard conditions.

Each line carries a document name and page number, so the estimator and commercial manager check claims instead of rereading the pack. That makes it a document processing workflow, not a chatbot.

How the extraction runs

Inputs
The request for tender, conditions of contract and special conditions, specification sections, drawing register, schedule of quantities, notices to tenderers and addenda, and the firm’s own notes on standard clause positions.
Systems
GETS or the client’s tender portal, the firm’s document store (often SharePoint), a model API under business terms, spreadsheet parsing code, and the tender register.
Steps
Inventory every document and version; extract dates and returnables; list each special condition with the clause it amends; pull risk terms; map specification sections to trades; parse schedules with code; compare each addendum with the previous version; assemble the brief.
Human review
The estimator checks scope and quantities. The commercial manager reads every special condition flagged as high risk. A director makes the go or no-go call. Before lodging, a person ticks off each returnable against the matrix.
Outputs
The tender brief, the returnables matrix, the risk register with clause references, a draft list of questions for the client, and an addenda change log.
Measures
Working hours from receipt to go or no-go, returnables missed at lodgement (target zero), questions lodged before questions close, and amendments the commercial manager found that the register missed.
Risks
A missed addendum, misread special conditions, scope that appears only on drawings, tender confidentiality terms, and wrong close times.

Illustrative example: a 45-person Tauranga commercial builder

Illustrative example, synthetic numbers: the firm considers six tenders a month. A typical pack has 850 pages of written documents (40 for the request for tender, 120 of conditions and special conditions, 600 of specification and 90 of schedules) plus 140 drawings. Today the estimator and commercial manager spend about 14 hours between them on a first read. With the brief and register, the firm assumes 6 hours of checking and targeted reading.

Line First read by hand With extraction brief
Tenders a month 6 6
Senior hours on first read 84 36
Bulk extraction per tender None about US$2.20
Special conditions comparison per tender None about US$0.72
Model fees a month None about US$18

Bulk extraction assumes 800,000 input tokens (the written pack in sections plus instructions) and 60,000 output tokens on Claude Sonnet 5.5 at US$2 and US$10 per million. The special conditions pass assumes 130,000 input and 10,000 output tokens on Claude Opus 5.5 at US$4 and US$20 per million. Prices are as listed on 19 October 2026. Addenda reruns add a little. Drawings are excluded from both lines.

NZ procurement and contract forms

  • GETS and the Government Procurement Rules. Public sector work is advertised on GETS, and the Government Procurement Rules, now in their 5th edition, set out when and how agencies must advertise openly. Notices to tenderers and addenda arrive through the same listing, so the workflow should watch it, not just the first download.
  • NZS 3910:2023. Standards New Zealand’s conditions of contract for building and civil engineering construction are the base for many tenders, with special conditions amending them. Standards New Zealand says its general conditions may not be adapted or repurposed without permission, so compare against your own clause notes and check licence terms before loading the standard itself.
  • Health and safety returnables. Under section 34 of the Health and Safety at Work Act 2015, businesses sharing a duty must consult, co-operate and co-ordinate. Extraction lists what the client requires; your safety team writes the response.

What estimators and directors keep

Price, margin, risk appetite, methodology and non-price responses, every question sent to the client, and the signature on the tender form. The system never contacts the client and never lodges anything.

Where tender extraction goes wrong

  • The brief is one addendum behind. Notice to tenderers 3 moves the close date or replaces a specification section. Version every document, rerun on each addendum, and print “as at addendum 3” on the brief.
  • Amendments by reference. “Clause 12.2 is deleted and replaced with the following” means little without the base clause. Where your notes do not cover the clause, the register says so instead of guessing.
  • Spreadsheets with hidden rows. Schedules of quantities with merged cells and hidden rows confuse models. Parse them with code and keep quantities out of the model’s hands.
  • Scope on the drawings. A note on a drawing adds work the specification never mentions. The brief states plainly that drawings were not reviewed.
  • Long-document blind spots. Models can miss terms buried in the middle of very long inputs. Process sections separately and test recall; see evaluating long context models.

Tender extraction pairs with the other construction workflows we map: a site safety knowledge assistant and quote preparation from job notes. The construction sector guide covers the wider picture, and our SharePoint integration notes cover document permissions.

Sources

  1. GETS: Government Electronic Tender Service, New Zealand Government Procurement, accessed 19 October 2026
  2. Government Procurement Rules, New Zealand Government Procurement, accessed 19 October 2026
  3. NZS 3910:2023 Conditions of contract for building and civil engineering construction, Standards New Zealand, accessed 19 October 2026
  4. Health and Safety at Work Act 2015, New Zealand Legislation, accessed 19 October 2026
  5. Claude API pricing, Anthropic, accessed 19 October 2026

Written by the Levy Street team in Wellington. Last updated 19 October 2026.

Questions people ask

Can AI read construction drawings for a tender?

It can read drawing registers, title blocks, revision letters and notes reliably enough to check that the set is complete and current. It is not reliable for quantity take-offs or for finding scope that only appears on a drawing. Treat the extraction brief as covering the written documents, say so on the brief, and keep drawings with the estimator.

Will AI help us win more tenders?

Nobody can promise that, and you should distrust anyone who does. What extraction changes is how quickly you reach a go or no-go decision, how early your questions reach the client before questions close, and how rarely a returnable is missed. Whether the bid wins still depends on price, methodology and relationships.

Is it safe to upload tender documents to an AI tool?

Read the tender conditions first: some restrict how documents may be copied or shared. Use a business plan or API account whose terms exclude training on your data, keep documents in your own store, and record where processing happens. Key personnel CVs in your returnables contain personal information, so the Privacy Act 2020 applies to them as well.

Does it work with NZS 3910:2023 contracts?

Yes, in the sense that it can list every special condition, the clause it amends and what changes. It works best against your own written notes on each standard clause position. Check your Standards New Zealand licence terms before loading the standard itself into any system, because Standards New Zealand restricts adapting or repurposing its general conditions.

How do we test it before relying on it?

Run it on four or five past tenders your team has already read closely, including one with heavy special conditions and one with several addenda. Compare its brief and returnables list with what your team actually submitted. Missed returnables and missed amendments are the two numbers to watch.

Have something to build?

Tell us about the workflow. We will map where AI can change speed, margin or decision quality, and what it would take to run it in production.