AI site safety knowledge assistant.

A worked workflow for contractors whose safety plans, method statements and data sheets exist but are hard to find on a phone, on a busy site, at the moment someone needs them.

Workflow5 min readUpdated 19 October 2026

Short answer

An AI site safety knowledge assistant answers workers’ and supervisors’ questions from a contractor’s approved safety documents, such as site-specific safety plans, task method statements, safety data sheets and toolbox talks, and shows the document and section behind each answer. On New Zealand sites it helps people find the rule quickly; it does not discharge any duty under the Health and Safety at Work Act 2015.

  • Answers come only from approved, current documents for the worker’s own site.
  • When the documents are silent, the assistant says so and points to the supervisor.
  • Emergencies and possible notifiable events bypass the assistant and go straight to people.
  • Document control and weekly answer review cost more than model fees.

What the assistant answers, and what it refuses

A site safety knowledge assistant is a question box on a worker’s phone that searches the contractor’s approved safety documents and replies in two or three sentences with a link to the source. Typical questions: what the exclusion zone is for this afternoon’s crane lift, which data sheet covers the epoxy in the store, what respiratory protection the cutting task needs, who holds first aid today, and how to log a near miss.

It refuses three kinds of question by design. Anything describing an emergency gets a fixed reply: call 111 and tell your supervisor now. Questions about injuries or symptoms are pointed to a person. Questions the approved documents do not answer get “I can’t find this in the approved documents for your site; ask your supervisor”, never a best guess. This is retrieval over a small, controlled library, the core of most RAG development work.

How the assistant is set up

Inputs
Approved site-specific safety plans, task analyses and method statements, safety data sheets, permit procedures, emergency plans, toolbox talk content, the guidance the business has adopted, and workers’ questions by text or voice.
Systems
The document control system (often SharePoint), a search index tagged by site, trade and version, a chat interface in the app crews already use, a model API, an answer log, and an escalation path to supervisors.
Steps
Ingest approved documents only; tag each by site and expiry; ask which site if unknown; retrieve passages; answer with citations; refuse when nothing relevant is retrieved; log question, answer and sources; review gaps weekly.
Human review
The safety advisor approves every document before it is indexed, reviews a weekly sample of answers, and reads every refused or flagged question. Supervisors make every decision on site.
Outputs
Cited answers, escalations to supervisors, a weekly gap report, and usage by site.
Measures
Share of answers with a valid citation, refusal rate, wrong answers in the weekly sample, and how quickly gaps in documents are closed.
Risks
Superseded documents, answers from another site’s plan, gap-filling from general knowledge, use during an emergency, and misheard voice questions.

Illustrative example: a 120-worker Tāmaki Makaurau civil contractor

Illustrative example, synthetic numbers: nine active sites, 1,400 pages of approved safety documents, and 2,000 questions a month. Each question retrieves about 3,000 tokens of passages and adds 1,300 tokens of instructions and question, with replies of about 250 tokens.

Line Claude Haiku 5.5 Claude Sonnet 5.5
Price per million input and output tokens US$0.10 and US$0.50 US$2 and US$10
Input cost a month (8.6 million tokens) US$0.86 US$17.20
Output cost a month (0.5 million tokens) US$0.25 US$5.00
Model fees a month about US$1.10 about US$22
Safety advisor review a month about 15 hours about 15 hours

Prices are as listed on 19 October 2026. The review line assumes 215 sampled answers at 2 minutes and 150 refused questions at 3 minutes. Test both models on your own questions before choosing; the cheaper model is often accurate enough for retrieval answers, but the review hours stay the same either way.

HSWA duties the assistant sits inside

  • The primary duty stays with the business. Section 36 of the Health and Safety at Work Act 2015 requires a business to ensure, so far as is reasonably practicable, the health and safety of workers, including safe plant and safe systems of work. An assistant is one way of getting information to workers. It is not a control in itself.
  • Shared sites. Section 34 requires businesses with the same duty to consult, co-operate and co-ordinate. On a site with a principal contractor and subcontractors, the assistant must answer from the principal’s site plan and say whose procedure applies.
  • Notifiable events. Section 56 requires notification to WorkSafe as soon as possible after a business becomes aware of a notifiable event, by the fastest means, and section 57 requires records to be kept for at least 5 years. WorkSafe asks for a call on 0800 030 040 immediately after a death. The assistant routes these to people.
  • Codes of practice are changing. WorkSafe says the Health and Safety at Work Amendment Act commences on 1 April 2027 and introduces a safe harbour for duty holders who follow an approved code of practice. Expect to re-index guidance when codes are reapproved.
  • Privacy. Questions sometimes include injuries or health details. The Privacy Commissioner’s AI guidance applies to the answer log.

Levy Street worked with BetterWorx on PainPal, an AI digital workmate for musculoskeletal pain and injury prevention. Workers describe a task and the discomfort they feel and get practical guidance and next steps. The design lesson carries over: answer from expert-approved content, and hand people to a person when a question goes beyond it.

What supervisors and safety staff keep

Stop-work calls, permits, incident response, deciding whether an event is notifiable, worker consultation, and approving every document the assistant can read.

Where safety assistants go wrong

  • Superseded documents. Last year’s method statement stays in the index and contradicts the current one. Index only the controlled folder and delete superseded versions nightly.
  • The wrong site. A worker on site B gets site A’s traffic plan. Make the site tag mandatory and ask when it is unknown.
  • Filling the silence. When the documents say nothing, a model reaches for a generic figure. Refuse when retrieval finds nothing relevant, and test this deliberately; see RAG retrieval evaluation.
  • Noise and numbers. A voice question on a loud site turns “1.5 metres” into “15 metres”. Read back the question and confirm any number before answering.

The other construction workflows we map are tender document extraction and quote preparation from job notes. The construction sector guide explains where each fits.

Sources

  1. Health and Safety at Work Act 2015, New Zealand Legislation, accessed 19 October 2026
  2. Notify WorkSafe, WorkSafe New Zealand, accessed 19 October 2026
  3. Approved code of practice: Safe farm vehicle operation (safe harbour and Amendment Act commencement), WorkSafe New Zealand, updated 1 October 2026
  4. AI and the Information Privacy Principles, Office of the Privacy Commissioner, 2023
  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 an AI assistant replace our site-specific safety plan?

No. The plan, method statements and permits are how the business manages risk, and the assistant only helps people find what they say. If the assistant gives an answer the documents do not support, that is a fault to fix, not a new rule. Supervisors and the safety team still own every decision on site.

Why not let workers ask ChatGPT safety questions?

A general chatbot answers from general knowledge, not from your site’s plan, your exclusion zones or your chemical register, and it cannot tell which site the worker is on. It may also receive details of incidents and injuries that belong in your own systems. An assistant limited to approved documents, with citations, is the safer pattern.

What happens when the assistant does not know the answer?

It says it cannot find the answer in the approved documents for that site and tells the worker to ask their supervisor. Each of those questions is logged and reviewed weekly by the safety team. Over time the log shows exactly which procedures are missing or unclear, which is useful in its own right.

Can the assistant tell us whether to notify WorkSafe?

It can show WorkSafe’s description of notifiable events and your internal reporting procedure, and it can tell the worker who to call. It should not decide whether an event is notifiable. Under section 56 of the Act, the business must notify WorkSafe as soon as possible, so the decision goes to a supervisor or safety manager straight away.

How do we keep answers current when procedures change?

Index only the controlled folder of approved documents, remove superseded versions automatically, and show the version and date in every citation. Plan a review when codes of practice change: WorkSafe says the Health and Safety at Work Amendment Act commences on 1 April 2027 and changes the legal status of approved codes of practice.

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.