AI policy template for construction companies and joint ventures
An AI use policy tells your people which AI tools they may use, with which data, and who answers for what the tools produce. Generic templates stop there. A contractor or a joint venture (JV) also holds documents that belong to others: the owner's drawings, the prices subcontractors sent for one bid, a partner's methods and rates. This template adds what those documents require. Each project, bid and JV gets a separate workspace. Every quantity, request for information (RFI), submittal or claim that leaves the company carries a named person's sign-off. Records are kept as long as the contract and the limitation period require, and vendors and subcontractors are held to the same terms. The template comes in three parts: the policy in twelve sections, an annex of eight clauses an owner can put in a request for proposals (RFP) or project agreement (both in one .docx), and a vendor evaluation checklist with a compliance matrix (.xlsx). Its structure follows NIST's AI Risk Management Framework, and its training section answers Article 4 of the EU AI Act.
- Published
- Updated
- Reading time
- 17 min
- Written by
- Jules Toussaint
This guide is for contractors and joint ventures in the US, Canada, the UK and the EU, and for owners' procurement teams. It cites laws and frameworks where the template relies on them, and you should adapt the text with your counsel. It is not legal advice.
Who wrote this, and why. Mirage Metrics builds AI agents that read construction documents. They run in the client's cloud, on its premises or hosted by Mirage, with a named approver set for each workflow (see security and approvals and audit). Mirage has no published deployment on AI policy work. Its nearest published construction proof is OCBM: three agents in production, with 100% of quantities tied to a source drawing. That is the traceability that section 6 of this template asks of any tool. The vendor checklist is neutral, so Mirage's own answers are in its procurement and security pack, not in the file.
What generic AI policy templates miss on a construction project
A contractor handles documents it does not own, and that is where a generic policy falls short. A policy written for a business that owns its documents can stop at approved tools and banned uses. Your estimating drive, document control system and project inboxes, by contrast, hold material under at least four other sets of terms. An AI tool that indexes all of it ignores those terms unless the policy says otherwise. The table below sorts what a construction team handles by whose it is and by what already governs its use. Its last column gives the template's default rule, which section 4 turns into data classes.
| What the team handles | Whose it is | What already governs its use | Template default |
|---|---|---|---|
| Drawings, specifications and models from the owner and its designers | The owner, or the designers under their appointments | The contract's confidentiality and intellectual property clauses | Project confidential: approved tools only, inside the project's workspace, never used to train a vendor's model |
| Quotes and prices received from subcontractors and suppliers for a bid | Each bidder | The bid invitation's terms and any non-disclosure agreement | Project confidential, held in that bid's workspace; reused for another bid only with the sender's written permission |
| A JV partner's methods, rates, resources and past bids | That partner | The joint venture or consortium agreement | Project confidential, held in the JV's workspace and used for the JV's project only |
| Workers' names, certifications, timesheets, site photos showing faces | The people concerned | Privacy law: the GDPR in the EU, national, state or provincial law elsewhere | Personal: tools under processor terms only, limited to what the use needs |
| The company's own estimates, cost history and procedures | The company | This policy | Internal: approved tools under a company contract with no-training terms |
| Purchased standards and paid cost databases | Their publishers | The licence | Only as the licence allows |
| Published tender notices and other public documents | Nobody in particular | Accuracy only | Public: any approved tool |
A generic template has no answer for the second and third rows. A quote that a subcontractor sent for one bid should not surface when the same estimator prices the next one. Nor should a partner's rates follow your staff out of the JV. Yet search and retrieval tools make both easy by default, because they index whatever they are pointed at. NIST's Generative AI Profile asks organisations to inventory every third party with access to their content and to keep approved provider lists (GV-6.1-007). It also asks them to extend acceptable use policies to contractors, consultants and other third-party personnel (GV-6.1-010) (NIST AI 600-1).
The policy template, section by section
Here is what the policy sets, section by section, before you download it. The .docx has twelve sections and three annexes. The annexes hold the data classes with a register of approved tools, the owner clauses, and the sign-off record. In the table, the right column says which framework or law each section answers. NIST's AI Risk Management Framework is voluntary and organises the work in four functions: GOVERN, MAP, MEASURE and MANAGE. A policy sits in GOVERN (NIST AI 100-1, Table 1). The laws, by contrast, are not voluntary.
| Section | What it sets | What it answers |
|---|---|---|
| 1. Purpose and scope | Everyone who works for the company or for a JV it belongs to, and every AI tool, features built into other software included, on any device | NIST AI 600-1, GV-6.1-010 |
| 2. Definitions | AI tool, approved tool, data classes, workspace, issued document, responsible person | The terms the other sections rely on |
| 3. Roles | Policy owner, executive sponsor, tool owners, a data owner for each workspace, approvers for each workflow | NIST AI RMF, GOVERN 2.1 and 2.3 |
| 4. Approved tools and data classes | A register of approved tools and the data classes each may hold | GOVERN 1.6 (inventory of AI systems); GV-6.1-007 (approved provider lists) |
| 5. Separation | One workspace per project, bid and JV; no retrieval across workspaces | The contract's confidentiality clauses; JV agreements |
| 6. Human sign-off | Every issued document checked against its sources and signed by a named person | GOVERN 3.2; AI 600-1 risks Confabulation and Human-AI Configuration |
| 7. Records | Sources, tool, reviewer, date and decision, filed with the project record | FAR 4.703; Limitation Act 1980, ss.5 and 8 |
| 8. Vendors and subcontractors | Minimum contract terms, flowed down to anyone who uses AI on company or project data | GV-6.1-004, GV-6.1-006, GV-6.1-009; GDPR Article 28 where it applies |
| 9. Prohibited uses | Six prohibitions and a line for your own, among them AI decisions about people without HR and legal review | AI Act, Annex III, point 4 |
| 10. Incidents | What to report, to whom and how fast; who decides on notifications to owners, partners and authorities | GOVERN 4.3; GDPR Article 33 |
| 11. Training | Before access to a tool, refreshed when tools change, completion recorded | AI Act Article 4; GOVERN 2.2 |
| 12. Review | At least yearly, and after a new tool, a new contract requirement or an incident | GOVERN 1.5 |
The file holds the policy, Annex A (data classes and the register of approved tools), Annex B (the eight owner clauses) and Annex C (the sign-off record). Every name, period and threshold is a bracketed field for you to fill in.
Approved tools and data classes
Section 4 lets data go only into a tool approved for its class. The template defines four classes: Public, Internal, Project confidential (what the owner, bidders and partners entrust to you) and Personal. Its default is strict where a generic policy is loose. Project confidential data goes only into tools that hold each project, bid and JV in a separate workspace, under a contract that forbids training on it. Internal data needs a tool under a company contract with no-training terms, and you can change that line if your risk appetite differs. Section 9 then spells out what follows: no Project confidential or Personal data goes into a public chat assistant used on a personal account. The Government of Canada applies a comparable rule to its own staff. Its public servants must not input personal information into publicly available online generative AI tools (Guide on the use of generative AI).
Separation between projects, bids and joint ventures
Section 5 gives each project, each bid and each JV its own workspace in every approved tool that holds Project confidential data. That workspace's data owner sets who has access. Indexes, embeddings, prompts and outputs stay where they were created. Quotes received for a bid stay with that bid unless the sender agrees otherwise. In the same way, a partner's documents serve the JV's project only, unless the JV agreement says otherwise. When a bid is lost or a JV ends, the data owner decides what is archived and what is deleted, under the retention rule of section 7.
Human sign-off on issued documents
The template defines an issued document broadly. It is anything that leaves the company or that others rely on, from quantities and bid prices to RFIs, submittals, non-conformance reports, schedules, claims and letters to the owner. None is issued without a named person's sign-off, and that person must be competent to issue it without the tool.
The reviewer checks the document against its sources, not for plausibility. NIST names the failure this guards against. On the tool's side it is confabulation, "the production of confidently stated but erroneous or false content". On the human side it is automation bias and over-reliance (NIST AI 600-1, section 2). That is why section 6 requires every figure, date and clause reference in an AI-assisted document to trace back to a sheet, page or clause and its revision. An output without sources is treated as a draft. The Government of Canada's guide puts the same duty in one word of its FASTER principles: Accountable, "take responsibility for the content generated by these tools" (Government of Canada).
Responsibility does not move to the tool. The AI vendor is not a party to the construction contract between owner and contractor, and section 6.5 of the template says so: using a tool changes nothing about who answers for a document under the contract or under professional rules. Whether you disclose AI assistance to the owner is set by the contract, or by clause B1 where the owner has adopted it. If your team is testing a takeoff tool before trusting it on a bid, see how to test AI takeoff software.
Records and retention
Section 7 sets what is kept for each AI-assisted issued document, and where. It keeps the sources used, the tool and its version where known, the reviewer, the date and the decision (Annex C). These are filed with the project record rather than left in the tool. How long you keep them is a contract question. On US federal contracts, FAR 4.703 has contractors make records available for 3 years after final payment. It counts as records "other data, regardless of type", computer data included (FAR 4.703). In England and Wales, an action founded on simple contract cannot be brought more than six years after the cause of action accrued. For an action upon a specialty, a contract made as a deed among them, the period is twelve years (Limitation Act 1980, s.5; s.8). Provincial and state periods differ, so the template leaves the number blank for counsel to fill.
Training and AI literacy
Section 11 covers training, and in the EU it answers Article 4 of the AI Act. Article 4 has applied since 2 February 2025 (Article 113(a)). As amended by Regulation (EU) 2026/1744 in July 2026 (EUR-Lex), it requires providers and deployers of AI systems to "take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf", taking into account their knowledge, experience and the context of use, without having to guarantee any individual's level (Article 4). A contractor that uses AI systems in its business in the EU is, in the Act's terms, a deployer (Article 3(4)). To meet Article 4, section 11 trains users before they get access, refreshes the training when a tool changes and records completion. NIST's GOVERN 2.2 asks the same for personnel and partners.
Clauses an owner can put in an RFP or project agreement
If you buy construction work, Annex B gives your procurement team eight clauses to adapt. They are written in contract register, with bracketed references to the contract's own defined terms. They do not assume the contractor uses AI: they apply if it does. The same clauses work further down the chain too. A contractor can propose them to its subcontractors, and a JV can write them into the agreement between its partners.
The clearest public model is the US federal one. OMB memorandum M-25-22 (April 3, 2025) governs AI that agencies acquire, from solicitations issued 180 days after the memo. It expressly does not govern "AI used incidentally by a contractor during performance of a contract", meaning AI a contractor uses at its own option (M-25-22, sections 2(b)(v) and 2(c)). Even so, its privacy paragraph asks agencies to set terms whenever a contractor's AI system handles Federal information containing personally identifiable information (3(d)). Annex B follows the memo's contract terms on four points. Contracts must "permanently prohibit the use of non-public inputted agency data and outputted results to further train publicly or commercially available AI algorithms" absent explicit agency consent (3(e)(iv)). Vendors collect and retain agency data only when reasonably necessary to serve the contract (3(e)(iii)). Ownership and intellectual property rights are delineated (3(e)). And when a contract ends, agency and vendor agree on the format and usability of the data and plan its transfer (4(f)(i)).
| Clause | What the contractor commits to | Public model |
|---|---|---|
| B1 Disclosure of AI use | A list of the AI tools used to prepare deliverables, on request and whenever it changes, with the vendor, the purpose and where owner information is processed | Canada's FASTER principles: Transparent |
| B2 Use of owner information | Owner information used only to perform the contract; no training or improvement of any AI model on it, or on outputs derived from it, without written consent | M-25-22, 3(e)(iv); NIST AI 600-1, GV-6.1-004 |
| B3 Human responsibility | Each AI-assisted deliverable reviewed and issued by a named, competent person; no obligation of the contractor reduced or transferred | Canada's FASTER principles: Accountable |
| B4 Audit trail | For each AI-assisted deliverable: sources, tool, reviewer and date, kept for the records period and produced on request | NIST AI 600-1, GV-6.1-008; FAR 4.703 on federal contracts |
| B5 Ownership of outputs | Deliverables follow the contract's intellectual property clause as if made without AI; no AI vendor acquires rights in owner information or deliverables | M-25-22, 3(e); NIST AI 600-1, GV-6.1-004 |
| B6 Confidentiality | AI vendors bound by confidentiality at least as protective as the contract's; the contractor answers for them | M-25-22, 3(e)(iii) |
| B7 Flow-down | Equivalent obligations in every subcontract and supply agreement under which owner information goes into an AI tool | NIST AI 600-1, GV-6.1-010 |
| B8 Return and deletion at exit | Owner information returned in an agreed format and deleted from AI tools, with written confirmation, unless law or the records clause requires keeping it | M-25-22, 4(f)(i); GDPR Article 28(3)(g) for personal data |
Two details make the clauses work in practice. For clause B2, ask for the commitment in the contract itself, not on a web page the vendor can change. For clause B4, set the records period to match the contract's records clause, so that the contractor keeps one set of records, not two.
The vendor evaluation checklist and compliance matrix
You use the checklist when you buy a tool: the .xlsx is the buyer's side of section 8. It lists 26 requirements in nine areas, each with why it matters and the evidence to ask for. The three sheets are identical, so you can score three vendors side by side. For each requirement you set a priority (Must have or Should have) and record the vendor's answer and the evidence received. Then you choose a status: Met, Partly met, Not met, Not applicable or Not yet answered. The Compliance matrix sheet counts the statuses per vendor and per area, with COUNTIFS on those labels. It also counts the Must have requirements not met, and gives the share met with its denominator (requirements less those not applicable).
| Area | What to require | Evidence to ask for |
|---|---|---|
| Data use | No training or improvement of any model on your data or outputs without written consent; how long the vendor keeps prompts, files and outputs | The clause in the signed contract or data processing terms, not a web page |
| Data location and subprocessors | Countries and cloud regions, backups included; named subprocessors, model providers among them, with notice before changes; your cloud or premises when a contract requires it | Written statement; subprocessor list; change-notice clause (GDPR Article 28(2) where it applies) |
| Separation and access | Workspaces per project, bid and JV with no retrieval across them; access by role through your identity provider; revocation effective everywhere | A demonstration on two test workspaces and a revoked test user |
| Traceability and audit | Every output cites its document, sheet or page and revision; an exportable log of who did what, where, with which sources | An output on your own documents; a log export |
| Human review | Outputs that need a check are marked; approvals recorded by name and date; a named approver for each workflow | The approval screen and its record |
| Security and incidents | Encryption, access control, incident notification within a stated time, service levels | Security documentation; an independent audit report or certificate if the vendor holds one, with its scope and date |
| Model changes and quality | Notice before changes that affect outputs, with time to test; known limits stated; a test on your own past documents before contract | Change-notice clause; written limits; pilot terms |
| Exit and continuity | Export of data, outputs and logs in usable formats; deletion with written confirmation | An export sample; the deletion clause |
| Legal role | Its role under privacy law and under the AI Act; no rights claimed in your inputs or outputs | Data processing agreement; written statement; intellectual property clause |
NIST lists the same ground for any generative AI purchase. It names contracts and service level agreements that specify content ownership, usage rights, quality standards and security requirements (GV-6.1-004). It adds clauses that let you evaluate a third party's processes (GV-6.1-006), and due diligence that covers intellectual property, data privacy and security (GV-6.1-009) (NIST AI 600-1). Where the EU's General Data Protection Regulation (GDPR) applies, a vendor processing personal data for you is a processor. The contract must then bind it to process only on your documented instructions and to keep its staff to confidentiality. It must also bind the vendor to secure the data, to use another processor only with your prior written authorisation, to delete or return the data at the end, and to allow audits (Article 28(2) and (3)).
Score each vendor on the evidence column, not on its answers. The checklist holds nobody's answers, Mirage's included. For a test on your own documents, the document automation pilot guide sets the protocol. For the security controls to put in the contract itself, see construction AI data security.
The file holds three vendor sheets with the 26 requirements, a read-me, and the Compliance matrix sheet with its formulas. The priority, answer, evidence and notes columns are empty, and every status starts at Not yet answered.
Putting it in place: the company first, then each project
Roll the policy out at company level first, then on each project. The order matters, because each step uses the previous one.
- Inventory the AI tools already in use, including features built into existing software. NIST's GOVERN 1.6 asks for exactly this inventory.
- Name the policy owner, the executive sponsor and the tool owners (section 3).
- Score each tool with the vendor checklist, record the outcome in the Annex A register, and retire the tools that fail.
- Adopt the policy through your usual route. Where staff are represented, check first whether employee representatives must be informed or consulted. In France, the labour code sets the procedure (see the French version of this guide).
- Train users before they get access (section 11) and record who completed the training.
- At the start of each project, bid or JV, create the workspaces, set the roster and name the approvers for each workflow. Read the contract and the JV agreement too, for AI terms that override the defaults.
- Review after the first project and then at least yearly (section 12), and after any incident.
When this template is not enough
The template covers ordinary use of AI on documents. In five situations, you need more than a policy.
- AI that makes or shapes decisions about people. Systems for recruitment, for decisions on work relationships, for allocating tasks based on behaviour or personal traits, or for monitoring and evaluating performance are high-risk under the EU AI Act (Annex III, point 4). Deployers' obligations for those systems apply from 2 December 2027 (Article 113(c)(i)). Among them are human oversight by persons with the necessary competence, training and authority, and informing workers' representatives and the affected workers before use at the workplace (Article 26(2) and (7)). Section 9 forbids such use without HR and legal review, and that review is where the work is.
- Operating infrastructure. A contractor may also operate assets, in a concession's operating phase for instance. It may then use AI as a safety component in managing road traffic or the supply of water, gas, heating or electricity, a use that Annex III, point 2, also lists as high-risk. That calls for an engineering safety case, not a document policy.
- Personal data at scale. Where the GDPR applies and a processing, particularly one using new technologies, is likely to result in a high risk to people, an impact assessment comes before it (Article 35(1)).
- An owner with its own rules. A US federal solicitation may carry terms written under M-25-22, and Government of Canada institutions have their own guidance. In either case, the solicitation prevails over your defaults.
- A management system. If an owner or a parent company asks for an AI management system, the reference is ISO/IEC 42001:2023, which specifies requirements for establishing, implementing, maintaining and continually improving one (ISO; the standard is sold). A policy is one document inside such a system, not a substitute for it.
Questions readers ask
An AI use policy for a construction company should include twelve parts: scope, definitions, roles, approved tools and data classes, separation, human sign-off, records, vendor and subcontractor terms, prohibited uses, incident reporting, training and a yearly review. The roles include a data owner for each project and an approver for each workflow. Separation means a separate workspace for each project, bid and joint venture. Every issued document needs a named person's sign-off, and records of AI-assisted documents are kept as long as the contract requires.
Under this template's default, staff may not put project drawings into a public AI chatbot. Owner drawings are Project confidential data, so they may only go into approved tools that keep each project in its own workspace, under a contract forbidding training on it. A public chatbot used on a personal account meets none of those conditions. The Government of Canada applies a comparable rule to its own staff for personal information.
The EU AI Act's Article 4, applicable since 2 February 2025, requires providers and deployers of AI systems to take measures to support the development of AI literacy of their staff and of others who use AI systems on their behalf. As amended in July 2026 by Regulation (EU) 2026/1744, it does not require any individual's level to be guaranteed. The template meets it with training before access to a tool and a record of who completed it.
The template's Annex B proposes eight AI clauses an owner can put in a construction contract: disclosure of AI tools, use of owner information, human responsibility, audit trail, ownership of outputs, confidentiality, flow-down, and return and deletion at exit. The contractor discloses the AI tools used for deliverables. Owner information is used only for the contract and never to train a model without written consent. A named person is responsible for each AI-assisted deliverable, and the audit trail records sources, tool, reviewer and date. Outputs are owned under the contract's intellectual property clause, confidentiality binds AI vendors, and the same obligations flow down to subcontractors.
The person who issues a quantity or an RFI drafted with AI is responsible for it, and so is the company behind them. The template requires a named, competent person to check every AI-assisted document against its sources and sign it. It also states that using a tool changes nothing about who is responsible under the contract. The AI vendor is not a party to the construction contract between owner and contractor.
Sources and scope
Checked on . Every source was read on 10 November 2026. The AI Act is cited in its consolidated text of 27 July 2026, as published on the European Commission's AI Act Service Desk. ISO/IEC 42001 is sold: the page relies on ISO's own description of it.
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023: the four functions GOVERN, MAP, MEASURE and MANAGE; Table 1, GOVERN 1.5, 1.6, 2.1, 2.2, 2.3, 3.2, 4.3 and 6.1
- NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024: section 2, the 12 risks (Confabulation, Data Privacy, Human-AI Configuration, Intellectual Property, Value Chain and Component Integration among them); suggested actions GV-6.1-004, GV-6.1-006, GV-6.1-007, GV-6.1-008, GV-6.1-009 and GV-6.1-010
- European Commission, AI Act Service Desk, Article 4 (AI literacy), consolidated text as amended by Regulation (EU) 2026/1744: measures to support the development of AI literacy of staff and others acting on the provider's or deployer's behalf; no specific level to guarantee
- European Commission, AI Act Service Desk, Article 3 (definitions): point (4), deployer: a person or body using an AI system under its authority, except in a personal non-professional activity
- European Commission, AI Act Service Desk, Article 113 (entry into force and application), consolidated text: Chapters I and II from 2 February 2025; Chapter III, Sections 1 to 3, from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems
- European Commission, AI Act Service Desk, Annex III (high-risk AI systems): point 2, critical infrastructure; point 4, employment, workers' management and access to self-employment
- European Commission, AI Act Service Desk, Article 26 (obligations of deployers of high-risk AI systems): paragraph 2, human oversight by persons with the necessary competence, training and authority; paragraph 7, informing workers' representatives and affected workers
- Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), Official Journal of 24 July 2026: amends Regulation (EU) 2024/1689, Articles 4 and 113 among others
- OMB, Memorandum M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, April 3, 2025: covered AI (2(b)) and the exclusion of AI used incidentally by a contractor (2(b)(v)); application 180 days after issuance (2(c)); privacy (3(d)); IP rights and use of government data (3(e)); vendor lock-in (4(c)(ii)); contract terms (4(d), items A to G); contract closeout (4(f)(i))
- Government of Canada, Guide on the use of generative artificial intelligence (page modified September 24, 2026): the FASTER principles; public servants must not input personal information into publicly available online generative AI tools
- Federal Acquisition Regulation, 4.703 Policy (contractor records retention), acquisition.gov: records, computer data included, made available for 3 years after final payment, with the exceptions of subpart 4.7
- Limitation Act 1980, section 5, legislation.gov.uk: six years for an action founded on simple contract
- Limitation Act 1980, section 8, legislation.gov.uk: twelve years for an action upon a specialty
- Regulation (EU) 2016/679 (GDPR), Articles 28, 33 and 35: processor contract terms (28(2), 28(3)(a) to (h)); breach notification within 72 hours where feasible (33(1)); impact assessment where a high risk is likely (35(1))
- ISO, ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system (standard page; the text is sold): edition 1, December 2023; requirements for establishing, implementing, maintaining and continually improving an AI management system
The policy, the clauses and the checklist are templates for general information. They are not legal advice and do not replace a reading of your contracts and of the law that applies to you by a qualified professional.
Related on this site
- Document automation pilot: how to evaluate a vendor: test a tool on your own documents
- How to test AI takeoff software: a protocol for quantities before a live bid
- Procore and AI document workflows: where an agent reads and writes in Procore
- Document control with AI agents: transmittals, registers and status codes
- Approvals and audit: how Mirage records a named approver's decision
- Security: Mirage's deployment and data commitments
- Procurement and security pack: Mirage's answers for buyers' reviewers
- Construction AI data security: security controls to put in the contract
Show us what your AI policy lets an agent do
Leave your work email, then pick a time for a 30-minute call on the documents your teams would hand to an agent and what your policy allows. You leave with a written plan: what an agent could prepare, where it would run and who signs off. No commitment.