What should a government contractor's AI use policy include?
No rule requires the policy. Several rules assume you have one. The seven sections a Costpoint shop's policy needs, with the clause or the vendor document behind each. Research as of November 4, 2026.
The short answer: Seven sections: a named owner with an approval path for new tools; a data-by-tool matrix that says which classes of data (public, internal business data, FCI, CUI, ITAR technical data) may enter which classes of tool, with CUI and ITAR confined to environments that meet DFARS 252.204-7012; the approved tools by tier, with consumer accounts barred from work data; the AI already switched on, which at a Deltek shop means Dela inside Costpoint and Copilot Chat in Microsoft 365, with who may use it and what may connect to it; a rule that a person checks AI output before it reaches a timesheet, a bill, a proposal or the government, with the source kept; an incident path that meets the 72-hour reporting clock in DFARS 252.204-7012; and a review date. As of November 4, 2026 no FAR clause requires the policy, but NIST SP 800-171 requirement 3.1.20, 32 CFR 170.19 and, as law firm Bradley reads it, Section 1532 of the FY2026 NDAA each assume one exists.
The board asked for an AI policy and IT has a vendor's template. This is the version a CMMC assessor, DCAA and a contracting officer would look for, written for a Costpoint shop.
No rule requires the policy, but several rules assume you have one
As of November 4, 2026 we found no FAR or DFARS clause that says a contractor shall have an AI use policy. What exists is a set of rules that cannot be met without one.
The first is NIST SP 800-171 Rev 2, still the baseline the Department of War named in the July 2026 memo that suspended CMMC Phase 2. Requirement 3.1.20 reads: "Verify and control/limit connections to and use of external systems." Its discussion names cloud services as external systems, and says organizations may set terms for their use and restrict personnel from using systems whose owners will not agree to them. A SaaS AI tool is an external system under that sentence, and a policy naming the approved ones and their terms is the control.
The second is the CMMC program rule. 32 CFR 170.19 says that before a Level 2 assessment the organization must consider whether an external service provider is a cloud service provider handling CUI; one that is "shall meet the FedRAMP requirements in 48 CFR 252.204-7012" and goes in the System Security Plan. That clause, DFARS 252.204-7012, defines rapid reporting of a cyber incident as "within 72 hours of discovery." A prompt that carried CUI to the wrong tool is the kind of event the policy has to route in time, whether or not it turns out to be a reportable cyber incident.
The third is new. Section 1532 of the FY2026 National Defense Authorization Act (Public Law 119-60), as the Congressional Research Service summarizes it, prohibits DoD from using or acquiring covered AI systems, including those from DeepSeek and High Flyer; law firm Bradley's August 2026 guide reads it as barring DoD contractors from using covered AI in performing DoD contracts. If the policy does not name the banned tools, nothing stops an engineer from pasting a deliverable into one.
None of these says how to write the policy. NIST's AI Risk Management Framework (January 2023) and its Generative AI Profile, NIST AI 600-1 (July 2024), are voluntary, and its action GV-3.2-003 is the plainest instruction we found: "Define acceptable use policies for GAI interfaces, modalities, and human-AI configurations." Deltek itself says an AI Council owns "our company-wide AI policy and review/approval of acceptable tool usage." Here is what we would put in yours.
Start with the data, not the tool
Most policies begin with a list of approved products, and that list is stale within a quarter. The durable part is the matrix: which class of data may enter which class of tool. For a Costpoint shop the data classes are public information, internal business data that is not CUI (rates, payroll, billing, project financials), federal contract information, CUI and ITAR technical data; the tool classes are consumer AI apps, enterprise AI under signed terms, and environments rated for CUI.
The tiers are the point, and the vendors state them. OpenAI's help page says that on Free, Plus and Pro ChatGPT "we may use your content to train our models" by default, while its enterprise privacy page says Business, Enterprise and the API do not. Microsoft calls the personal-account Copilot "a consumer version" and says to "never add sensitive or proprietary work information in a prompt," while Copilot Chat with a work account carries enterprise data protection. One product name, two promises. The policy names the tier.
| Data class | Costpoint examples | Consumer AI app | Enterprise AI, signed terms | CUI-rated environment |
|---|---|---|---|---|
| Public information | Website copy, press releases | Yes | Yes | Yes |
| Internal business data, not CUI | Pool rates, payroll, billing, project status | Depends: consumer terms may allow training; most policies say no | Yes: terms must rule out training and set retention | Yes |
| Federal contract information | Contract documents not for public release | No | Depends: the 15 safeguards in FAR 52.204-21 still apply | Yes |
| CUI | Technical data, some deliverables | No | Usually no: needs the FedRAMP Moderate requirements in DFARS 252.204-7012 | Depends: the service must be in your SSP and assessment scope |
| ITAR technical data | Export-controlled drawings | No | No | Depends: US-person operations and support in the contract |
Sources: NIST SP 800-171 Rev 2 requirement 3.1.20; 32 CFR 170.19; DFARS 252.204-7012 paragraph (b)(2)(ii)(D); FAR 52.204-21; OpenAI's help page on training data; Microsoft's enterprise data protection and consumer Copilot pages; Deltek's AI and Trust page. "Yes" means the data class may enter the tool class under the stated condition; the contract text and the contracting officer control. Our reading, not legal advice. Dela inside Costpoint sits in the middle column for commercial cloud and on-premises customers; see the next section for GovCon Cloud Moderate.
Two notes on the middle column. Costpoint's finance data is sensitive but usually not CUI, so an enterprise tier under signed terms can handle most of a controller's questions. And "signed terms" means the contract, not the web page: Deltek's AI and Trust page says "our customers' data will not be used to train large language models," and that promise belongs in your file next to OpenAI's and Microsoft's. We drew the CUI line in Is it safe to put CUI or contract documents into ChatGPT or Copilot?
Name what is already switched on in Costpoint
A policy that covers ChatGPT and misses the AI inside the ERP has missed the one every Costpoint user can reach. Dela is included with the Costpoint license (on-premises customers bring their own Microsoft model license, per Deltek's FAQ), and the Costpoint 8.2 help for the Dela tab in Configure System Settings lists choices an administrator has already made, knowingly or not. "When Dela is enabled initially, it is enabled for all users," with a subtask to restrict it to specific users or groups. One checkbox clears Ask Dela chat history on logout; another switches Dela off inside Microsoft Teams. A Help Content Links subtask lets you point Ask Dela at your own documents, which is a place to put the policy itself.
Then the connections. The same tab has an Enabled checkbox for the Dela MCP server connection, which lets outside AI clients connect to Dela, with a Trusted MCP Clients subtask to decide which. A separate MCP Servers (Beta) subtask runs the other way: it registers outside MCP servers that Dela may call, with an authorization type of No Auth, API Key or OAuth, and a Restrict User Access checkbox. Deltek's pages do not state what an outside client can reach beyond "Dela's capabilities" as the signed-in user, or whether either direction exists in GovCon Cloud Moderate. The policy records the settings you chose and who may change them; GCCM customers get Dela's status in writing, since Deltek's own documents disagree.
Say the same about Copilot Chat, included with eligible Microsoft 365 plans. Deltek's commitments for Costpoint (stateless requests, nothing stored on Deltek's side, no training) are the standard any added tool has to meet.
Say what happens to the output
This is the section a DCAA auditor would read. AI drafts and flags; people approve and sign. Put that sentence in the policy, then say what it means wherever a figure enters a record. DFARS 252.242-7006 lists "Approval and documentation of adjusting entries" among the criteria of an acceptable accounting system, and the clause does not care whether the entry came from a spreadsheet or a model. A voucher, a bill or a journal entry that AI prepared is still posted by an accountable person. Costpoint's own Dela features keep that trail: the 2026.3 release notes describe Dela recommending or, if you allow it, approving vouchers above a threshold while storing the confidence level, the date and the reason. Whether to allow the approval is a policy decision.
The same rule covers what leaves the building. A proposal statement no one checked is, on our reading, the reckless-disregard case the False Claims Act describes, as we set out in Can AI draft government proposals from past performance without compliance risk? So every AI output that reaches a submission gets a named reviewer and keeps its source. FAR 52.215-2 defines records to include "computer data," and FAR 4.703 keeps them three years after final payment; a vendor's 30-day prompt retention is not your records policy. Every answer shows where it came from. No black boxes.
Give it an owner, an incident path and a review date
The remaining sections are short. An owner: one named role that approves tools, keeps the matrix current and answers the assessor; the ABA's Procurement Lawyer guide (Gatto and Bourne, summer 2024) puts it as "in all cases, companies must ensure proper AI governance." An incident path: who an employee tells when CUI went into the wrong tool, and how that reaches the person who files with DoD inside 72 hours. Banned categories, because the tools a policy never names are the ones it cannot control: browser extensions that read the screen, and meeting note-takers. Mississippi's state IT agency, in its November 2025 acceptable use policy for AI, requires notice before an AI note-taker joins a meeting, a rule worth borrowing.
And a review date, because this field moved three times this year. The Department of War suspended CMMC Phase 2 on July 13, 2026. GSA revised its proposed AI safeguarding clause, GSAR 552.239-7001, in June; comments closed August 3. Deltek's Costpoint release notes dated April and July 2026 added MCP connections, and the 8.2 help now documents them in both directions. For a Level 2 contractor the annual affirmation under 32 CFR 170.22 is the natural anchor: the Affirming Official attests, by name, that the controls this policy implements are in place.
A policy written this way is also the specification for anything built later: which data an assistant may read, where it runs, how it reaches Costpoint, and that it answers with sources and never posts. We read it before the first task. The client's data stays the client's. Nothing leaves their control.
Common questions
Does the FAR require a government contractor to have an AI use policy?
As of November 4, 2026, we found no FAR or DFARS clause that requires one. The requirement arrives sideways. NIST SP 800-171 Rev 2 requirement 3.1.20 says to verify and control or limit connections to and use of external systems, and its discussion names cloud services and says organizations may set terms for their use and restrict personnel from using systems whose owners will not agree to them; a SaaS AI tool is such an external system. 32 CFR 170.19 puts a cloud service that handles CUI into your CMMC assessment scope and your System Security Plan. Section 1532 of the FY2026 NDAA, as law firm Bradley describes it, bars covered AI such as DeepSeek from use in performing DoD contracts. A written policy is how you show an assessor, or a contracting officer, that you meet all three.
Can employees at a government contractor use ChatGPT or Copilot for work?
The policy answers by tier and by data, never by brand. OpenAI's pages say Free, Plus and Pro ChatGPT may use your content to train its models by default; Enterprise, Business and the API do not. Microsoft tells organizations that consumer Copilot is a free consumer version and to never add sensitive or proprietary work information in a prompt, while Copilot Chat with a work account carries enterprise data protection. So consumer accounts are out for anything but public information. Enterprise tiers under signed no-training terms can handle internal business data that is not CUI, which covers most Costpoint finance data. CUI needs a service that meets the FedRAMP Moderate requirements in DFARS 252.204-7012, and ITAR technical data needs US-person operations as well.
Does Deltek's Dela need to be in the policy if it is already inside Costpoint?
Yes, because once enabled it is on for every Costpoint user. The Costpoint 8.2 help for the Dela tab says that when Dela is enabled initially it is enabled for all users, with a subtask to restrict it to specific users or groups, a checkbox to clear Ask Dela chat history on logout, a checkbox to disable Dela for the Microsoft Teams integration, and an Enabled checkbox for the Dela MCP server connection with a Trusted MCP Clients subtask. Deltek says customer data is not used to train large language models and that Deltek does not store information processed through Dela. The policy should record which of those settings you chose and who may change them. GovCon Cloud Moderate customers should also note that Deltek's two documents on Dela in GCCM disagree, and get the answer from their account team in writing.
How often should a contractor review its AI use policy?
At a fixed date at least yearly, and whenever a rule or a vendor setting changes. The rules and the vendor settings moved three times in 2026 alone: the Department of War suspended CMMC Phase 2 on July 13, GSA revised its proposed AI safeguarding clause in June with comments closing August 3, and Deltek's Costpoint release notes dated April and July 2026 added MCP connections to Dela. For a Level 2 contractor the annual CMMC affirmation under 32 CFR 170.22 is a natural anchor, because the Affirming Official is attesting to the controls the policy implements. NIST's Generative AI Profile also suggests updating acceptable use policies to address proprietary and open-source AI technologies and data.