Before a project team puts drawings, specifications or correspondence into an artificial intelligence (AI) tool, two questions need answers: where that information will go, and whether the project permits that use.
Running a model on the company's own hardware can look like the answer to both. It can help control where processing happens. It also makes the company responsible for operating, securing and maintaining the system, and it settles neither the contract question nor whether the tool does the work well enough.
The practical decision is narrower. Does this specific deployment meet this job's requirements, and does it perform the intended task to a standard a qualified reviewer will accept? This article sets out Atlas's recommended approach to that decision, with two records a team can use before any restricted information moves.
Map the whole information path, not only the model
A model is one component of an AI application. Around it sit document conversion, a search index, the chat interface, logs, backups, update services, telemetry, plug-ins and remote-support tools. Any of them can create a copy of project information or send it somewhere else.
A search index is the store of processed text the application searches to find relevant passages before the model answers. It is derived from the project documents, so it needs the same handling rules they do.
Hosted products need the same feature-level reading. Bluebeam's AI transparency page distinguishes features processed locally, features processed in the cloud, and sharing through third-party integrations that depends on permissions.[1] Procore publishes the model providers and infrastructure behind its AI features, and its retention disclosures differ by product.[2] Neither a product name nor a general privacy statement settles what happens to a particular file under a particular account.
Information-boundary record
| Field | What to record | Who confirms |
|---|---|---|
| Information in scope | Document types, source revisions, and any contract or client markings | Project manager or document control |
| Permitted users | Named roles, including any outside parties who will see outputs | Project manager |
| Processing locations | Where each component runs: model, document conversion, search index, interface | IT or system administrator |
| Stored copies | Prompts, outputs, logs, search index and backups; retention period and deletion method | IT or system administrator |
| External connections | Updates, license checks, telemetry, plug-ins, connectors and remote support | IT, using the vendor's documentation |
| Administrative access | Who can administer the system, from where, and how that access is removed | IT or security lead |
| Requirements basis | The contract clauses and client requirements that apply, and who established them | Contract, security or legal personnel |
| Open questions | Each unknown, its owner and the date it must be resolved | Project manager |
| Acceptance | The person authorized to accept the arrangement, and the date | Named decision owner |
Treat local operation as a location choice
Running a model locally changes where processing happens and who is responsible for it. Patching, access control, backups, monitoring and incident handling become the company's work. Local operation does not, by itself, establish confidentiality, contract compliance or export-control compliance.
Which obligations apply depends on the specific information and the agreements and regulations that govern it, not on a sector label. Knowing that a job is a hospital, a semiconductor plant or a defense-adjacent facility does not tell you which rules apply to which documents.
One clause shows why the details matter. Where Defense Federal Acquisition Regulation Supplement clause 252.204-7012 applies to covered defense information, it sets security obligations, including specific requirements when an external cloud service provider handles that information.[3] Whether the clause applies to a given subcontract or document is for the responsible contract, security and legal personnel to establish. Moving processing onto an office server does not show that every applicable obligation is met.
Test the task the team will actually delegate
Readable output is not the same as correct output. The National Institute of Standards and Technology (NIST) generative AI profile identifies confabulation, data privacy and information security among the risks to manage, and describes confabulation as confidently stated but erroneous or false content.[4] The profile is a risk-management framework, not a certification that any model is safe for construction decisions.
We recommend evaluating the actual work. Build a small set of authorized tasks whose correct answers a qualified reviewer establishes before seeing the tool's output. Candidate tasks include summarizing an approved specification section with page references, locating the clause that governs a product, listing the requirements a submittal must address, drafting a request for information from supplied facts, or extracting an equipment list from a schedule. Treat drawing interpretation with particular caution and test it directly rather than assuming it works.
If both a hosted and a local option are permitted, run them on the same task set. Do not assume they perform alike.
Task-evaluation scorecard
| Measure | How to check it | Decide before the trial |
|---|---|---|
| Factual correctness | Compare each statement with the source document and grade errors by consequence | Which error grade blocks acceptance |
| Reference accuracy | Confirm each cited page, section or sheet exists and says what the output claims | Whether any wrong reference is tolerable |
| Omissions | Compare the output with the reviewer's answer key | Which omissions matter for this task |
| Unsupported content | Flag statements with no basis in the supplied documents | How flagged content is handled and reported |
| Consistency | Rerun the same task and compare the answers | What variation is acceptable |
| Review effort | Record reviewer minutes to verify and correct each output | The baseline, such as the current manual method |
| Time to an accepted result | Record active labor and elapsed time separately | Which of the two the decision rests on |
| Boundary behavior | Check logs for any copy or connection not on the boundary record | Whether any deviation stops the trial |
The companion piece on introducing a new tool on a live job sets out the same bounded-trial approach. The companion piece on measuring coordination effort develops the distinction between active labor and elapsed time.
Size the hardware after defining the work
Hardware requirements depend on the model, its memory needs, the document workload, response length and the number of people using it at once. A storage appliance, a workstation and a server can differ widely in what they can run, even when each is described as able to run AI.
We recommend defining the task set and the boundary first, then asking for:
- A demonstration of the exact configuration on representative, authorized work
- Dated price and performance figures for each configuration under consideration
- Operating costs, including support, updates, backup, recovery and administration time
- The update path, and who approves a new model version before it reaches users
- What happens to stored copies when the system is retired or replaced
A result from an unrelated public benchmark cannot establish how quickly your configuration produces a deliverable your reviewer will accept.
Make acceptance observable
Acceptance should leave a record another person can check. Document the approved model version and settings, the tested tasks and their scores, the permitted information paths and the review responsibilities.
Name the changes that trigger re-evaluation: a new model version, a new document type, new users, a new connector or a contract change. For a hosted service, find out how the provider announces model or feature changes. For a local installation, record who approves updates.
A successful deployment gives the team a working tool inside an understood boundary. It does not remove the need for access control, technical review or maintenance. That is the standard for comparing a local option with any permitted hosted option.
Sources
- Bluebeam, AI Transparency, web page, accessed September 28, 2026. Bluebeam AI transparency disclosure. Supports that processing arrangements differ by feature: local, cloud and permission-based integrations. Does not establish the settings of any account or how a particular file is handled; features and terms can change.
- Procore, AI Transparency, web page, accessed September 28, 2026. Procore AI transparency disclosure. Supports that the company publishes model providers and infrastructure for its AI features, with retention disclosures that vary by product. Does not establish retention or contract terms for any specific account.
- Defense Federal Acquisition Regulation Supplement, 252.204-7012 Safeguarding Covered Defense Information and Cyber Incident Reporting, clause text accessed September 28, 2026, paragraph (b)(2)(ii)(D). Official clause text. Supports the conditional statement that, where the clause applies to covered defense information, it sets security obligations including requirements for external cloud service providers. Does not determine whether it applies to any project, document or deployment.
- Autio and colleagues, National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024, introduction and risk overview, PDF p. 8. Primary publication. Supports that the profile identifies confabulation, data privacy and information security among generative AI risks, and its definition of confabulation. It is a risk-management framework, not a certification of any model, product or deployment.
SCOPE
Atlas Construction Consultants is a construction technology consultancy. We help teams document workflows, configure and test tools, including AI-assisted workflows, and train the people who use them. This article is general guidance. It is not legal advice, an export-control or contract determination, a security certification, or a statement that any product or installation is compliant, or that a local model matches a hosted one.