A procurement checklist for construction data and AI
7 min read · Research checked 20 August 2026
Part I introduced the 3Ps of Data Control: Permission, Promise and Prevention. Part II showed how they apply to real construction technology products. This final part turns the framework into a practical checklist for evaluating software before sharing project information.
The checklist should be applied to a specific feature, deployment and data flow, rather than to a supplier as a whole. A provider may have permission to reuse one type of data, promise to restrict the processing of another, and offer a customer-controlled environment for a third. The questions are therefore cumulative: the aim is to understand what happens to each type of information, who can use it, and what control the customer retains.
This checklist is intended to support procurement decisions. It is not a substitute for legal, cybersecurity or professional procurement advice.
| Model | What it means | Questions to ask |
|---|---|---|
| Permission | The supplier may use customer data beyond providing the service. | What data is covered?For what purposes?For how long?Does this include anonymous or derived data?Did we have the right to grant that use? |
| Promise | The supplier processes the data, but its use is restricted by contract and controls. | Who receives the data?Are subprocessors named?What are they allowed to do with it?How long is it retained?What happens when processors or models change? |
| Prevention | The customer controls whether project data leaves its environment. | Who controls administrators and keys?Where does the AI actually run?Can the system work without an external connection?When can data leave?Does the customer deliberately approve that handoff? |
* The same product may sit in different positions for different data or workflows.
** “Vendor does not process” only applies when the customer controls who can access the data and when it can leave. A private environment run by the vendor is still Promise, not Prevention.
Permission covers any right to use customer information beyond returning the customer’s own result. That may include training, product development, benchmarking, research, analytics or commercial data products.
The first task is to establish exactly what has been authorised, including rights over anonymous and derived data.
A customer may be willing to permit reuse without having the right to do so for every element in the file.
Do not accept “anonymised” as a complete answer. Ask what has been removed, what remains, and what new data is created from the original.
Anonymising the project does not necessarily anonymise the knowledge contained in it.
Promise applies when project data enters infrastructure controlled by the supplier or by providers it selects, while contracts, policies and controls restrict what those organisations may do with it.
The customer needs evidence that the supplier, and the providers it relies on, can keep that promise.
For the named feature, map the full path from source file to result. Training is only one possible use; data may still move through several systems even when training is prohibited.
For customer-controlled boundary, follow one file through a feature
Questions to ask at each stage
For each box, establish:
Training is one possible side effect of this journey. Exposure to another system exists even when training is prohibited.
Security evidence should be credited for the question it answers, and no more.
Security protects data from unauthorised use. It does not limit authorised use.
An assurance should be credited for the question it answers and no other
| Assurance signal | Useful evidence of | What it does not prove |
|---|---|---|
| SOC 2 Type II report | Designed controls operating over a stated period, subject to the report’s scope and exceptions | That customer content cannot be used for training; that every product or subprocessor is in scope |
| ISO/IEC 27001 certificate | An information-security management system and risk-management process within the certified scope | A particular data-use restriction, zero retention, or customer-controlled infrastructure |
| Penetration test | Resistance of tested systems to selected technical attacks at a point in time | Contractual confidentiality, processor behaviour, or the security of untested services |
| DPA | Roles and obligations for regulated personal data | Protection of every form of commercially sensitive non-personal construction information |
| NDA/confidentiality clause | A duty to restrict use and disclosure of confidential information | That systems technically prevent access, or that anonymous derivatives remain confidential |
| Data residency commitment | Where specified information is stored or processed | Ownership, no-training, local-only support, or identical residency for logs, backups and AI calls |
| Subprocessor list | The named organisations performing stated functions at the date shown | That the list is complete for project content, or that changes require customer approval |
| No-training statement | A restriction on model development, if the subject and scope are clear | No retention, no human review, no inference processing, no derived analytics, or no additional recipients |
| Zero-data-retention configuration | The relevant processor should not persist submitted content beyond transient inference | That the construction vendor, its index, its logs or its support systems retain nothing |
| Customer-managed keys | Greater customer control over decryption and revocation, depending on implementation | That plaintext is never processed, or that the vendor has no privileged path while the service is running |
| Single tenancy | Reduced risk of accidental cross-tenant access and stronger isolation | Customer control of administrators, keys, backups or outbound processing |
| Local/open-weight inference | The model can operate without sending prompts to a hosted foundation-model API | That the runtime has no telemetry, remote tools, update channel or other outbound connection |
Prevention applies when the architecture allows project information to remain inside a customer-controlled boundary rather than relying only on a supplier promise.
A “private cloud” or single-tenant label is not enough on its own. The practical question is who controls access and when data can leave.
Prevention is about making the boundary visible and enforceable: where the AI runs, when data leaves, and who can access it.
A buyer should not need to reconstruct the data path from multiple policies and security pages. Ask the supplier to complete the following statement for the specific feature and deployment being considered:
For named feature and deployment, the customer provides data classes.
The original is stored in systems and regions for period.
The following derived artefacts are created: list.
They are retained for period and may be used for purposes.
The following organisations receive project information: names and roles.
No recipient may use it for training / improvement / benchmarking, except clearly stated permissions. Human access is possible only for reasons under controls.
Provider changes are governed by notice / approval mechanism.
Deletion removes systems and derived artefacts within period, subject to backup / legal exceptions.
The customer can operate without this external processing by local / customer-managed / non-AI mode, or “not available”.
If the supplier cannot complete that statement clearly, the buyer does not yet understand the data boundary well enough to approve sensitive project information.
Before a drawing, model, estimate, programme, contract or project history enters a new service, answer five questions:
If any answer is “we assume”, stop and resolve it before uploading the data.
The 3Ps are not three labels to assign to a supplier. They are three questions to ask of every important data flow: what have we permitted, what has been promised, and what can the architecture prevent?
Either bring the model to the data, or make the data’s departure explicit.
The full 3P Data Control Check, the five-minute pre-upload check and the supplier statement template, in one printable PDF.
About the research
The profiles are based on first-party terms, privacy policies, security pages, product documentation and AI notices reviewed on 20 August 2026. They are a public-document audit, not a penetration test or legal opinion.
Questions about the research, or about how your product handles customer data? We’d like to hear from you.