Control Before Capability.

• AI may analyze.

• AI may compare.

• AI may identify contradictions.

• AI may summarize.

• AI may calculate.

• AI may identify patterns.

• AI may surface risk.

• AI may prepare recommendations.

• AI may assist in producing reports.

But AI does not become the responsible party.

Final commercial, contractual, estimating, operational and professional decisions remain subject to human authority.

There must always be an identifiable person or business function accountable for consequential decisions.

AI can inform authority.  It cannot become authority.

Access to information does not create ownership of that information.

Client information is to be used only for authorized operational purposes.

The governing standard prohibits unauthorized:

• Sale of client information

• Repurposing of client information

• Distribution of client information

• Cross-client exposure

• Public disclosure

• Contribution of client information to public AI training systems

• Use outside the authorized business purpose

Possession of data is not permission to use it without boundaries.

Where information resides matters.

Our infrastructure strategy is based on controlled data placement, jurisdictional awareness, system segmentation and defined access boundaries.

We do not believe sensitive project intelligence should move indiscriminately between systems simply because technology makes that possible.

Data placement is evaluated according to factors including:

• Sensitivity

• Business purpose

• Client requirements

• Jurisdiction

• System classification

• Operational need

• Access requirements

• Retention requirements

• Integration risk

Detailed physical data center locations and infrastructure information are not publicly disclosed when doing so could unnecessarily increase security exposure.

Where appropriate, infrastructure region and data-handling information can be addressed directly with customers.

Transparency does not require publishing a roadmap to the infrastructure.

Not every piece of information should be treated the same.

Our governance model recognizes different levels of data sensitivity.

A system being technically capable of processing information does not automatically mean it is authorized to receive it.

Classification comes before convenience.

Our architecture is designed around a simple principle: information should exist only where it has a reason to exist.

Infrastructure controls are expected to address:

SQUARES AI™ is positioned as the Construction Intelligence Engine operating behind our ecosystem.

It was not developed around the idea of giving every employee unrestricted access to an AI chatbot.

It was developed around the opposite idea: control the intelligence layer.

SQUARES AI™ is intended to bring information together from approved sources, analyze that information within defined workflows and produce intelligence that remains subject to human governance.

The system may support analysis involving:

• Construction drawings

• Specifications

• Addenda

• RFIs

• Finish schedules

• Scope information

• Estimating information

• Project correspondence

• Historical project intelligence

• Commercial documents

• Approved enterprise information sources

Connectivity does not eliminate governance.

Every additional data source expands intelligence — and expands responsibility.

Employees and contractors should not be free to move company or client information into whichever AI application happens to be convenient.

Enterprise AI governance requires a distinction between:

This includes apparently harmless activities such as:

• Summarizing an email

• Reviewing a contract

• Uploading drawings

• Checking pricing

• Rewriting an RFI

• Analyzing specifications

• Summarizing meeting notes

• Uploading customer spreadsheets

A five-minute productivity gain is not sufficient justification for uncontrolled data exposure.

AI should receive the information necessary to perform its authorized function.

Not everything available to the enterprise needs to be exposed to every model, workflow or employee.

Before information is made available to AI, the governing questions are:

• Does the system need this information?

• Is the system authorized to receive it?

• Is the user authorized to expose it?

• Is there a less sensitive way to accomplish the same objective?

More data does not automatically mean better governance.

One of the greatest risks in artificial intelligence is not simply a wrong answer.

It is a wrong answer presented with confidence.

Our Construction Intelligence philosophy therefore places a premium on evidence.

Where practical, material findings should be traceable to their underlying source.

Examples include:

• Drawing sheet

• Detail

• Specification section

• Addendum

• RFI

• Finish schedule

• Contract provision

• Email

• Approved project record

• Other identifiable source material

When evidence is unavailable, uncertainty should not be disguised as certainty.

Inference must remain distinguishable from fact.

AI-generated information may contain:

• Errors

• Missing context

• Misinterpretations

• Conflicting conclusions

• Outdated information

• Unsupported assumptions

• False confidence

• Fabricated information

For that reason, AI-generated output should be evaluated according to the consequence of being wrong.

The greater the commercial, contractual, financial, legal, safety or operational consequence, the greater the expected level of human review.

Across the enterprise, AI should not be used to independently:

• Execute contracts

• Accept contractual obligations

• Approve material commercial commitments

• Submit bids without authorized human review

• Create unauthorized scope commitments

• Represent unverified information as established fact

• Circumvent company security controls

• Circumvent access restrictions

• Upload confidential information into unapproved AI systems

• Conceal the origin of fabricated information

• Impersonate individuals without authorization

• Make discriminatory employment or personnel decisions

• Provide final legal determinations

• Override required human approvals

• Disable required audit or logging controls

• Bypass established document-control processes

Artificial intelligence is not a justification for eliminating established controls.

AI systems operating within our ecosystem are not granted independent authority to bind the company or its clients.

Unless explicitly authorized through a separately controlled process, AI may not independently:

• Commit pricing

• Approve change orders

• Accept substitutions

• Approve contracts

• Accept scope

• Release contractual notices

• Authorize payments

• Execute purchases

• Sign documents

• Waive rights

• Modify contractual obligations

Commercial authority remains human authority.

The same AI Constitution applies across the organization, but implementation depends upon the function of each division.

Being part of the same corporate ecosystem does not mean every division automatically receives unrestricted access to every piece of information.

Access should remain purpose-driven.

Cross-division information sharing should consider:

• Client authorization

• Business necessity

• Data classification

• Contractual limitations

• System permissions

• Conflict concerns

• Confidentiality obligations

• Scope of engagement

Enterprise integration does not mean enterprise-wide exposure.

AI models change.  Providers change.  Capabilities change.  Risk changes.

A model that was appropriate for one workflow may be inappropriate for another.

AI governance therefore includes the model itself.

Models and major AI components should be evaluated for factors such as:

• Intended purpose

• Data handling

• Reliability

• Security

• Model behavior

• Known limitations

• Integration requirements

• Vendor dependency

• Output quality

• Change risk

Material model or architecture changes should be treated as governed system changes rather than invisible software updates.

A powerful AI system should not quietly become a different AI system overnight without anyone understanding what changed.

Material modifications to:

• Models

• Prompts

• Agents

• Data sources

• Permissions

• Integrations

• Automated actions

• Decision logic

• Retrieval systems

should be subject to appropriate testing and change control.

The question is not only: Does the new version work?

It is also: What new risk did the new capability create?

Traditional AI produces information.  Agentic AI may be capable of taking actions.  That changes the risk profile considerably.

As systems gain the ability to:

• Send communications

• Modify records

• Create documents

• Trigger workflows

• Access multiple systems

• Execute transactions

• Make decisions

the control requirements must increase accordingly.

More autonomy should never mean less governance.

Documents, websites, emails and other data sources may contain content designed to manipulate AI behavior.

Therefore, information entering an AI system should not automatically be treated as trusted instructions.

Content is data.

System authority is separate.

An external document should not be able to rewrite the rules governing the system analyzing it.

We reject a common technology assumption: If something can be automated, it should be automated.

No.

Automation should exist only where the benefit justifies the risk and appropriate controls exist.

Some functions should remain intentionally human.

Every consequential AI-supported process should preserve the ability for an authorized human to:

• Challenge the result

• Reject the result

• Correct the result

• Escalate the result

• Disable the workflow

• Require additional evidence

AI systems should support judgment — not trap users inside automated conclusions.

AI incidents are not limited to traditional cybersecurity breaches.

An AI incident may include:

• Unauthorized data disclosure

• Improper access

• Incorrect automated action

• Model malfunction

• Significant hallucination

• Cross-client information exposure

• Compromised integration

• Prompt injection

• Credential compromise

• Unapproved system use

• Unexpected model behavior

Where necessary, the organization must be capable of isolating:

• A user

• A model

• An integration

• A data source

• An automated workflow

• An entire AI function

Every intelligent system needs an off switch.

Trust should not require blind faith.

Material AI operations should be designed so that authorized personnel can investigate what occurred.

Depending upon the system and risk level, this may include:

• User activity

• Source data

• Model or workflow version

• Generated output

• Human changes

• Approvals

• Overrides

• System actions

• Timestamped events

The more consequential the action, the more important the audit trail.

AI governance is not a policy written once and placed in a binder.

Systems should continue to be challenged.

Testing may include:

• Accuracy evaluation

• Hallucination testing

• Permission testing

• Data-boundary testing

• Security testing

• Adversarial testing

• Prompt-injection testing

• Cross-client leakage testing

• Workflow testing

• Human-review testing

• Failure-mode testing

We are interested in what the system does when everything works.

We are equally interested in what happens when something goes wrong.

AI governance cannot stop at systems we build ourselves.

Every external:

• Model

• Platform

• API

• Connector

• Cloud service

• Software provider

• Data processor

may introduce additional risk.

Convenience does not eliminate the need for vendor evaluation.

Technology alone cannot govern artificial intelligence.

People must understand its limitations.

Employees operating AI-supported systems should understand concepts including:

• Hallucination

• Data sensitivity

• Confidentiality

• Prompt injection

• Verification

• Human accountability

• Approved versus unapproved systems

• Escalation

• Appropriate reliance

A user who blindly trusts AI can defeat even a sophisticated technical control environment.

We will not tell customers that any connected technology environment is:

• "Unhackable."

• "Risk free."

• "100% secure."

Those are marketing statements, not serious security principles.

Security is the continuous reduction, management, detection and containment of risk.

The correct question is not: Can risk be completely eliminated?  It cannot.

The correct question is: How intelligently is risk controlled?

It is sustainable AI.

The construction industry does not need another uncontrolled layer of technology producing more information.

It needs intelligence that can be:

• Traced

• Challenged

• Verified

• Governed

• Contained

• Owned by accountable humans

That is the standard we are building toward across the CCFE Group ecosystem.