CEDARS - A framework for using AI in Construction Quality

From the PQS Team.

CEDAR - A framework for using AI in Construction Quality was our first attempt to grapple with the idea of AI in Construction Quality. In it we laid out five principles on how to use AI in our industry.

In the only four months since that article, we have found an urgent need to update this framework because of a new understanding of how AI Agents actually work. The original blog post focused on Claude Desktop as the example of an AI agent. Since then we have been using AI coding agents to audit Construction Quality records with agents like Claude Code, Codex, and even Github Copilot CLI.

Claude Desktop, specifically its Cowork feature, brings Claude Code's agentic capabilities into a graphical interface, while Claude Code, Codex, and Copilot CLI expose similar agentic workflows more directly. The important distinction here isn't branding: it's that these tools can now read files, run programs, and potentially modify the material they are reviewing silently or with barely any trace of the edit.

Let's restate the CEDAR principles with the new "Strict Open Loop" principle added.

The CEDARS Framework for using AI in Construction

Letter	Principle	Definition
C	Conservation	Preserve the original source documentation, records, images, scripts, and other evidence behind AI-assisted work. AI outputs must remain traceable back to the materials and methods that produced them.
E	Energy	Use AI inference only where inference is appropriate. Deterministic tools such as scripts, calculations, MCP servers, AI-skill scripts, and databases when they can perform the task more accurately, efficiently, and repeatably.
D	Diligence	Audit AI-assisted work to a level appropriate for the importance and risk of the task. Humans remain responsible for ensuring that information provided to others can be reasonably supported.
A	Accuracy	Use the tool that produces the most accurate result. AI should be used when it improves accuracy, speed, or reliability. Use AI for scan transcription, but scripts for structured, vector-graphic documents.
R	Responsibility	AI must not approve or reject installed work or inspection results in the name of a human. It may identify issues and raise alarms, but under CEDARS the responsible human retains authority for acceptance, rejection, and other accountable quality decisions.
S	Strict Open Loop	AI may detect errors, flag discrepancies, recommend corrections, and prepare proposed changes, but it must not autonomously modify authoritative jobsite records. A responsible human must explicitly authorize any correction, and the original value, correction, authorization, and reason for the change must remain traceable.

We use “Open Loop” here in a specific sense: AI may open a correction workflow, but it cannot close that workflow by writing back to the authoritative record. Only an accountable human can close the loop.

Preserve the boundary between evidence and inference.

Strict Open Loop does not mean that AI cannot create or modify data. AI should be free to create summaries, extract data, perform calculations, build indexes, populate discrepancy lists, propose corrections, and create other derived information. In fact, AI agents seem to do these tasks very well and at very low cost. The protected boundary here is the authoritative record. Signed test and inspection reports, NDE reports, field-entered weld-log entries (even if it is digital), inspection photos, or an instrument's original outputs are evidence of what was observed and recorded. Electronic forms, including Custom Forms in QC Database, should be treated as human records when they represent observations or entries made by a human inspector. AI may analyze that evidence, but it must not silently rewrite it. Any form populated by deterministic metrology or AI systems must be clearly labeled as such.

This simple workflow explains the goal of CEDARS:

Source Evidence → AI Analysis / Derived Information → Human Decision.

AI can do enormous amounts of work in the middle of that workflow. It CANNOT be allowed to silently rewrite the source evidence or impersonate the responsible party who created the evidence, made observations in the field, or approved the installed work.

If you think this limit is too restrictive, consider a hypothetical: An intern finds a site photo showing a A105 stamp on a flange. He knows that A105N flanges are required on that system. Would it be appropriate for the intern to open photoshop and add the "N" to that photo?

Absolutely Not!

The same idea goes for AI. If an intern cannot photoshop your site photos to make them look conformant, neither should an AI be allowed to modify your heat/MTR log to conformatize it.

Important Note: Strict Open Loop is more restrictive than simply saying a system is “human in the loop.” In the AI coding world, a human skimming code before pushing to github is considered 'human in the loop'. In reality, AI coding agents do all the coding work and humans just act as a sleepy gatekeeper. Strict Open Loop means that AI NEVER writes to an authoritative record that is presented as human-attributable or metrology-attributable.

Why Strict?

Your IT department may have a meltdown if you tell them you want a coding agent auditing project records. They have legitimate reasons: these agents may be able to execute programs, install dependencies, access the network, and read or modify files. But CEDARS addresses another category of risk that is specific to construction quality…

1) Coding Agents cannot stand typos

Coding agents are trained to regard inconsistencies as things to fix. In their native coding environment, fixing typos is exactly what we want. This helps avoid annoying debugging sessions related to spelling the word "from" as "form" (I can't type). That is NOT something I want to happen in my weld log. If the AI (Authorized Inspector from Hartford Steam Boiler) finds out my weld log was written or modified by AI (Artificial Intelligence), that inspector will ask if the coding agent inspected the welds in the boiler. Awkward.

It is far easier to explain mistakes caused by human fat-fingering than it is to explain AI pencil-whipping. AI weld records would be more consistent and less trustworthy at the same time.

2) Coding Agents are even more eager to please than other AI tools

AI Coding Agents will just say "I fixed that bug, and some other ones I found." You could say they have a "can-do" attitude. How many times have you been dinged by an auditor because a weld was x-rayed 3 days before it was welded? A general-purpose agent tasked with “cleaning up” that log may reasonably understand its job as reconciling those contradiction. No dishonesty is required. It is doing exactly what software agents are normally rewarded for doing: resolving an inconsistency and completing the task. That is precisely why the control cannot depend only on prompting the AI to behave. The system should technically prevent the agent from closing that loop on its own. That arrangement preserves the value of the NDE log and your other quality records.

Imagine the following questions:

Was the NDE log changed? - Yes.

What was changed? - Weld W-104 NDE date, July 12 → July 15.

Who authorized the correction? - John Human, July 15, 2026 at 8:15 AM.

What proposed the correction? - Claude Code.

What evidence supported it? - Weld completion record X and NDE report Y.

What was the original value? - Preserved in the audit trail.

This is better than:

Was the NDE log changed? - It looks perfect, so we assume AI cleaned it up.

Who changed it? - Probably Claude.

Who told it to do that? - We have no idea.

3) Coding Agents own the code base, not the Construction Quality Database.

If you have ever tried vibe-coding, you will know that it involves telling your AI assistant build an app, finding that it works pretty well, and then not really caring how it got there. You can let the IT and security guys figure out how it works later when they put it in production. Who built the app? You? Claude? If you look at github, you will probably find both of your names there even if you didn't write any code at all. This is fairly sloppy attribution, but it is the standard today.

Construction quality records, however, MUST remain attributable to the person, instrument, or process that actually produced the observation. An inspector can ask an AI assistant "What welds did I inspect yesterday?". The key word there is "I". The inspector is the entity that did the work and the AI assistant just helps with paperwork. The AI can never be trusted to modify the inspection log without a Strict Open Loop. If the AI finds an issue with the inspection log, the inspector must approve every change verbatim or simply make the AI-recommended fixes manually if appropriate.

4) Coding Agents are not just the model

When we say “AI,” we tend to focus on the model. An AI agent is a larger system. The model plots the course, but the surrounding software determines what files it can access, what tools it can call, what commands it can execute, what requires approval, and what changes are automatically accepted. This means that no matter how smart AI models become, there is still risk of error and overzealous fixing in agentic systems. CEDARS therefore has to govern the whole agentic system, not merely the model.

AI-generated records

Many manufacturing processes use machine vision, automated metrology, sensors, and other systems in which the machine itself is intentionally the source of the inspection result. CEDARS is not intended to prohibit those machine-originated records.

The dividing line is not whether a record is digital, whether AI touched it, or even whether a human eventually approves it. The question is what the record claims happened. If the record says a human inspected something, preserve that human observation as human evidence and let AI assist around it. If the inspection was actually performed by an automated system, the system may create its own record as long as that provenance is clear. Once either a human-originated or machine-originated record becomes authoritative, Strict Open Loop protects it from silent revision by AI.

In Conclusion

AI is going to become extraordinarily good at finding mistakes in construction records. We should use it. It should search every document, cross-check every table, question every impossible date, flag every missing report, and point out every discrepancy a tired human auditor might miss, but it should never quietly make the evidence agree with itself. CEDARS does not reserve clerical work for humans. Quite the opposite. Instead, where a quality record represents a human observation, inspection, or decision, CEDARS preserves the distinction between the person who created that evidence and the AI that assists them.

• DO allow automated systems to document their own observations.

• DO record the provenance of evidence, including who or what made the observation.

• DO allow AI systems to audit human evidence and create documentation of its insights.

• DO assume that AI systems, regardless of how intelligent the model is, can work from incorrect assumptions.

• DO NOT allow AI systems to impersonate a human observer.

• DO NOT allow AI systems to obscure or misrepresent who or what created an observation.

• DO NOT rely on AI alone to decide whether a discrepancy is a harmless typo or evidence of a real inconsistency.

Construction quality is an evidence-heavy discipline. CEDARS is not about reserving inspection work for humans or preventing automation. Humans may make observations. Automated systems may make observations. AI may analyze either. What must remain unambiguous is who/what actually made the observation, what the original record said, and who authorized any later correction.

That is the line CEDARS is intended to preserve.