Five things the people building AI admitted

“The AI applications of 2026 and 2027 will be doers.”

That was Jack Clark, co-founder of Anthropic, the company behind Claude, talking to Ezra Klein on his podcast two months ago. Clark runs policy at one of the main companies pushing the frontier of this technology. He is not a critic. He is one of the builders. And throughout the conversation, he talks a lot of sense.

Worth an hour of your time if you have not already heard it.

Five things stood out and each one matters for surveying

Clark was direct about where we are in the technology cycle.

“The AI applications of 2026 and 2027 will be doers.”

Up until now, most surveying firms have used AI as a talker. A chatbot that drafts text. A tool you converse with and then copy and paste the output. The supervision model was simple. You wrote a prompt. It wrote a draft. You checked the draft. You signed it off or you binned it.

That model is finished.

Agents are systems that go away and do things on your behalf, over time, using tools, making decisions, taking actions in sequence. Clark described colleagues running multiple agents simultaneously, with other agents monitoring those agents. He described his own use of Claude Code, where one instruction produced a working piece of software that included subsystems the AI had built for itself.

For a surveying firm, this is not a small shift. The governance question changes entirely. You are no longer supervising a tool that drafts text. You are accountable for a system that takes action. The point of failure moves from the output to the process. The audit trail becomes the artefact.

If your firm is still treating AI oversight as “we read what it writes before it goes out,” you are governing the wrong thing.

Two. Knowledge is now raw power

Clark was describing what is coming next. Systems that can read vast amounts of material and synthesise insights in real time. Things that would have taken a team of researchers weeks, done in minutes.

For surveying, this cuts both ways.

Your firm holds commercially sensitive information. Valuation reasoning. Defects identified and the rationale behind decisions. Client identities, building details, occupier circumstances. Historical reports going back years. Most of that material has, until now, been safe through sheer obscurity. Nobody had the time or the tools to mine it.

That is no longer true.

Anyone with access to your data, including former staff, lapsed system permissions, mis-shared cloud folders, or compromised email accounts, can now ask an AI to read everything and tell them what is interesting. Patterns of valuation. Inconsistencies. Pricing logic. Client lists. The raw material for a complaint, a dispute, a competitor pitch, or worse.

The question for your firm is not whether AI can read your data. The question is who else can now ask AI to read it for them, and what controls you have in place to know if they do.

Three. The value of junior people is becoming dubious

This was the most uncomfortable admission in the interview.

“The value of more senior people with really well calibrated intuitions and taste is going up. The value of more junior people is a bit more dubious.”

He was talking about Anthropic’s own hiring. The basic tasks Claude Code can now do are the same tasks that, traditionally, junior staff did. Reading source material. Drafting initial outputs. Running first-pass analysis. The work that built experience by doing it.

When asked whether graduate unemployment would be higher in three years, Clark said yes.

For surveying, this is the succession conversation most firms have been avoiding. Ninety percent of surveying firms are SMEs. The profession already has an ageing demographic problem. The pipeline of new entrants was thin before AI arrived. Now the entry-level work, the inspection write-ups, the comparable evidence gathering, the initial defect noting, the standard letter drafting, is exactly the work AI is fastest at automating.

If juniors are how seniors are made, and AI is doing the junior work, where does the next generation of senior surveyors come from?

This is not a question for ten years’ time. It is a hiring decision for next year. Firms that automate junior work without redesigning how junior surveyors learn the craft will run out of senior surveyors. The training ladder does not rebuild itself.

Four. The people building it are still working out how to oversee it

Clark was open about how Anthropic now runs its own engineering.

“I would say comfortably the majority of code is being done by the system. Some of our systems, Claude Code, are almost entirely written by Claude.”

The humans at Anthropic are now spending significant time building monitoring tools to track what the AI is doing, because the speed of generation has outstripped traditional code review. Clark put it like this.

“The biggest thing that is happening across the company and on teams that I manage is basically building monitoring systems to monitor this.”

He said it calmly. As a description of the present, not a prediction.

If the company at the frontier of the technology, the company with the most resources and the deepest expertise, is still building the systems needed to oversee its own AI use, every surveying firm using these tools needs to ask the same question of itself.

What does your oversight actually look like?

Not what you tell the regulator. Not what is in the policy document on the shared drive. What does it actually look like, on a Tuesday afternoon, when a surveyor is using AI to draft a report, and the firm needs to know whether the output is reliable, where the source material came from, and who is accountable if it is wrong?

If you cannot describe that process in one sentence, you do not have oversight. You have hope.

Five. The thing the safety community has been warning about is already happening, slowly

Klein asked Clark directly about recursive self-improvement. The scenario, long discussed in AI safety circles, where AI systems start improving themselves faster than humans can review the improvements. The point at which the loop closes.

Clark’s answer was honest in a way that should make every reader pause.

“I think right now it is happening in a very peripheral way. Researchers are being sped up. Different experiments are being run by the AI system. It would be extremely important to know if you are fully closing that loop.”

Translation. We are not yet measuring the thing the safety community has been warning about for a decade. We know it is starting. We do not yet have the instruments to track it properly.

This is not a problem your surveying firm is going to solve. But it is a problem that should reframe how you think about the tools you are buying and deploying. The pace of change inside these systems is faster than the pace of governance around them. The vendor pitching you their AI-powered inspection tool today is building on top of a foundation whose own creators admit they are still working out how to monitor.

The implication is not “do not use AI.” It is “build your governance for a moving target.” The tool you assessed six months ago is not the tool you are using today. The capability you trained your staff on last quarter has likely changed. The risk profile of the system is not static.

Governance that assumes the technology is stable will fail. Governance that assumes the technology is changing under it has a chance.

Where this leaves your firm

Five admissions. From one of the people building the technology. Said openly, on a major podcast, two months ago.

The agents are doing things now. Your data is more exposed than it was. The juniors are being automated before they can become seniors. The oversight gap is real and it exists inside the AI companies themselves. The pace of change is outstripping the ability to measure it.

There is no version of “wait and see” that survives contact with this list.

What firms need is not another tool. It is a governance framework that holds up regardless of which tool is in front of you, which model has been updated overnight, or which staff member has decided to try something new this week. Governance, Use, Accountability, Risk, Documentation. The five things every firm needs to be able to evidence, in order, for any AI system in their business.

That is what GUARD is, and that is the framework I will be discussing on 9th June.

In the meantime, the most useful question you can ask of your firm this week is the one Clark could not answer cleanly about his own company.

What does your AI oversight actually look like, today, in practice, if someone asked you to show them?

If the answer is uncomfortable, you are not alone. Ninety-two percent of surveying firms have no documented process for verifying AI outputs. That is the headline number from the readiness assessment, and it is the anchor for the free 60-minute briefing on Tuesday 20 May at 1pm BST.

Link to register: https://www.rics-ai-briefing.surveyors-uk.com

I look forward to seeing you there.

Nina