How to Think About AI — Edition 3

A different way of thinking about replacement

The clearest way to see this is to look at how it has worked elsewhere.

Consider self-driving cars.

When the industry first started taking the problem seriously, almost no one suggested the best route would be to build a humanoid robot that could sit in the driving seat of a conventional car. No one proposed giving it hands to hold the steering wheel and feet to operate the pedals. No one tried to build a machine that drove the way humans drive.

The actual approach looked nothing like that. Engineers built vehicles powered by cameras, lidar, radar, GPS, and machine learning. They do not look like drivers. They do not need to, because the goal was never to recreate the driver. The goal was to deliver the outcome the driver had historically delivered: getting from A to B safely.

In February this year, Uber’s CEO Dara Khosrowshahi went on the The Diary Of A CEO podcast and said something most chief executives will not say in public. He acknowledged that AI will displace the 9.4 million people who drive and deliver for Uber globally. He went further. He estimated that AI will replace the work that 70 to 80 per cent of humans currently do, with knowledge and intellectual jobs going within a decade. He also said something pointed about his peers. Most tech executives, he said, are saying one thing on stage and admitting something very different in private. He named the dissonance because he was tired of pretending it did not exist.

Read that paragraph again. The CEO of one of the largest employers of contractors in the world is forecasting the end of his own workforce. He is telling a public audience that the role he employs millions of people to do will not survive at scale. And the same prediction, in the same interview, was extended to knowledge work. Most professional services. Most office-based roles. Most of what people reading this newsletter do for a living.

His reasoning on driving matters because it points at something most professionals overlook. He is not predicting replacement only because autonomous vehicles are cheaper, although they are.

He is predicting it because they are becoming safer. Self-driving systems do not get tired. They do not get distracted. They do not check their phones. They process thousands of inputs per second across a 360-degree field of view, never blink, and never slow down at the end of a long shift. Human drivers are remarkable in many ways. Continuous, undistracted attention is not one of them.

The assumption embedded in “a machine cannot do what I do” is not only that the machine cannot replicate the process. It is also that humans will always do the work better. The self-driving car example puts pressure on both assumptions. Increasingly, the machine does not need to be as good as a human. It needs to be better. On the metrics that matter to the people commissioning the work, it already is in many situations.

The same logic is reshaping medicine.

Take radiology

Reading a mammogram for early signs of breast cancer has been one of the most demanding tasks in modern medicine. It requires years of training, calibrated judgement, and the ability to spot subtle anomalies on grainy images. Until recently, the safest approach was for two radiologists to read each scan independently, with a third brought in if they disagreed. That is how the NHS Breast Screening Programme has worked for decades.

In February 2026, The Lancet published the results of the MASAI trial in Sweden, a study of more than 100,000 women. AI-supported screening detected 29 percent more cancers than the standard two-radiologist process, without increasing false positives. The AI was not replacing the human reader. It was running alongside as a second reader, drawing on patterns from millions of prior images, and it caught cancers the human eye missed. The same study reported a 44 percent reduction in radiologist workload.

The AI does not look like a radiologist. It does not interview the patient. It does not exercise clinical judgement in the form a doctor would recognise. But for the question the screening exists to answer, namely, whether this woman has early-stage breast cancer, it is now part of producing a better answer than the established human process delivered on its own.

Radiology is one of the most highly trained, highly regulated, highly defended professions in modern healthcare. Radiologists are not being replaced by humanoid robots holding stethoscopes. They are being routed around by a process that does not look like radiology at all, produces better outcomes than they did, and cuts the volume of work that flows through them by nearly half. This is the shape that replacement takes in regulated professional work. Not robots. Software that does the job better.

Applying this to surveying

Now look at surveying through the same lens. The right test is whether the outcomes a surveyor delivers can be reached by other means.

Some of those outcomes are already being reached differently.

Automated valuation models do not value properties the way a surveyor does. They draw on transaction data, planning records, comparable sales, and property characteristics, and produce valuations at scale. They do not need to inspect the property. They do not need to apply professional judgement in the form a surveyor would recognise. But for a substantial and growing share of lender decisions, they produce a valuation good enough to act on. The work is being done.

Structural sensors do not assess buildings the way a building surveyor does. They monitor movement, moisture, temperature, and vibration continuously, and flag anomalies as they emerge. They do not need a site visit. They do not need professional intuition. But for the question of whether a building is moving, they answer it more accurately and more cheaply than periodic human inspection.

None of these systems looks like a surveyor. None of them works like a surveyor. They are doing part of surveying work anyway.

What the assumption costs

My motto is life is “Never Assume” and the same applies here. Holding on to the picture of AI as a humanoid replacement has a real strategic cost.

It produces false comfort. A firm that believes AI has to replicate the surveyor will look at the current state of the technology, conclude it is not close, and move on. The threat they are watching for is not the one arriving.

It directs attention to the wrong places. The same firm will spend its strategic energy debating whether AI tools can write a survey report, when the more useful question is whether the work that created the need for a survey report is being routed around entirely.

It creates a delayed response to a quickly moving problem. By the time a firm realises that the work has been lost, the loss is structural. The clients who used to instruct them are now getting their answers through different routes. The energy that should be going into repositioning the firm around what genuinely requires a surveyor is going instead into proving that AI cannot do something it was never trying to do.

A better picture

A more useful way to imagine the threat is this. Picture every outcome your firm currently delivers. For each one, ask not whether a machine can replicate your process, but whether that outcome could be reached without you.

For some outcomes the answer will be no, even at the limits of foreseeable technology. Those are the outcomes worth building the firm around. They are the load-bearing parts of professional value. They are more likely to be there in 2030.

For others, the honest answer is yes. The outcome can already be reached without you, or will be within a few years. Those are the outcomes you cannot afford to centre its identity on. They are the work that contracts as the market shifts.

It is worth being honest about the limits of this exercise. No one can predict with confidence how far AI will reshape surveying, or any profession. The capabilities arriving next are unlike anything that has come before, so historical patterns can guide thinking but cannot provide certainty. Anyone claiming to know exactly what the profession will look like in five years is either selling something or has not thought hard enough about the question.

That uncertainty cuts both ways. It is a reason to act, not a reason to wait. Firms that wait for clarity will get the clarity they seek, but only after the market has already moved. Firms that move now do so with imperfect information, but they are the ones who will still have agency when the picture sharpens.

The work is not safe because the technology does not look threatening. The work is safe only where the outcome genuinely requires what only a regulated, accountable human can provide. This is the third edition in a six-part series on how to think about AI in surveying. Each Monday, I set out one principle that is challenging the profession, and what to do about it.

Next Monday: why automation is the smallest of the changes coming, and what most firms are missing about the deeper transformation already underway.

Free monthly AI Briefing – one hour informal CPD

My next free briefing is this Wednesday, 17 June, 1pm to 2pm UK time, online. The June edition covers the regulatory landscape now shaping the profession: the EU AI Act, the RICS Standard, the DUAA changes from 19 June, and more.

One hour. Practical, plain English, designed for surveyors. Free. Counts as informal CPD.

Register here → https://www.rics-ai-briefing.surveyors-uk.com

If you cannot attend, join The Surveying Room community to access the recording – join here