LO
Real estate agent
Gets a transaction over the line between two people who do not trust each other — and is paid a share of a price they both have reason to dispute.
This is not a probability of losing your job. It combines how much of the role's task load is exposed to automation with how far adoption has actually gone — useful for comparing occupations on one consistent basis, and for nothing else.
Written for agents handling residential resale and rental — the people paid on commission per transaction. New-build sales teams, commercial brokerage and property management are different jobs with different pressures. The regulatory setting matters more here than on most pages: what an agent is legally required to do differs sharply between markets, and so does how they are paid.
What is actually changing#
The unit of analysis is the task, not the job title. A role is not replaced — its task mix shifts.
Is this your job? Say so and this page narrows to your share of it.
A job title is a bundle of tasks bought together, and no two people hold the same bundle. Nothing is sent anywhere — it stays in this browser.
Knowing what is for sale
Automating≈ Platform inferenceHolding the list: what is available, what just came on, what quietly came off.
This was the profession's original asset and it has been draining away for twenty-five years, long before anything called AI. Public portals put the list in the buyer's pocket; what remains of the advantage is timing and off-market knowledge, both of which shrink as listings become more complete. Machine matching and search extend the same trend rather than starting a new one.
Says nothing about markets where listings are not public or not centralised, which is a large share of the world. It also says nothing about whether losing this asset reduced the number of agents — in several markets agent numbers rose while this advantage fell, which is a fact this judgement does not explain.
Putting a price on it
Being augmented✓ Evidence-backedSaying what it is worth, and being believed enough that a seller lists at it.
Automated valuation has been ordinary for years and is genuinely good at the average house in a liquid market — enough that the agent's estimate is now checked against it rather than taken on faith. What has not transferred is standing behind the number. The clearest evidence on this site is a company that tried exactly that and stopped: Zillow wound down its algorithmic home-buying business, its CEO writing that the unpredictability in forecasting home prices far exceeded what they anticipated, alongside a roughly 25% workforce reduction.
That record is about one company taking balance-sheet risk on its own valuations in one country during an extraordinary housing market; it does not say automated valuation is inaccurate, and it does not say agents price better. What it establishes is narrower and more useful: an estimate and a commitment are different things, and only the second was scarce.
Standing in the room
Being augmented≈ Platform inferenceOpening the door, reading the buyer, noticing what the photographs left out.
Virtual tours, self-showing lockboxes and remote access genuinely removed a share of the trips, and that share is the part of this task that was pure logistics. What is left is the part that was never logistics: noticing the neighbour's extension, the damp smell, the thing the buyer went quiet about. That is observation in a specific room, and no amount of imagery substitutes for it.
Nothing here measures how many viewings still happen in person, and the answer moved sharply during the pandemic and has partly moved back. It also says nothing about rental, where self-showing has gone much further than in resale.
Getting two sides to agree
Still human-led≈ Platform inferenceThe offer, the counter, the thing the survey found, the buyer whose finance fell through on a Friday.
This is the part people actually pay for, and it is not information work. It is holding two anxious parties through the largest transaction of their lives, absorbing bad news on their behalf, and knowing when a deal is dying versus merely stalling. A model can draft the message; it cannot be the person the seller shouts at.
A judgement about the nature of the work, not a measurement of its value. It also does not defend the commission: that this part is human does not establish that a percentage of the sale price is what it is worth, and the fee structure is under pressure in several markets for reasons that have nothing to do with automation.
Finding the next client
Automating≈ Platform inferenceProspecting, farming a neighbourhood, staying the name someone thinks of in three years.
The mechanical half — lists, sequences, targeted advertising, follow-up that never forgets — is now bought rather than done, and it is bought by everybody, which is the point. When the tooling is identical for every agent in a market, it stops being an advantage and becomes a cost of staying in business.
Says nothing about referral, which in most markets is where the majority of business actually comes from and which no tool has moved. It also does not address the oversupply of agents relative to transactions, which is the real pressure on income in several markets and predates any of this.
Answering for what you told them
New task≈ Platform inferenceDisclosure: what you must tell a buyer, what you may not say at all, and what happens when a generated listing describes a property that does not exist.
New work created by the tooling rather than removed by it. Generated listing copy, enhanced photography and automated outreach all produce statements a licensed person is answerable for, and fair-housing style rules make some of those statements actionable. Somebody has to read what the machine wrote before it goes out under their licence.
New work appearing is not new income, and in a commission business unpaid duties are absorbed rather than compensated. Requirements also differ enormously by market, and this site holds no verified record of enforcement against agents for generated material.
Which technologies matter here#
Four separate signals. They are deliberately not added together — a job exposed to two technologies is not twice as exposed.
How it got here#
The index is not a static number. This is where it would have sat at each capability checkpoint since ChatGPT — reconstructed, and labelled as such.
The highest starting point of any occupation added in 2026, and none of it is about AI: public listing portals had already taken this profession's original asset — knowing what is for sale — over the previous twenty-five years, and automated valuation was ordinary well before 2022. The period's rise is the prospecting half becoming something you buy rather than something you do, and identically for every agent in a market, which turns it from an advantage into a cost of staying in business. It flattens because the remaining work is not information: holding two anxious parties through the week a deal nearly dies. Read the height as saying most of the erosion already happened, not that this is about to go — and note what the curve cannot see, which is that agent numbers rose in several markets while this line was climbing.
A flat line is not a forecast of safety. It says which tasks automation has reached so far — the occupations that moved least here are the ones where the constraint is physical or regulatory, and both of those can change.
Recent changes#
One US company taking balance-sheet risk on its own automated valuations, during an unusually volatile housing market and alongside stated capacity and supply-chain constraints. It says nothing about the accuracy of automated valuation as a tool, and nothing about whether agents price better — the company kept publishing estimates after closing the buying business. What it establishes is about who carries the consequence of a valuation, not who produces one.
Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.
What this means for you#
The two things easiest to learn — knowing the listings and working the prospecting tools — are the two this page marks as the least defensible, and everyone entering beside you has the same tools. What is scarce is being the person two frightened strangers will listen to on the worst day of a transaction, and that is learned by being present for deals that went wrong, not by taking a course.
Your position depends on which half of your income comes from information and which from trust. If clients come to you because you know what is on the market, that advantage has been eroding for twenty-five years and tooling is only the latest instalment. If they come because of what you did when a previous deal went wrong, that is the durable half — and it is worth noticing that it is also the half you cannot advertise.
Your options#
Four directions, each with its real constraints and one thing you can test this week. Continuing as you are is a legitimate choice — it just has to be a chosen one.
Move your income to the transaction
The information half is commoditised and the tooling is identical for everyone. What people still pay for is somebody carrying a deal through the week it nearly died.
It is slower to build than a marketing funnel and it does not scale — which is exactly why it holds.
Take your last ten closings and write down, honestly, why each client chose you. If most answers are about information, that is the finding.
Own the part with a legal duty
Disclosure, fair-housing style rules and what a licence makes you answerable for are all getting harder as more of what goes out is machine-written. Being the person who checks is a position, not a chore.
It is unpaid in a commission model until something goes wrong, at which point it is the only thing that mattered.
Take one listing your office published with generated copy and check every factual claim in it against the file.
Go to where the deal is financed or surveyed
Mortgage broking, surveying and conveyancing sit on the same transaction and are bound by duties an automated system cannot hold. Your knowledge of how deals actually fail transfers directly.
Most of these are licensed separately and pay differently — often steadier and lower.
Ask a conveyancer what proportion of their queries come from errors in the listing. The number is usually higher than agents expect.
Common questions#
The wrong frame for this occupation, because the erosion started long before AI and has been running for twenty-five years without removing the job. Public listing portals took the information advantage; the transaction survived. A signal you can check yourself, and a better use of the worry: of your last ten closings, how many clients came to you because you knew something they could not look up. That number is the part under pressure, and it has been under pressure since before anyone reading this started.
Worth separating two things that get said as one. An algorithm can produce an estimate; the question is who stands behind it. The clearest test anyone has run was Zillow buying homes on its own valuations, and it stopped — its CEO writing that the unpredictability in forecasting home prices far exceeded what they anticipated, with about a quarter of the workforce cut. That does not show the estimates were bad. It shows an estimate and a commitment are different things, and the agent was never selling the estimate.
This site cannot tell you, and the honest reason is worth stating: the pressure on agent income in most markets comes from fee structure, regulation and the number of agents relative to transactions — none of which is automation. Confusing the two makes it harder to see either. Watch what your market does to the fee, separately from what the tooling does to the work; they move for different reasons and on different timescales.
They mean fewer trips, which is not the same thing. The part of a viewing that was pure logistics — driving across town to open a door — is the part that moved, and losing it makes the remaining work denser rather than smaller. What did not move is being in a specific room and noticing what the photographs left out, including the things a seller would prefer you did not notice.
Method and sources#
- Assessment date
- 2026-09-12
- Basis of the task judgements
- 1 evidence-backed · 5 platform inference · 0 not enough evidence
- Verified events
- 1