88% of Your Job Is Going to AI. The 12% Left Is the Bit You Love

I gave AI 100 typical tasks that a quantity surveyor does on an infrastructure project.

It got through 88, in three hours and only £27.67 of tokens.

Before you update your CV or write off the profession, read the rest. Because the 88% it can do is the admin you have always resented. The 12% it cannot touch is the part you became a QS to do.

This is not a prediction about five years from now. Every rating below reflects what is possible today, with tools that already exist.

How I tested it

I listed 100 typical tasks across the full project lifecycle, from tender to final account. Then I rated each one against a single question: who actually does the work, and who carries the consequence?

Three ratings.

Automate. AI does it end-to-end. You check the output rather than produce it. A genuine removal of manual effort.

Augment. AI does the heavy lifting, the extraction, the drafting, the calculation, the pattern-finding. But you make the commercial call and own the result. Hours become minutes. The responsibility does not move.

Human. The task is about judgement, negotiation, relationships or being physically on site. AI helps at the edges, nothing more.

A word on why I say automate, not replace. You stay responsible for checking the output on every one of these. The machine does the task. The accountability stays with you. Nobody is replaced. The work is.

One more honest point before the numbers. When I say AI completed 88 tasks, I mean it produced credible output on 88. It can be trusted to finish 35 of those on its own. The other 53 still need a QS reading what it produced and deciding what to do with it. That distinction is the whole article, so keep it in mind.

The honest split

Of 100 tasks:

• 35 can be automated today. AI does them, you verify.
• 53 can be augmented. AI does the legwork, you make the call.
• 12 stay human.

Add the first two together and AI touches 88 of 100 tasks. That is where the number comes from. Only 12 are essentially untouched.

Now the full list, stage by stage, so you can see exactly how I got there.

Stage 1: estimating and tendering

The early commercial groundwork. Heavy on data, which is why almost all of it is at least augmentable.

Stage 2: procurement and subcontracts

Comparison and tracking lean automatable. Anything involving a price negotiation or a scope dispute stays with you.

Stage 3: measurement and valuation

The core number-crunching of the role. AI does the maths. You agree the figure with the other side.

Stage 4: change and contract administration

This is where the money leaks, and where AI potentially earns its place fastest. Spotting the change is automatable. Standing behind the notice is not.

Stage 5: payment and cash

Rules-based and reconciliation-heavy. This is the most automatable stage in the whole role.

Stage 6: cost control and reporting

AI assembles the report in seconds. You write the bit that explains why the margin moved.

Stage 7: records and site interface

The capture and review is automatable. Being on site, and talking to the people on it, is not.

Stage 8: risk, final account and dispute

AI builds the substantiation pack. You win the argument across the table.

The tally

• Automate (35 tasks): AI does it end-to-end, you verify.
• Augment (53 tasks): AI does the heavy lifting, you decide and own it.
• Human (12 tasks): judgement, negotiation, presence and accountability.

88 of 100 tasks are materially changed by AI. 35 are genuinely automatable now. Only 12 are essentially untouched.

What "completed" actually looked like

Numbers are easy to argue with, so here is what completed meant in practice on three of the tasks. It also shows exactly where the Automate line gets drawn.

Dayworks (task 29, Automate). Fed a week of allocation sheets, AI pulled every entry of People and Equipment hours, matched them to the contract daywork rates, applied the percentage additions and produced a priced daywork sheet ready to submit. What used to be an evening of cross-referencing took under a minute. The QS checked the rates were the right ones and signed it off.

Compensation event identification (task 35, Automate). Given a month of site diaries, it flagged a run of weather entries that breached the contract weather tolerance, a note about late information that disrupted a follow-on trade, and a plant breakdown affecting the programme. Three potential compensation events a tired reviewer would most likely have filed and forgotten.

Notifying that compensation event (task 36, Augment). It then drafted the notice, dated it and flagged the notification deadline. But it did not send it. Deciding to notify, and standing behind that notice when the Project Manager pushes back, is the QS's call and the QS's risk. That single step is the line between Automate and Augment, drawn inside one task family.

That last example is the whole framework in miniature. AI found the compensation event and wrote the notice. A QS decided to send it.

But what about when it gets it wrong?

This is the rebuttal every QS reaches for, and it is the right one. Yes, AI gets things wrong. It will misread a clause, pull last year's rate, or confidently produce a number that is simply off. Anyone who tells you otherwise is a utopianist.

But look at what that objection actually proves. It is the reason the verify step exists, and it is exactly why only 35 tasks are Automate and 53 are Augment. The whole framework is built on the assumption that AI is fallible, which is why a QS checks the dayworks rates before signing, and why the notice gets drafted but never sent on its own.

A QS who blindly trusts the output is misusing the tool. A QS who uses it to do an evening of cross-referencing in a minute, then spends two of those saved minutes checking it, is doing the job properly. The error rate is not an argument against AI. It is an argument for keeping a surveyor in the loop, which is the entire point, and from 9 March 2026 the RICS standard on responsible use of AI makes that loop a documented professional requirement.

Where the line actually sits

The pattern in the data is consistent, and it is worth saying plainly.

AI automates the work, not the responsibility

Look at the change administration stage. Identifying a compensation event from the records can be automated (task 35). Deciding whether to notify it, and standing behind that notification, is augmentation (task 36). Assessing the Project Manager's assumptions, where you are forming a professional opinion that might be challenged, is human (task 40).

The same task family runs across all three ratings. What moves it up the scale is not difficulty. It is who carries the consequence.

The four tasks AI looks like it can do, but can't

Four tasks look automatable and are not: notifying a compensation event (36), notifying an early warning (37), certifying a subcontractor's payment application (53), and issuing a pay less notice (54).

AI drafts all four perfectly. The words are never the problem. The accountability is. Notifying, certifying and issuing are contractual and statutory acts with consequences attached. When you certify a payment, you own that number. When you issue a pay less notice you had better be right, because it is the one most likely to end up in front of an adjudicator.

On those four, AI is not doing the task while you check. You are doing the task while AI does the typing. That is augmentation, and calling it anything else is how you lose a technically sharp reader in one line.

I could have moved them into the Automate column and shown you a bigger number. The number would have been a lie, and the first commercial manager to read it would have known.

The 12% it can't do is the 12% you became a QS for

Every one of the 12 involves a person on the other side of a table, a junior who needs developing, or a pair of boots on the actual site.

Negotiating the final account. Talking a settlement over the line. Advising the project team when it genuinely matters. Bringing on a junior who reminds you of yourself ten years ago. Agreeing on measurement face to face. None of that is a data problem.

Be honest with yourself for a second. You did not qualify to reconcile ledgers or chase certification. You qualified to make the commercial call, win the argument, and own the outcome. That is the part of the week you would keep if you could bin the rest.

And there is a tier above even these twelve. Commercial strategy. Building the client relationship. Settlement strategy on a major claim. Influencing the stakeholders who decide whether a project lives or dies. Leading and growing a team. Those are not tasks on a hundred-item list at all. They are what the list frees you up to do. Strip out the admin and you do not just protect the human work, you finally get the time for the work that was always the point.

The automatable third is the third you hate

Reconciliation. Ledger postings. Deadline tracking. Record review. Substantiation packs. Cost coding.

This is the work that fills your week and gets blamed when a notice slips. It is also the least satisfying part of the role, and the reason so many QS teams are drowning in admin. Removing it does not threaten you. It hands you back the 53 augment tasks and the 12 human ones, which is where the enjoyment and the career actually live.

What this means in practice

Strip the admin out of the role and three things follow, depending on where you sit.

For contractors, you capture more of what you are owed. The tasks AI is best at, reviewing records and spotting the change, are exactly the ones that decide whether entitlement gets claimed or quietly absorbed. Automate them and the leakage stops. More compensation events get notified in time. More cost gets recovered. The commercial position improves without hiring a single extra person. This is the same leak we sized up in the real cost of a missed compensation event.

For commercial managers, each QS covers more ground. This matters because there are not enough quantity surveyors, and there have not been for years. RICS market surveys have repeatedly flagged the shortage, and every commercial team already feels it from the inside. Roles sit open for months. The good people are stretched across too many projects. When AI absorbs the task list, the commercial work that used to need three surveyors can be run by one experienced QS and the right tools. Not because two people were replaced, but because you were never going to hire them in the first place. The shortage stops biting.

For junior QSs, the admin grind disappears. No more evenings lost to cost coding, reconciliations and dayworks. Which sounds like an unqualified win. It is not quite, and that is the part worth sitting with.

The question nobody wants to ask: what happens to junior QSs?

Here is the uncomfortable bit, and it might be the most important question in the whole piece.

How does a junior QS actually learn the trade?

Traditionally, through exactly the tasks AI is about to take. Measurement. Payment applications. Cost coding. Reconciliations. Dayworks. The repetitive work was never just output. It was the training ground. You learned what a rate really means by building a thousand of them. You learned where money hides by reconciling it line by line.

Automate that away and you solve a productivity problem while quietly creating a development one. The senior QSs are fine. They already did their ten thousand hours. The question is how the next generation gets theirs when the traditional route disappears.

I do not think the answer is keeping juniors on manual work for the sake of it. That is sentimental, and the market will not allow it. The answer is that the learning has to move up the stack. Juniors spend less time producing the records and more time interrogating them. Less time building the notice and more time deciding whether to send it. They reach the judgement work earlier, with AI doing the legwork they used to grind through by hand.

That is a real shift in how the profession trains, and the firms that work it out first will have a serious edge. The ones that simply cut juniors because AI does that now will wake up in ten years with no seniors. We went further into this in whether AI kills off the graduate QS.

Where Gather's QS AI Agent fits

Fair question, given who built this. Here is the honest answer, mapped straight onto the list above.

The QS AI Agent does not try to do all 35 automatable tasks, and it does not pretend to touch the 12 human ones. It owns one spine: the records to change pipeline.

What it does for you

It works in the background on nine of the 35 Automate tasks, the ones most expensive to get wrong:

• Reviews every site diary and shift record as it is logged
• Cross-references each entry against the contract
• Surfaces the compensation events hiding in routine records
• Keeps the Early Warning Register current
• Tracks notification deadlines so none expire
• Records weather and its programme impact
• Maintains the record of delay and disruption
• Compiles the contemporaneous evidence as it goes
• Builds the final account substantiation pack

Where it stops

It drafts the paperwork, then hands the decision straight back to you. It drafts the compensation event notice, dates it and flags the deadline. It does not send it. Same with the early warning, the quotation and the contractual letter.

AI does the typing. The QS makes the call and carries the accountability. That is the line between Automate and Augment, and Gather is built to respect it.

What it never touches

The 12 human tasks stay entirely yours: negotiating the final account, agreeing measurement across the table, advising the project team when it counts, and bringing on your junior.

Nor should it touch them. That is the part you became a QS for.

Why this spine and not another

Because it is where the money leaks. Reviewing diaries and spotting the change usually falls to whoever has time, usually a junior, usually at month-end, usually too late.

Gather's own analysis of beta projects suggests manual diary review captures only around 60% of legitimate change events. The missing 40% is revenue the Contractor absorbs, because by the time anyone notices, the notice period has gone and the negotiating position with it.

That is the gap the QS AI Agent closes. Not by replacing the surveyor, but by doing the nine things that should never have been left to a tired junior at half past five on a Friday.

The bet worth making

Roughly a third of a quantity surveyor's task list can be automated by AI today. Another half can be heavily augmented. About an eighth stays firmly human.

And that third is a floor, not a ceiling. This is the worst AI will ever be at your job. The tools only sharpen from here, and the line keeps moving. Several of today's augment tasks, drafting a notice, certifying an application, will become safe to trust outright, and the automatable share will climb past a third towards half and beyond.

What does not move is the shape of the 12. People, judgement and presence are not solved by a better model. So the safest bet a QS can make is not that AI will stall. It is the opposite. The admin keeps falling away, and the human core keeps mattering more.

The headline that survives scrutiny is not that AI replaces quantity surveyors. It is the opposite.

There are not enough of us, there have not been for years, and AI strips out the manual effort so the surveyors we do have can cover more ground and spend their time on the judgement, the negotiation and the relationships that were always the point of the job.

The QSs who win this are not the ones who fight it. They are the ones who let it take the 88% they never enjoyed, and get genuinely brilliant at the 12% they became a QS to do.

Sources

• RICS UK Construction Market Surveys, which have consistently reported quantity surveyor and labour shortages
• Gather QS AI Agent capability analysis, beta project data. The 60% capture figure reflects Gather's own project analysis, not an independent industry study