What Exactly Can You Do With an AI QS?

In summary

• The AI QS is available right now. It is 40 QS skills, 8 agents and 4 always-on hooks running on your actual site records, and it works in Claude, ChatGPT and Microsoft Copilot.
• The ROI is substantial: the delay register that takes 2 days takes 2 minutes; the valuation that takes a day takes minutes; the half-day application check takes 1 minute, line by line with evidence.
• Hooks are the sleeper feature. They never stop reading your diary, scanning every record for compensation events, fatigue risks and data gaps, and tap you on the shoulder before the clock runs out.
• Routines turn the commercial calendar into scheduled output. The Monday pack builds itself; you review it.
• What separates it from a chatbot is the methodology: NEC, JCT and FIDIC clause knowledge, fabrication guards, and every line traceable to a shift record.
• Honest limitations included, as always. It does not replace judgement, and it is only as good as your records.

Monday, 8:47am

Picture your next Monday morning.

Instead of three hours pulling last week together, there is a one-page commercial pack waiting for you. Valuation movement. Every subcontractor's position. The idle plant you are still paying for, ranked by what it is costing you per week. Material flags. The prelims position against the tender allowance. Actions ranked by money.

You did not ask for it. It ran itself. And underneath it, quietly, something has been reading every diary entry that came in over the weekend, checking for the things that become compensation events, and it has flagged two.

That is not a 2030 vision. That is what an AI QS does today, on live site records, on the kind of NEC and JCT jobs most commercial teams work on. Every week someone asks me: I get that AI is coming for commercial work, but what would an AI QS actually do? So this is the concrete answer.

The example I will use is the one I know best: the Gather plugin. It is a fair test case because it is not a chatbot with a construction skin. It is a combination of 40 individual QS skills, 8 multi-skill agents and 4 automatic hooks, each one encoding the analytical steps a QS would follow, connected to your project records through a secure, read-only connection. You ask a question in plain English. It routes to the right skill, pulls the records, and applies the methodology.

What is it, exactly?

Before the list, a plain definition.

The AI QS is a plugin that sits on top of your Gather site records and does quantity surveying with them. Under the bonnet, it is three things working together.

Skills are focused analyses. Each one encodes the steps a QS would follow for a single task, a valuation, a delay register, a daywork sheet, and automates them against your records. There are 40 of them. You invoke one with a command or just by asking the question in plain English.

Agents are multi-skill workflows. When the job needs several analyses assembled together, a full claim package, a portfolio review, a Monday-morning commercial pack, an agent orchestrates the right skills automatically and compiles the results. There are 8.

Hooks run without being asked. They watch your records as you work and surface things you should know: fatigue risks, data quality gaps, potential contractual events. More on these below, because they might be the most underrated part of the whole thing.

All of it connects to your project data through a read-only, OAuth-secured connection. Nothing the AI does can modify your records. And every output carries a reference number with each line traceable back to the shift record it came from. It runs on the same open standard we covered in the piece on MCP for quantity surveyors.

The record-heavy work, timed

The honest pitch is not that the AI QS does anything a good QS cannot. It is that it does the record-heavy work at a speed no human can, and shows its working. Same task, same records, same evidence standard. Read the middle column: that is where the week goes.

Added up over a month, that is the better part of a working week handed back to judgement, not admin. And it goes to the only things that were ever really the job: the decisions, the negotiations, the judgement calls that no amount of record-reading was ever going to make for you.

Here is what that looks like in practice.

The commercial routine, automated

Start with the work that eats the week.

The monthly valuation. The measure for an interim application for payment is a day of arithmetic on most jobs. The valuation skill builds it from progress records in minutes, quantities done this period against your rates, with every line traceable back to the underlying shift record. Traceability is the point. If the PM queries a line, you can show the record it came from. That is the same discipline behind building an amount due the PM can't cut.

Subcontractor application checks. This is the one that has landed hardest with commercial managers I have shown it to. A subcontractor submits an application; the check runs line by line against recorded attendance, hours and output from the site diary. What used to be half a day of cross-referencing timesheets becomes a one-minute verification with evidence attached to every line. Not trust me, but here are the shift records that do and do not support this claim.

The weekly commercial report. Three to four hours of pulling labour, plant, materials and progress from different sources becomes a single command. There is even a Monday-morning agent, the Weekly QS pack, that runs the valuation movement, subcontractor position, plant off-hire review, material reconciliation and prelims monitor in one pass.

Dayworks. Signature-ready daywork sheets built from contemporaneous shift records, labour, plant and materials already captured on the day. No reconstructing a Tuesday three weeks later from memory and a photo of a whiteboard.

Cost control. Cost to complete forecasts based on actual recorded productivity rather than the tender allowance you stopped believing in months ago. A prelims monitor tracking staff headcount on site against the tender allowance week by week. A plant off-hire review that ranks idle hired plant by weekly cost exposure and tells you what to send back. A material reconciliation that compares delivered against placed quantities and exposes waste before it becomes a month-end surprise.

The claims work, evidenced

This is where structured records earn their keep.

Delay registers. Building a delay register for a claim normally means one to two days reading diary entries. The delay analysis skill identifies, categorises and quantifies delays from shift narratives in a couple of minutes, and outputs a numbered register you can reference in submissions.

Narrative search. Find every mention of access problems across the whole job used to be half a day of scrolling. Now it is a ten-second query across every shift narrative on the project. For anyone who has assembled claim evidence the old way, this alone justifies the exercise.

As-built programmes. Reconstructing actual start, finish, active and dormant periods per activity from records, the raw material of an EOT submission, drops from days to about two minutes.

Compensation events, end to end. An events skill scans records for potential contractual events. A CE builder assembles a structured narrative from the diary evidence. A CE quotation skill prices it as a clause 62 quotation under NEC, changes to the Prices and delay to Completion. And an NEC deadlines skill runs the deemed-acceptance clock, computing every live clause 61.3 time bar and 61.4 reply deadline so nothing quietly times out. If you have ever had a near miss on the 8-week time bar, you know why that matters.

Disruption analysis. A measured mile skill compares baseline against impacted period output using the recognised methodology, not a hand-waved percentage.

40 QS skills, six groups

Each skill is one focused analysis, run with a single command or a plain-English question.

8 agents: skills chained into workflows

Use a skill for one focused output. Use an agent when the job needs several analyses assembled together.

Contract knowledge and guardrails

Here is what separates this from pasting your diary into a generic chatbot.

The first is contract framing. The plugin carries NEC3/NEC4, JCT and FIDIC knowledge as auto-invoked skills with clause references. Ask about a delay on an NEC job and you get compensation event language and the eight-week notification window. Same question on a JCT job and it becomes Relevant Events and Relevant Matters. On FIDIC, the 28-day claim notice. It also supports a contract profile, so if your contract has amended timescales, and whose does not, the deadlines run off the contract as executed, not the standard form.

The second is guardrails. Every skill follows shared patterns: never guess at a project name, explain when data is missing rather than inventing it, and gate outputs on evidence. There are fabrication guards precisely because the failure mode everyone fears, an AI confidently citing a shift record that does not exist, is the one thing a commercial tool cannot be allowed to do.

The hooks: a QS that never stops reading the diary

This is the part I want to stress, because it is easy to skim past and it changes the nature of the tool.

Skills and agents answer questions you ask. Hooks work the other way round. They run automatically, constantly scanning your records for things that need action, whether or not anyone thought to look.

Think about what that means in practice. Every QS knows the failure mode: the compensation event that was sitting in the diary for six weeks before anyone noticed, quietly running down the clause 61.3 clock. The access restriction mentioned in a Tuesday narrative that nobody connected to the programme slippage until the claim was already compromised. The events were recorded. Nobody was reading for them. That is exactly what a missed compensation event really costs.

The event detection hook reads for them. As records come in, it scans for potential contractual events, the things that under NEC become compensation events and under JCT become Relevant Events, and surfaces them to you unprompted. It is the difference between a filing cabinet and a colleague who reads everything and taps you on the shoulder.

The same applies to the other hooks: fatigue alerts flag at-risk individuals from workforce scores before it becomes an incident, and data quality checks flag thin records before they become a hole in your evidence six months from now. None of it requires you to remember to ask.

Think about the asymmetry here. One missed 61.3 notification can cost more than a year of software ever will. The hook that catches it does not get tired on a Friday afternoon, does not go on leave during the busiest month, and does not skim the narratives because the valuation is due. For a profession whose biggest commercial losses come from things noticed too late, an always-on reader of your own records is not a nice-to-have. It might be the single strongest argument for the whole thing.

Routines: the commercial calendar, scheduled

The other shift in thinking is that once QS tasks are commands, they can be routines.

The commercial calendar is rhythmic. The valuation is monthly. The commercial report is weekly. The subcontractor applications land on a cycle. The prelims position, the idle plant review, the material reconciliation, all of it repeats. Which means all of it can be scheduled rather than remembered.

The Weekly QS agent is built for exactly this: a Monday-morning pack covering valuation movement, subcontractor position, idle-plant cost exposure, material flags and the prelims position, one page, actions ranked by money. Run it as your Monday routine and the week starts with the commercial position in front of you instead of the first three hours spent assembling it. Pair it with the Data Assurance agent as a preflight before anything high-stakes, and the assessment-date workflow before each valuation, and you have a commercial operating rhythm where the assembly work happens on schedule and your time goes on the decisions.

That is the mental shift worth making. Not that AI answers my questions faster, but that the routine commercial workload runs itself and you review the output. The QSs who make that shift first will run more projects, catch more entitlement, and spend their days on work that actually compounds. The ones who do not will still be assembling Monday's report at Wednesday lunchtime.

Your AI QS works wherever you do

One connection to your site records. The same AI QS in Claude, ChatGPT and Microsoft Copilot.

One Gather connection. 40 QS skills. Your records, your rates, your contract. Read-only, workspace-scoped, permission-aware, OAuth 2.1 secured. AI access can never modify your data.

Limitations

As ever, the caveats matter as much as the capabilities.

It is only as good as your records. Every output traces back to site diary data. If your narratives are thin or your progress records patchy, the analysis will be too. There is a data quality skill for exactly this reason, and running it first is not optional on a real job. Garbage in still applies; it just applies faster now.

It does not exercise judgement. It is a large language model after all. It will build the CE narrative and price the quotation, but whether to notify, how to play the negotiation, and what the commercial relationship can bear remain your call. It is a very fast assistant QS with a perfect memory, not a commercial manager.

Contract outputs still need checking. Clause references and time bars are encoded, but amended contracts are exactly where standard-form assumptions break. Set up the contract profile properly and review anything that leaves the building. From 9 March 2026 the RICS standard on responsible use of AI makes that review a documented professional requirement.

What to do with this

If you want to see what an AI QS does rather than read about it, the path is short. It works with Claude, ChatGPT and Microsoft Copilot; all you need is a Gather account with project access and the assistant your team already uses.

Then try the question you actually care about. Build a delay register for January. Verify this subcontractor application. Are we on programme? The answer will come from your records, with the working shown.

And if your records are not yet structured enough to support any of this, that is the real starting point, and honestly the more important one.

The AI QS is downstream of the site diary. Every capability above, the two-minute delay register, the evidenced application, the hook that catches the CE before the clock runs out, exists because the records were captured properly at the point of work. Get the records right and everything above becomes available to you. Leave them in a paper diary, and none of it does.

The gap between commercial teams that work this way and teams that do not is already opening. It will not close on its own.