A solo quantity surveyor cut a three-hour monthly review to twenty minutes

I'm a chartered quantity surveyor running a small cost consultancy on my own. In construction, a quantity surveyor is the professional who manages the money and the contract on a building project: setting budgets, running tenders, valuing the work each month, and settling the final account. It's detailed, numbers-heavy work where a mistake has a real cost, so accuracy matters more than speed. Over the last few months I've built a library of small, purpose-made AI skills that take the repetitive parts of that job and do the legwork, while I keep every professional judgement. No big platform, and no automation for its own sake. Just steady, low-risk wins on real fee-earning work. Achievements •    Valuation cross-check skill. Reconciles a contractor's monthly payment application against my own assessment, matching lines that have moved, calculating every quantity and value difference, and drafting the list of changes for me to classify. It has cut the monthly review from two to three hours down to about twenty minutes of checking only the lines that need my eye. Used live on real jobs. •    Tender comparison skill. A specialist comparison of subcontractor quotes that used to take most of a day now takes under an hour. •    A spread of smaller skills across the practice: [a Passivhaus early-design checker, a tax-return helper, CV tailoring, blog formatting and publishing, spreadsheet-to-document conversion, and brand-consistent document drafting. Trim to the ones you want to name]. Tools: Claude (desktop and browser), Claude skills, Python for the arithmetic that runs inside the skills, the Claude for Excel add-in, Microsoft 365 (Excel and Word), and Airtable. I stay tool-agnostic and pick whatever fits the task. Methodology: My habit is simple. When a task keeps eating my time, I try the quickest AI version of it first. If it works and I know I'll do it again, I build it into a reusable skill there and then, rather than letting the method disappear when the chat closes. One rule shapes all of it: anything with a number in it runs as code, not as the model reasoning in prose, so the arithmetic is repeatable and I can show my working. The tool does the legwork; I make and sign off every professional call. As a chartered surveyor I carry the liability for what goes out, so the human checkpoint is designed in, not bolted on. That is also the evidence the RICS responsible-use standard expects me to be able to show. Learnings: The value came from a lot of small, repeatable wins rather than one big automation. Keep the human on the judgement and let the tool take the grind. For anything with a number in it, make it deterministic, because a surveyor cannot hand accuracy to a language model. And capture the win as a skill immediately, or it's gone by tomorrow. Direction of travel: This is early, but it's systematic. I've mapped the repetitive QS tasks across the whole life of a project where AI can safely take the strain, and I'm working through them in order. The areas identified so far: •    Fee stage: bid or no-bid triage from a tender enquiry, fee proposal drafting, and scope clarification schedules. •    Cost planning: benchmark data from past projects, project-specific assumptions and exclusions, a cost plan quality checker, and priced risk schedules. •    Procurement and tendering: drawing register reconciliation, tender completeness checks, query log management, and scope-gap analysis between trade packages. (The tender comparison here is already built.) •    Post-contract: interim valuation preparation, payment and notice date timetabling, a contract clause interrogator, variation assessment, and final account reconciliation. (The valuation cross-check here is already built.) •    Disputes and advisory: chronology building, loss and expense records, and Scott schedules. That's around two dozen workflows, which I'm grouping into roughly eight reusable skills. Two are built and proven, several are in build, and every one keeps a surveyor's checkpoint and runs its sums as code so the numbers can be verified. I'm timing the savings conservatively as I go, rather than quoting best-case figures. I've already found the kind of error these checks are built to catch: a ten-million-pound final account sitting in my own archive whose total reads one penny different in figures than in words, caused by a single missing rounding formula. Real, not planted.