Three AI Agents, One Quantity Surveying Practice, 58% More Cost Plans Delivered

<p class="lead">Caldwell &amp; Webb is a five-person quantity surveying practice based in Birmingham. They work across residential development, commercial fit-out, and housing association projects — the unglamorous, structurally essential work that keeps UK construction moving. In February 2026, the practice's director Tom Caldwell ran a time audit that produced an uncomfortable number: 27 hours a week were disappearing into cost plan drafting, tender documentation, and fee proposals. The kind of work that looks like professional practice but follows the same repeatable pattern every single time. Three AI agents changed the maths. By the end of May 2026, the practice was delivering 58% more cost plans per month, had recovered 22 hours of weekly capacity, and was running the entire system for £80 a month.</p> <figure> <img src="https://images.unsplash.com/photo-1486325212027-8081e485255e?w=1200&q=80" alt="AI operating system for a UK quantity surveying practice — three AI agents delivering cost plans, tender documents, and fee proposals for a Birmingham QS firm" width="1200" height="630" loading="lazy" /> </figure> <h2>The 27-Hour-a-Week Problem Written in Yellow Highlighter</h2> <figure> <img src="https://images.unsplash.com/photo-1568992688065-536aad8a12f6?w=1200&q=80" alt="Quantity surveyor buried in NRM cost plan documents and spreadsheets — the repeatable admin work that AI agents are designed to handle" width="1200" height="800" loading="lazy" /> </figure> <p>Tom did not start the time audit expecting to find a 27-hour problem. He started it because the practice was turning down work. New enquiries were being politely declined or deferred, not because the team lacked the skill, but because they lacked the hours. A five-person QS practice with a senior director, two qualified surveyors, and two assistants should not be running at capacity with the same project volume as three years ago.</p> <p>When the hours were mapped, three categories dominated:</p> <ul> <li><strong>Preliminary cost plans.</strong> Every new commission started with a preliminary cost plan — taking architects' drawings and specifications, applying NRM elemental cost rates, building the summary sheet, and producing the client-facing document. An average preliminary cost plan took 3.5 hours. With six to eight new commissions a month, this was consuming 21 to 28 hours of monthly capacity before any other work was touched.</li> <li><strong>Tender documentation.</strong> Once a cost plan was approved and the project moved to procurement, the practice assembled the tender documentation package: bills of quantities, invitation-to-tender letters, contractor briefing notes, and evaluation criteria. A typical package took 2.5 to 3 hours to assemble, following a structure that changed very little from project to project.</li> <li><strong>Fee proposals.</strong> Every new enquiry required a bespoke fee proposal — scoping the service, referencing comparable projects, building the fee schedule, and formatting the client document. Each proposal took 90 minutes to two hours. With eight to ten new enquiries a month, this alone represented 12 to 20 hours of monthly capacity.</li> </ul> <p>Total: 27 hours a week on work that followed predictable patterns. The judgment, the expertise, the client relationships — none of that was the bottleneck. The bottleneck was the assembly work wrapped around it.</p> <blockquote><p>The irony is that the work we were losing most time to was the work we're best qualified to do quickly. A qualified QS can apply NRM rates faster than anyone — but the assembly, formatting, checking, and document production still takes 3.5 hours. That's the part we needed to change.</p></blockquote> <h2>Three Agents, One AI Operating System</h2> <p>The design principle was the same one we apply to every AI operating system: identify where the pattern is consistent enough for an agent to run it, identify where human judgment is genuinely required, and build the handoff between the two. For Caldwell &amp; Webb, the pattern work was the document assembly. The judgment work was the rate selection, value engineering advice, and client communication. The agents handle the first; the surveyors own the second.</p> <h3>Agent 1: The Cost Plan Agent</h3> <figure> <img src="https://images.unsplash.com/photo-1504307651254-35680f356dfd?w=1200&q=80" alt="AI cost plan agent for quantity surveyors — automated NRM preliminary cost plan generation from architect drawings and specifications, cutting 3.5 hours to 40 minutes" width="1200" height="800" loading="lazy" /> </figure> <p>The Cost Plan Agent triggers when a new project folder is created in the practice's shared drive. The surveyor uploads the architect's drawings, specification document, and a brief project data sheet. The agent reads the inputs — project type, gross internal area, specification level, and location adjustment — and applies NRM elemental cost rates from a knowledge base the practice maintains and updates quarterly.</p> <p>Within twelve minutes, it produces a draft preliminary cost plan: the elemental breakdown, the summary sheet, a contingency and professional fees section, and an executive narrative explaining key assumptions and risk factors. The draft is routed to the responsible surveyor in Slack for review. They check the rate selections, adjust any elements where project-specific knowledge changes the number, and approve or edit the document before it goes to the client.</p> <p>What previously took 3.5 hours now takes 40 minutes — twelve minutes of agent runtime and 28 minutes of surveyor review and rate adjustment. The professional judgment is still entirely in the hands of the qualified surveyor. The assembly, formatting, and first-draft production is handled by the agent.</p> <h3>Agent 2: The Tender Document Agent</h3> <p>Once a cost plan is approved and the project moves to tender stage, the Tender Document Agent assembles the full procurement package. It pulls the approved cost plan, the client's contractor selection criteria, the practice's standard invitation-to-tender template from Notion, and any project-specific clauses flagged by the surveyor. It produces the bill of quantities draft, the invitation-to-tender letter, the contractor briefing note, and the evaluation matrix.</p> <p>The package is assembled in under fifteen minutes and routed for a 20-minute surveyor review before going out. Previous assembly time: 2.5 to 3 hours. Current total time: 35 minutes including review. The agent applies the practice's standard clauses consistently — something that, under manual production, occasionally varied between team members depending on which template they opened last.</p> <h3>Agent 3: The Fee Proposal Agent</h3> <p>New enquiries come in through the practice's website contact form and are automatically processed by the Fee Proposal Agent. It reads the project brief, identifies the service type (full QS service, employer's agent, cost management only, or project monitoring), pulls comparable projects from the practice's Airtable database, and drafts a fee proposal: scope of service, fee schedule, relevant experience references, and next steps.</p> <p>The draft goes to the director for a ten-minute review and personalisation before sending. The whole cycle from enquiry to sent proposal: under two hours, compared to the previous two-business-day average. Speed matters here beyond efficiency — faster proposals win more work. The practice's proposal-to-commission conversion rate improved from 38% to 49% within the first 90 days, partly from speed, partly from the improved structure and consistency the agent applied to every proposal.</p> <h2>The Build: What We Actually Put Together</h2> <p>The technical stack follows the same architecture we use across all QF builds. n8n handles the orchestration — trigger monitoring, agent routing, Slack notifications, and Airtable updates. Claude handles the drafting and reasoning work: Opus for cost plan narratives and fee proposals where quality and nuance matter, Haiku for document assembly and formatting tasks where speed and cost matter more. The practice's knowledge base — NRM rate tables, project database, proposal templates, standard clauses — lives in Notion and Airtable, connected to the agents via API.</p> <p>The build took seven weeks: two weeks to design the architecture and map the exact inputs and outputs for each agent, five weeks of parallel testing with live projects before the team stopped producing any of this work manually. The <a href="/blog/human-in-the-loop-ai-agents-uk">human-in-the-loop design</a> was non-negotiable — all three agents route their outputs to the responsible surveyor before anything reaches a client. Professional liability means this judgment step stays human.</p> <p>Running costs:</p> <ul> <li>n8n Cloud: £55/month</li> <li>Anthropic API (Claude): £20/month at current volume</li> <li>Airtable Pro: £5/month (already in use before the build)</li> <li>Total: £80/month</li> </ul> <p>The payback period was under five weeks. At the practice's blended rate, recovering 22 hours of weekly capacity in the first full month of operation generated over £4,000 of productive time that could be directed to billable work.</p> <h2>The Results: 90 Days In</h2> <figure> <img src="https://images.unsplash.com/photo-1590283603385-17ffb3a7f29f?w=1200&q=80" alt="AI operating system results for a UK quantity surveying practice — 58% more cost plans delivered, 22 hours recovered weekly, proposal conversion up from 38% to 49%, £80 per month running cost" width="1200" height="800" loading="lazy" /> </figure> <p>By the end of May 2026, the numbers were clear:</p> <ul> <li><strong>Cost plans delivered per month:</strong> Up from an average of 6.5 to 10.3 — a 58% increase with no additional headcount.</li> <li><strong>Weekly admin hours:</strong> Down from 27 to 5. The five remaining hours are genuinely professional work: rate judgment, value engineering advice, client meetings, and quality assurance.</li> <li><strong>Proposal turnaround:</strong> From an average of two business days to under two hours.</li> <li><strong>Proposal conversion rate:</strong> Up from 38% to 49%.</li> <li><strong>Running cost:</strong> £80/month for the full three-agent system.</li> </ul> <p>The downstream effect was what changed the direction of the practice. With 22 hours of weekly capacity recovered, the team took on three housing association framework agreements that had previously been out of reach due to volume constraints. Within 90 days, the AI operating system had paid for itself many times over in the new work it made possible.</p> <blockquote><p>We're not producing AI-generated cost plans. We're producing surveyor-approved cost plans that the AI drafted. The distinction matters — both professionally and practically. The judgment is still ours. The assembly is the agent's. That is the right division of labour.</p></blockquote> <h2>What QS Firms Get Wrong About AI</h2> <figure> <img src="https://images.unsplash.com/photo-1553877522-43269d4ea984?w=1200&q=80" alt="AI strategy for UK quantity surveying firms — the right division of labour between human professional judgment and AI agent document assembly" width="1200" height="800" loading="lazy" /> </figure> <p>Most QS practices that engage with AI fall into one of two traps. The first is using AI as a writing tool — prompting ChatGPT to write a paragraph of the cost plan narrative, then copying it in. Useful, but it saves minutes rather than hours. The second is dismissing AI entirely because "every project is different." Both responses miss the architecture.</p> <p>Yes, every project is different. The cost plan rates change, the specification level changes, the client changes. But the structure of a preliminary cost plan follows NRM. The structure of an invitation-to-tender package follows the same framework every time. The structure of a fee proposal has five components in the same order on every commission. The variation lives in the inputs. The assembly structure is entirely predictable — and predictability is the condition for delegation to an agent.</p> <p>The other common mistake is thinking that professional liability creates a barrier to AI use. It does not. It creates a requirement for human review — which is exactly what the <a href="/blog/human-in-the-loop-ai-agents-uk">human-in-the-loop architecture</a> provides. The agent drafts. The qualified professional checks, approves, and takes responsibility for the output. The liability framework and the AI architecture are compatible by design.</p> <p>If you run a quantity surveying practice, a building surveying firm, a project management consultancy, or any professional services business in the built environment where document production follows a consistent pattern, the question is not whether AI can help. The question is which workflow to tackle first.</p> <p>The <a href="/blog/ai-delegation-matrix-uk-service-businesses">AI delegation matrix</a> is the right starting point — it maps every task against repeatability and professional consequence to identify where agents should run first. For most QS practices, preliminary cost plan drafting lands in the highest-priority zone: high repeatability, clear human review checkpoint, immediate impact on capacity.</p> <p>If you want to understand what an AI operating system for your construction consultancy or professional services practice could look like, <a href="/contact">get in touch</a>. We design and build AI operating systems for UK service businesses — starting with the workflows that deliver the fastest return and building from there.</p>