A six-person architecture practice in Bristol was losing 28 hours every week to three processes none of them had gone into practice to do: bid writing, planning documentation drafting, and client progress reporting. Three AI agents changed the maths — 65% more tenders submitted, four to five wins per quarter instead of two to three, and 24 hours recovered every week. The system runs at £80 per month. This is what they built and how.
The Admin Trap That Caps Every Small Architecture Practice
Architecture is a skills business. Clients pay for design thinking, planning expertise, and project leadership. What they do not pay for — and what consumes the majority of non-billable hours in any small practice — is the documentation layer that wraps around every project.
For the Bristol practice in this study, the admin trap broke down into three categories:
- Bid writing. The practice was responding to eight to twelve tenders per quarter — public sector frameworks, design competitions, developer RFPs. Each required pulling company CVs, assembling project case studies, writing method statements, and formatting output to the client's specification. Average time per bid: five to eight hours. With three to four senior people involved in each one, the real cost was closer to twelve hours of combined professional time.
- Planning documentation. Most projects required a planning statement, a design and access statement, and at least one technical report. Drafting these from scratch took six to ten hours per application, even for sites with no unusual constraints. The content was often 70–80% templated, but pulling together, adapting, and formatting it was purely manual.
- Client progress reports. Monthly reports for each live project took between two and three hours per project. The practice had nine active projects. That was up to 27 hours of report-writing per month — work that required accuracy and professional presentation but very little architectural judgement.
Total: 28 hours a week of skilled professional time consumed by structured, largely repeatable documentation. The practice was not growing because its directors were too busy writing instead of designing and winning work.
The practice was not growing because its directors were too busy writing instead of designing and winning work. The output was good — the problem was the input it required.
Three Agents, One Architecture AI Operating System
The solution was not a document tool. It was an AI operating system built around the three highest-volume documentation processes, each handled by a dedicated agent connected to the practice's existing project data.
Agent 1 — The Bid Writer
The bid writer agent monitors the Find a Tender service and the practice's priority client portals for new opportunities daily. When a relevant notice appears, it pulls the specification and scores it against a qualification matrix — project type, contract value, location, submission deadline — and flags opportunities that meet the threshold to the relevant director. No manual monitoring, no missed notices.
When the director approves pursuit, the agent drafts the full response. It reads the tender's evaluation criteria, pulls matching case studies from the practice's project library in Notion, selects relevant team CVs, and generates a first draft of the method statement and quality responses — structured to the tender's format and word count. The director's job is to review, refine, and add the strategic layer. Not to start from a blank page.
The practice moved from spending five to eight hours per bid to spending ninety minutes to two hours reviewing an AI-drafted submission. They went from submitting eight tenders per quarter to thirteen. Win rate stayed consistent — which meant more wins, not just more submissions.
Agent 2 — The Planning Documentation Agent
The planning documentation agent connects to the practice's project management system and reads the project brief, site address, planning history, and any technical reports already completed. From this, it drafts a planning statement, a design and access statement, and — where the project constraints require it — a heritage statement or transport note.
The agent knows the structure each relevant local planning authority expects: which LPA requires a design principles section before material considerations, which requires the planning policy framework addressed under each objective separately. This institutional knowledge is embedded in the agent's configuration and updated when LPA validation requirements change.
What previously took six to ten hours per application now takes twenty minutes to generate and sixty to ninety minutes for a qualified planner to review, adapt, and approve. The practice submits planning applications faster and with less principal time tied up in drafting. For larger projects requiring multiple technical reports, the saving is proportionally higher.
Agent 3 — The Project Progress Agent
The project progress agent runs automatically on the 25th of each month. It reads notes from each active project in the practice's Notion workspace, pulls milestone data from the project schedule, and generates a structured monthly progress report for each client — covering work completed, decisions outstanding, upcoming milestones, and budget position.
Each report is formatted to the practice's house style, addressed to the correct client contact, and sent to the project architect for a ten-minute review before dispatch. The practice went from spending up to 27 hours per month on progress reports to under three hours total — the time to review nine AI-drafted documents, not write them.
The Numbers After 90 Days
Ninety days after deploying the three-agent system, the practice measured the following:
- Tenders submitted: Up from 8 to 13 per quarter — a 65% increase. The bottleneck had been time, not opportunity. With bids taking ninety minutes instead of six hours, the practice could pursue work they had previously been forced to decline.
- Planning applications processed: Up from 6 to 9 per quarter. Projects held up waiting for planning statements moved faster. Two projects that had been sitting at design stage for six weeks got their applications submitted within two weeks of the agent going live.
- Hours recovered weekly: 24 hours per week across the whole practice — equivalent to adding 0.6 of a full-time professional without recruiting. The directors reclaimed 12 of those hours; the rest came from senior associates no longer pulled into bid formatting and report drafting.
- Client satisfaction: All nine active clients received their first AI-assisted progress report within 48 hours of month end. The previous average had been 12 days. Two clients remarked on the improved consistency without knowing the process had changed.
- New project wins: Four new commissions in the 90-day period, against a quarterly average of two to three in the prior year. The practice attributes at least two of the additional wins to bids they would not previously have had the capacity to pursue.
What the System Costs to Run
The practice runs three AI agents on a self-hosted n8n instance with Claude as the underlying model. Monthly running costs break down as follows:
- n8n cloud: £30/month (Starter plan, covers all three agent workflows)
- Claude API: £35/month (bid drafts, planning documents, and progress reports combined)
- Notion integration: £0 (already in use as the practice's project management tool)
- Total: £65/month, rounded to £80 with a buffer for higher-volume months
The initial build — scoping the agents, integrating with Notion, training the planning documentation agent on LPA requirements, and testing against real bids and applications — took approximately four weeks of part-time implementation. The practice worked with an AI operating system specialist to scope and deploy, then ran independently from month two onward.
At £80/month running cost, the system pays for itself within the first additional tender win it enables. For this practice, the payback period on the full build cost was six weeks.
What Architecture Practices Are Missing
RIBA's 2026 survey found that 74% of UK architecture practices now use AI for most projects — up from 59% the year before. But the largest productivity gains are concentrated in visualisation and rendering, not in the administrative layer. Practices are using AI to generate images faster; they are not yet using it to systematically recover the 25–35% of professional time that goes into documentation that does not require architectural skill.
That is the gap this case study closes. The Bristol practice did not use AI to design better buildings. It used AI to free the people who design buildings from the documentation processes consuming a third of their working week.
The compounding effect matters here. Recovered professional time redirected to design, client relationships, and business development does not just save money — it generates revenue that was not previously possible. The compounding advantage post covers this in detail: for firms that deploy AI systematically and redirect recovered capacity, the productivity premium typically reaches 40–60% within 24 months.
For architecture practices, the highest-value applications of freed capacity are:
- Pursuing more tender opportunities the practice had previously been forced to decline
- Deepening client relationships on live projects rather than managing paperwork
- Spending more design development time on projects where quality differentiates
- Building the practice's portfolio of case studies and credentials — which in turn makes the bid agent's outputs stronger over time
The AI workforce model covers how to think about this sequencing: where to deploy agent capacity first, how to measure the redirect, and when recovered hours translate into revenue versus efficiency gains. For architecture practices, bid writing and planning documentation are almost always the right starting point — high volume, high time cost, highly structured, and directly connected to revenue through tender wins and planning approvals.
If you want to understand what a three-agent system would look like for your architecture practice — what it would automate, what it would cost, and what the realistic return looks like over 12 months — book a free 30-minute call. We will map your current documentation processes, identify the three highest-volume targets, and give you a clear picture of build cost and expected return before you commit to anything.