By Dave Fuller, Managing Director, Accent Design
I have been building things for the web since 1998. I started as a developer, became Technical Director, and now run Accent. In all that time the most expensive part of running an agency has never been the code. It has been the slow leak of context: the decision made on a Teams call that nobody wrote down, the Slack thread that drifts off into a different problem, the email chain forwarded with five attachments and no summary. By the time the developer sits down to build, half the original intent is gone.
Early in 2024 I sat down with our monthly time logs and totted up how much of my project management work was actually managing the project, versus chasing context that should already have existed. The honest answer was uncomfortable. So we set ourselves a target: keep the standards, lose the busywork. Eighteen months on, our monthly project management hours are down by roughly 70%, measured against early 2024 and excluding account handling and client meetings (those have stayed where they were, because they are the work, not the admin around it).
Here is what we actually did, why it worked, and where it still needs a human.
The problem we were trying to solve
Most agencies our size use the same toolkit we do: Microsoft Teams for client calls, Zoom or Google Meet when a client prefers it, Slack internally, Office 365 for email, plus the usual board management in Jira or Trello. None of those tools talk to each other in a way that builds a single, reliable picture of a project. They are channels, not memory.
Our longest-running client is Agency Express , the property board service used by around 62% of UK estate agencies. We have been developing and maintaining their SignMaster system for over fifteen years. SignMaster is a .NET and ReactJS web application that handles job management, route planning, invoicing through Sage, and payments through WorldPay, with companion mobile apps for agents and operators. It is the sort of long-lived, integration-heavy product where context is everything. A change to the postcode lookup might affect routing, which affects the operator app, which has knock-on effects for invoicing. Forget one of those threads in a meeting and you create a bug three months later.
That is the project we used as our test case.
The workflow, plainly
We capture, we synthesise, we store, we instruct. None of those four steps is glamorous. All four matter.
Capture. Every project meeting is recorded with the client's permission and transcribed. Teams, Zoom, and Meet handle their own calls. For in-person meetings we use Apple Voice Memos and transcribe afterwards. The raw transcript is the source material, not the deliverable.
Synthesise. We feed transcripts to Claude Opus 4.7 with a prompt that asks for decisions made, decisions deferred, action owners, and any technical constraint that was raised. The output is short and structured. Anything Claude is unsure about is flagged rather than guessed. That last bit matters more than the summary itself, and I will come back to it.
Store. Each active project has its own Claude Project. Summaries, specs, integration notes, and historical decisions all sit inside it. When a developer or PM asks a question, the answer comes from the project's own history, not from a generic model trying to be helpful.
Instruct. Every project has a system prompt that tells Claude what this project is, what it absolutely is not, who the stakeholders are, and what house style to use. For SignMaster the system prompt makes Claude aware that we are working inside a fifteen-year-old codebase, that backwards compatibility is non-negotiable, and that integrations with Sage, WorldPay and Acquaint CRM must never be assumed to behave like generic equivalents.
For email we run Microsoft Copilot inside Office 365 to summarise threads by date. Before any of that summary content moves into our Claude environment, Copilot is also used to scrub passwords, API keys, and personally identifiable information. We keep separate markdown files per manager, so an exported summary for one project lead is emails-Dave.md and another's is emails-Karen.md . It looks pedantic. It is pedantic. It also means accountability never blurs.
Why the human-in-the-loop is the whole point
The mistake I see most often when other agencies ask us about this is treating AI summaries as truth. They are not. They are a strong first draft of truth.
Every Claude summary on an Accent project is read by the project manager before anything is acted on. We look for three things: factual hallucinations (Claude inventing a decision that was not made), omissions (a quiet but important point buried in the transcript), and tone drift (a client's "we'd quite like to consider" turning into "the client requires"). On a typical hour-long Agency Express call, the PM spends maybe ten minutes verifying the summary against the recording. That ten minutes is the difference between a useful tool and a liability.
This is the part that earns the trust. A model that flags its own uncertainty, plus a human who actually checks the flags, gives you something you can defend to a client when they ask how you remembered a detail from six months ago.
What changed for the client
The 70% number is the headline, but it is not the most interesting result. Three things shifted that we did not predict.
Meetings got shorter, because less time is spent re-establishing what was agreed last time. With Agency Express we now go into a fortnightly call with a one-page brief that the client has seen in advance, and we leave with action points already drafted. A call that used to run ninety minutes runs forty-five.
UI and UX feedback got faster. When a client says "can we see what that looks like" mid-meeting, we have the project history in front of us and can mock something up the same day rather than the following week. Clients feel included rather than managed.
We have quietly become what one Agency Express stakeholder called our "unofficial software department". Their team rotates, ours does not. After fifteen years we hold more institutional memory of SignMaster than anyone else, and the AI workflow is now the place that memory lives. That is not a marketing line, it is a responsibility, and it is the reason the security and scrubbing steps are not optional.
What we reinvested the time into
Time saved on admin has not gone into doing more billable hours. It has gone back into the work itself. More prototype iterations before a build starts. More accessibility testing, which matters to me personally given my background in accessible design. More R&D into the AI tooling we now offer to other businesses as a consultancy service, because the lessons we learned on our own projects translate well to clients who are drowning in their own digital noise.
If you are thinking of trying this
A few honest cautions. The setup is not free; the licences for Claude, Copilot, and the meeting platforms add up, and the time to write good system prompts is significant. Expect a month of awkwardness before anyone trusts the output. Keep the scrubbing step. Never let the AI be the only memory of a decision; the human-verified summary is.
And do not, under any circumstances, let the tool become an excuse to talk to clients less. The clients I have kept for fifteen years are the ones I have kept showing up for. The AI is what frees up the hours to do that properly.
Dave Fuller is Managing Director of Accent, based at the Enterprise Centre, University of East Anglia. He has worked in web development for 27 years, holds qualifications in web accessibility and design, and was previously Technical Director at Accent.