The call distilled to the decisions that matter — what Simon's team wants to build, where the data lives, the approach Andrew would take, and how an engagement would actually start. Read top to bottom.
Right now the team are heavy users of ChatGPT and Claude — projects spun up, documents loaded in, answers bounced between tools. Useful, but as Simon's boss put it, that's twenty chatbots doing twenty jobs.
The goal is an operating system above them: an intelligence layer one or two people can ask anything across the whole business — plus a handful of discrete agents for the thorny functions. Realistically three or four: clients & new business, finance & operations, and so on.
A 140-person marketing group across London, the US and expanding into Asia — and the knowledge is scattered. About 90% sits in Google Drive, in ordinary drives. The rest is the usual legacy spread: some on Dropbox and a legacy Microsoft system, mostly used by finance.
The cheapest, highest-leverage move isn't new infrastructure. It's taking the existing folders and documents and making small tweaks — a couple of map files, cleaner naming conventions — so an AI agent can explore them the way a new hire would.
Talk to enough people and you'll hear two approaches. A database has real benefits but a lot of setup and cost. The lighter path: structure what you already have, and keep the database in the back pocket until something proves you need it.
A few files and naming rules so Claude can walk the drive: this folder is clients, each has meeting notes, each note is named by its date. Low effort, fast to value.
Powerful, but heavy on setup and ongoing cost. Reach for it only if the simpler approach hits a wall — not as the default first step.
The architecture is a discovery route. At the top sits a simple map — there's a clients folder, a finance folder, an operations folder. When you ask something, Claude reads the map, then goes folder by folder, loading a specific file in each so it understands the context.
Ask to change a statement of work for a client and it knows to open that client's folder, load the context, and become the expert on them. Bite-sized steps are the whole trick — they keep the context small so Claude never gets overloaded.
Cowork is friendlier, but it runs locally on one machine, which makes it hard to share context across a team. For a layer the whole company leans on, Claude Code is the fit — it's what walks the folders on that discovery route.
Runs the folder-by-folder discovery, manages context in small pieces, and is the better base for something a team relies on.
Cowork is local and harder to share in a team setting. n8n likely plays only a small part in this build.
Beneath the top layer sit the working agents — each scoped to one function, with access only to the folders it needs. A few that came up directly on the call:
Jigar's clearest example: a finance agent watching the accounts-payable and accounts-receivable inboxes. Set up roughly twenty rules; once an incoming invoice is authenticated against them, the agent flags it, logs the activity, picks up the invoice and posts it into Xero — to the right one of seven companies.
Two ways to keep this alive: teach AI fluency across the org so people set up their own agents, or have Andrew act as the part-time AI person. The honest answer is a mix — he'd be on for the journey, six to nine months, while people get up to speed. Not a train-and-leave.
What makes or breaks the timeline is the amount of data and how well structured it is — getting things into shape can be most of the work. The plan: a short paid discovery first, and if it isn't viable, the money comes back.