Workflow automation
Workflow automation that knows your organisation
For any team with work that comes round every week. Built around what your organisation knows, in five parts: the brain, AI agents that draft and watch, a workspace where a person approves every draft, your website where you need it, and training.
The work that comes round every week.
Each needs judgement, and hours of gathering, drafting and formatting around it. In a small team, that desk work crowds out the work only people can do.
AI can take much of it on, but only with the right context. Most organisations have no structured set of files that captures what they do, so a chat window knows only what someone remembers to paste in, and the drafts come back generic.
- The weekly report
- The month’s figures, reconciled
- Meeting notes turned into action items
- The inbox, read and replied to
- The proposal or quote
- The briefing before a meeting
- Chasing what’s outstanding
Back again next week
One build, five parts.
Workflow automation is the part of the system that takes on the routine. We build it in five parts: the brain, the agents, the workspace, your website where you need it, and training. What one part does feeds the next, and what changes flows back to the brain each day. That loop is why the build gets better with use. Bought as separate tools, each would need someone to keep it fed; built as one, the everyday work keeps it current.
The five parts of the automation build. At the centre, the brain: everything the organisation knows, in plain files. It feeds the agents your facts and tone of voice, and their drafts go to the workspace for a person to approve. What people change in the workspace goes back to the brain, kept for next time. Where the organisation needs one, its website sends sign-ups and enquiries to the workspace and what it says in public to the brain. Training sits with the team, covering how they work with all of it.
What is the workspace?
A custom CRM, built for what a CRM is for: members, contacts, renewals. Connected to the brain and the agents, it becomes the home screen for the whole build. One page shows what needs you first, then links to your organisation’s knowledge, relationships and work. Drafts arrive there for review, and records update as people work, without anyone filling in a form. What you know about members and stakeholders sits in one place, so you can see how they’re likely to react before it happens.
Workflow automation is one of three settings for the same science, with training, and research and evaluation. How the whole system fits together
How a build runs.
We put it into use early and hand it over unfinished, on purpose: a month of real material is worth more than a month of guessing.
- 1
Discover
With your leadership: the work, the people, the risks, what matters most.
- 2
Build
On your own knowledge, starting with the work that costs the most time.
- 3
Test
On real work, before anyone depends on it.
- 4
Hand over
With training and an adoption plan: a champion, and results people can see.
Training comes with every build
A build only delivers if people keep using it, and use is what keeps the brain current. So training comes as each part goes live, with a named champion and results people can see, and reviewing drafts well is taught in its own right. How we train
Your data
Your data, your content and everything written for you stay yours, and nothing goes out until a named person approves it.
The brain: everything your organisation knows, in plain files.
Every other part reads from it, so it’s where we start.
The brain growing with use, in three circles. On day one, a few plain text files, each a node, seeded with the organisation’s documents, with meetings, drafts, emails and posts flowing in as arrows. A few weeks in, the files are connected by use. A few months in, they form a dense network, with the same everyday work still flowing in.
At its core, the brain is powered by the same two simple things that power a human brain: cells and experiences. Its cells are plain text files, the native language of large language models (the AI behind tools like Claude and ChatGPT), held in a structured system of folders. Its experiences are ‘the everyday’: meetings with staff and stakeholders, documents in draft, emails sent and received, social posts that land or fall flat.
And just as in a human brain, when the relationships between these simple components are mapped, a far more sophisticated system emerges. By seeding folders from day one with your existing documents, carefully curating how each file connects to the wider network, and automating a continuous feed of new information from ‘the everyday’, the outcome is an ecosystem of knowledge that grows in lockstep with your organisation.
What’s inside it, and how is it filled?
Folders built around how your organisation actually works, set with you at the start. Into them goes what you already hold: positions, past submissions, board papers, your style guide, your website’s content, each turned into a plain text file and linked to the notes it informs. From then on, every recorded meeting, every draft and every filed email adds to it. Several people can work in it at once, and where their notes disagree, a person decides which stands.
An evolving knowledge base is the bedrock of automation that works well, and keeps working. Agents fed stale knowledge make more mistakes over time and become harder to work with. Get the structure of your knowledge right from the outset, and automation gets better over time.
Kept current by the work
Once it’s built, the brain practically organises and grows itself as people use it: a wrap-up at the end of each day, an inbox that files new material, a briefing before each meeting.
The brain is plain files. Any AI assistant can read them: Claude, ChatGPT, Gemini, Copilot. Change tools and the files stay, with everything they’ve learned.
Autonomy, too. No one can predict how AI will change, and free models are closing the gap on paid tools fast.
What if we change AI tools?
Nothing you know is held inside one AI product. The agents’ instructions are plain files too, documented, so if you move to another tool, they move with your knowledge rather than being rebuilt from scratch.
The brain’s files at the centre, with four AI assistants around them, each on one connection of the same kind. Claude, Gemini and Copilot are plugged in. ChatGPT’s plug is pulled out and labelled unplugged. The files stay the same whichever tools are connected.
The agents, and what sets them apart.
When we build an agent, we go beyond a generic template. Each one is built for one kind of output, from your organisation’s own best work, so it drafts to your template and in your voice. What sets them apart is the expertise behind them. We’re cognitive scientists, with research in judgement and decision-making, and we build in an evidence-based approach, which is the difference between an agent that runs through a process and one that crafts. And every draft comes back to a person, with a prompt built from that draft on what to check, because people rarely review well on their own.
How is each agent built to do the job well?
- Discovery. With your leadership, we map how each piece of work is done now, what good looks like and where the time goes. That decides which specialists come first.
- Your best work. Each agent learns from your own finished work: the reports that landed, the posts that worked.
- Your house style. Your voice, templates and positions are written down once, in the brain, and every agent works from the same copy. Improve it once and every agent improves.
- Tested on real work. Each agent is tested on your real material and judged against your finished work before handover, then tuned again on real use.
How a job runs, shown with an illustrative report to a funder. A person briefs the writing lead, an AI agent, which reads the brain for the organisation’s positions and evidence and briefs one of several specialist agents, the report writer. A separate check scores the draft and checks every claim against its source. The draft arrives in the workspace with a prompt. A person reviews it well and approves it, and what they changed goes back into the brain.
A report to a funder, from brief to reviewed draft
Brief
Someone briefs the writing lead, an agent: your position, the figures, the deadline.
The lead plans
It reads the brain for your positions and evidence, and briefs a specialist.
A specialist drafts
- Report writer
- Proposal writer
- Update writer
- Briefing writer
The report writer drafts to your template, in your voice.
Checked
A separate check scores the draft against what good looks like, and checks every claim against its source.
It lands in the workspace with its prompt.
A person approves
You make the calls, edit, and approve.
The brain learns
What you changed is written back, so the next report starts closer to yours.
The prompted review
Report to the funder: community transport
Demand for community transport among our clients has risen by a third in two years, while program funding has stayed flat.
We recommend that funding be indexed to demand, and that eligibility rules be simplified.
Illustrative: an invented draft and its prompt. The questions change with every draft, because they’re built from it.
How is the prompt built?
From the draft itself. The independent check scores the draft against what good looks like for that kind of document, and flags any claim it can’t match to a source. The prompt gathers what matters most from both: where the draft scored lowest, which claims were flagged, and the decisions only a person can make. It opens with a question for the reviewer to answer before seeing the system’s view, so their own judgement comes first.
First, your view
Before you see the system’s view: which claim here would you least want the funder to test?
Two claims to check
- “Risen by a third in two years”: the client survey behind it gives a range, not one figure.
- “Funding has stayed flat”: no source found in your files. Add one, or soften it.
Where it’s weakest
The recommendations: neither yet gives the funder a first step it could act on.
Yours to decide
How hard to push on indexing, given where your relationship with the funder stands.
Approve · Edit · Send back
Only a signed-in person can approve.
Will it make things up?
It can, which is why it’s built not to let that reach anyone. Every draft is grounded in your own material. An independent check reviews it, and every factual claim is checked against its source before a person sees it. In the review, claims and sources sit side by side, so checking one takes moments.
Every agent is built from your best work and tested on your real material before handover.
What it adds.
We don’t put a headline figure on what a build adds, and we’d be wary of anyone who does. A number in hours or dollars saved is only honest where you can compare how long a task took before and after, and most builds add capacity where there was no one to do the work at all, so a confident figure would be a guess presented as a fact. What we can do is show where the value falls, so you can judge it for your own organisation.
Time
The desk work shrinks; the judgement stays with people.
Reports, proposals and updates start from a checked draft, and what’s left is the human part: deciding, and a careful review.
Reach
What would have slipped past, found.
An agent that reads every source every morning finds the funding round or announcement a busy person would miss. Nobody can know what they would have missed, so we don’t count it.
Coordination
One source of truth, current as people work.
Fewer handovers, searches and “where’s that up to?”s, and knowledge that stays when people move on: real, and hard to put a number on.
Time back for the work that needs judgement is the goal, not a guarantee.
A build can start with the one piece of work that costs your team the most time, and grow from there. Each is scoped with you and priced before work starts, with its monthly running cost beside it.
Start with a conversation.
Bring one piece of work that comes round every week, and we’ll talk through how it could be automated and what that would take.