For established organisations
Your organisation grew. Its systems didn’t.
We connect what your organisation knows, the tools it runs on and the people who make decisions, and put AI where it actually helps.
- Knowledge lives in documents, drives and a few people’s heads.
- Ten tools, and the same information typed into four of them.
- Staff spend their mornings answering the same questions.
- An AI pilot impressed everyone in a meeting and then never reached real work.
How the pieces connect.
The same shape works for a knowledge assistant, a document pipeline or a whole operations system. What changes is which pieces you need.
Your sources
- Documents and teaching material
- CRM and contacts
- Course or membership platform
- Inbox and forms
- Spreadsheets
- Stage 1
Knowledge layer
One searchable index over what your organisation knows, with permissions, so people and assistants only see what they’re allowed to.
- Stage 2
Workflows and agents
Each with one clear job: read the application, draft the reply, update the record. Most problems need a workflow, not an agent.
- Stage 3
Human approval
People sign off anything that matters, from the screen they already use, with the sources shown alongside.
- Stage 4
Actions in your tools
Replies sent, records updated, documents filed, dashboards current. In the systems your team already lives in.
Monitoring
Every run is logged: what it read, what it did, what it’s waiting on.
Evaluation
Tested against your real examples before launch and after every change.
Security
Least access by default. Your data stays in accounts you own.
Monitoring, evaluation and security run alongside every stage, not after it.
Watch one message go through it.
Pick a message, or write your own. The system reads it, decides what it is, routes it, and stops for a person where it matters. It’s the shape of a real one, running on simple rules in your browser.
Read
Turned into a record
Decide
Which kind of message
Route
Who handles it
Approve
A person, if it matters
Act
What actually happens
$ waiting for a message
Simulated. These rules run in your browser: plain keyword matching, no AI model, and nothing is sent. In a real system a model does the reading. The routing, the approval step and the log look like this.
What we build.
Every one of these can be small. A single assistant over one body of material is a real system, and often the right first one.
- Knowledge assistants
- Staff or members ask in plain language and get answers from your own material, with the source attached.
- Document processing
- Applications, invoices and forms read, summarised and routed before anyone opens them.
- Operations systems
- One record of clients, students or members, and the workflows that keep it current.
- Agent orchestration
- Several narrow agents passing work between them, when the job really has several stages.
- Staff-facing dashboards
- What every automated process did today, what needs a decision, and what’s stuck.
- Education platforms
- Courses, cohorts and teacher tools that connect to the rest of the organisation.
Try the approval step.
This is the moment that makes AI safe to use in front of your members or customers. The assistant drafts, shows what it based the answer on, and waits for a person.
Message from a member
Hi! I missed Thursday’s live session. Is the recording up yet, and do I need to finish lesson 3 before the next one?
Sofia, member since March
Draft reply
Waiting for you
Hi Sofia, the recording from Thursday is in the members’ area under Live sessions. You don’t need to finish lesson 3 first: the next session builds on lesson 2, so you’re ready. See you there!
Based on
- Live sessions page, updated Friday
- Course outline, lesson 3 notes
Audit log
Nothing yet. Every decision you make here is recorded with its sources.
Simulated. No model is called and nothing is sent.
How we decide what to build.
Integrations first
Most of the value is in connecting the tools you already pay for. We use their APIs, and MCP services where an assistant needs safe, defined access to a system.
People approve what matters
Anything that reaches a customer, moves money or changes a record someone relies on goes past a person first. Routine steps don’t.
Boring where it counts
Standard databases, standard hosting, code in your accounts. If we disappeared tomorrow, another engineer could pick it up.
Measured, not admired
We agree what good looks like before we build, and check it with real examples. If it isn’t helping, we say so.
We’ll talk you out of complexity.
Multi-agent systems are sometimes the right answer. Usually a single, well-built workflow is cheaper, faster and easier to trust. We start there and add agents only when the work needs them.
We know where that line is because we run our own agent experiments, like Orbis, before anything similar goes near a client.
How an engagement runs.
- 1
A conversation
Thirty minutes, or a $400 working session if you want something concrete straight away.
- 2
A map
Where your knowledge lives, which work repeats, and who decides. It’s a working session, $400 or $800 for a full day, credited toward what follows. Written up so you can use it with or without us.
- 3
One working system
We build the one piece that saves the most time, and put it into real use with your team.
- 4
Then the next
We measure it, fix what’s rough, and connect the next piece only when the first is earning its keep.
What it costs
- Defined projectfrom $1,500
A single assistant, automation or integration.
- Bigger buildfrom $3,500
Connected automations, a migration, or several pieces together.
- Intelligent systemfrom $12,000
Substantial connected systems across your knowledge, tools and team, with approvals and monitoring.
- Complex and enterprise projectsCustom quote
Complex platforms, multi-agent and enterprise systems.
Smaller AI, automation, integration and infrastructure projects are quoted as defined projects or bigger builds. After the map, you get one fixed price before any build work begins.
Or build it month by month: Build at $3,000 or Embedded at $6,000 a month →
Tell us what’s in the way.
Where the knowledge lives, which work repeats, what you’ve tried. We read every message ourselves and reply within two working days.