“A repetitive task follows rules we could write down, but a person still does it by hand.”
A business assistant grounded in your company, helping you think, plan and draft, with tools and scheduled jobs added under your control.
Sound familiar?
The knot.
Your team wastes hours on repetitive work a machine should handle.
There's work that needs doing but it's repetitive, tedious, and eats time your team should spend on real work. You're stretched thin, your team is stretched thin, and the answer isn't just "work harder." But off-the-shelf AI feels risky without the right controls in place.
Three signs it’s this
- 01 The task follows rules you could write down on one page.
- 02 Skilled people spend hours on it that they could spend on judgement work.
- 03 Mistakes happen because it is repetitive, not because it is hard.
T40 answer · Custom AI Agents
The fix.
We set up a business agent around a role in your business, not a one-task chatbot. It runs inside a harness: the working environment and controls around the AI, not the AI model itself. The harness gives the agent approved company knowledge, maintained memory, reusable skills, the tools the job needs, a schedule and clear points where a person approves. The foundation is a context-aware assistant for everyday conversation, planning and drafting across subjects. We name an owner, keep private and shared context separate, and scope connected systems and action permissions independently. Another conversation topic is not another chargeable job.
After An assistant for that defined job, with your people setting the rules and approving what matters.
What you get.
The deliverables, plainly listed. If it's not here, ask — we'll tell you honestly whether it's included.
- 01 Business and owner context, a named owner, everyday conversation and a written definition of what good looks like
- 02 An agent harness configured or extended on a framework chosen to fit the job, such as Hermes, OpenClaw or pi
- 03 Approved company knowledge, maintained memory and reusable skills, with live records ranked above memory
- 04 Tools and integrations limited to what the job needs, starting read-only, in accounts you own
- 05 Schedules, logging and human approval for anything external or any material change
- 06 Testing, a controlled release, user training and a handover covering running costs and rollback
How this one works.
No discovery phases that outlive the project. Just the steps that earn their place.
- 01 Set up the foundation We learn the business, the owner goals and working preferences, and agree approved sources, authorised users and what good looks like. Everyday conversation and drafting are the foundation. Numbers it reports are calculated from source records, not estimated by the AI.
- 02 Choose and shape the harness We choose the framework and AI model for the job. Hermes and OpenClaw give a running start; pi is a small, extensible building block for jobs that need more engineering. No framework supplies every control on its own, so we build the ones that are missing.
- 03 Test before trust We test on normal, awkward and failure cases, including attempts to make the agent ignore its rules. Access is checked, actions are logged and there is a way to pause or roll back.
- 04 Release, then expand Named people pilot it before wider release, and we train the team. A new job or a new write permission is added the same way, one at a time, after the last one has proved itself.
Why this matters
Open-source agent frameworks put business agents within reach of small teams. The model is the easy part. The value sits in the harness around it: the right knowledge, the right tools, sensible limits and a person accountable for the result. We run our own business on Hermes with scheduled jobs and human review, which is where our controls come from. That is our own experience, not a client result.
Who this suits
- Small teams with more recurring work than capacity to handle it
- Businesses with a repeated job that follows rules and draws on records they already keep
- Owners who want an agent that works across their own tools, not another standalone chatbot
- Leaders who will name an owner and approve what the agent sends or changes
When we would advise against it
- Every case needs a fresh judgement call. An agent would need so much checking it saves nothing.
- There is no one to approve what the agent does. We build with a named human owner, or not at all.
The questions everyone asks.
If yours isn't here, ask it — we answer emails within 24 hours.
Q1 What is the difference between an agent and a chatbot?
A chatbot waits for questions. A business agent has a role and jobs to do, such as a morning pipeline check or a weekly report. It works inside a harness: the working environment and controls around the AI, including its knowledge, tools, schedule and approval rules.
Q2 Which framework do you use?
The one that fits the job. We run our own business on Hermes and also select and configure OpenClaw. pi is a small, extensible agent toolkit we build on when a job needs more engineering; it is a building block, not a ready-made business setup. Whichever we use, we add and test any control the job needs that the framework does not supply.
Q3 Will it sound like us?
We give it your approved documents, terms and examples, and a person reviews drafts before anything goes to a customer. We do not present it as a person.
Q4 What if the agent makes mistakes?
All AI makes mistakes sometimes, and an agent does not fix itself. We design for this: read-only first, numbers calculated from source records, human approval for anything external, logging and a way to pause or roll back. When something goes wrong, a person reviews it and we change the setup.
Q5 How much does a custom AI agent cost?
An initial defined agent setup starts from £2,500 +VAT. A larger harness with more connected systems or scheduled work is quoted separately. AI provider usage is paid separately, straight to the provider, and we set a budget and alerts up front.
Sound like your problem?
Tell us where it hurts. We’ll tell you honestly whether we can help — and what we’d do first.
First paid step · T40 Systems: AI Strategy Session · £395, credited against the build