Everyone has a chatbot, nobody has a system
Staff use whichever AI tool they like, with company and customer data, and the results never make it back into the systems that run the business.
AI for non-regulated businesses
AI for growing businesses
Minibit helps non-regulated businesses, from owner-run firms to product teams, put AI to work on the jobs that eat the week: admin, documents, phones, reporting and the features customers keep asking for. The same discipline as my regulated work, without the overhead you do not need.
The starting point
What I usually find
Most growing businesses are not short of AI ideas. They are short of someone who has shipped them and can say which ones will pay back.
Double keying between tools
The CRM, the accounts package, the project tool and the inbox each hold part of the truth, and a person copies between them every day.
The phone and the inbox
Enquiries arrive at all hours, get answered late or not at all, and the ones that convert are the ones that happened to be picked up.
Reports built by hand
The weekly numbers take someone half a day to assemble, which is why they are usually late and occasionally wrong.
An AI feature backlog
Customers are asking for AI in your product, the team has not built agents before, and the risk of shipping something embarrassing is real.
No idea what any of it is worth
Pilots start and stop with no baseline, so nobody can say whether AI saved money or just moved the work around.
What I build
What I build
Six things that come up again and again. Each one is scoped in discovery, piloted on live work, and measured in hours and pounds.
AI assistants on your templates
First drafts of proposals, reports, letters and replies in your voice, from your templates and your records, waiting for a person to approve.
Automations between your systems
Data that moves itself: the enquiry into the CRM, the signed quote into the accounts package, the chaser sent on day seven without anyone remembering.
Voice agents that answer the phone
A receptionist that picks up every call, answers the common questions, books appointments and hands off to a person with a summary. The pattern is proven in Mini Calls, my own voice agent.
AI reports and dashboards
The weekly numbers assembled from your systems, written up in plain English and sent before anyone asks.
AI features in your product
Agents, assistants and MCP connectors inside your own product, with evals, approval steps, spend caps and prompt-injection defence so you can ship with confidence.
Launch checks for AI-built apps
If your team is building with AI coding tools, Ship Keys checks the result for the security, privacy and quality problems those tools tend to leave behind, before launch.
Product and tech teams
Agentic AI done properly, by someone who has shipped it
Most teams building agentic AI have never worked where mistakes are regulated, and most people from regulated industries have never built agentic systems. I have done both, and I bring that to your team as an architect, a builder or a reviewer: MCP servers and clients, Claude tool use and structured output, OAuth 2.1 connectors, headless agents in CI, voice agents, LangGraph and RAG pipelines with human approval checkpoints.
Engagements range from a two-week architecture review to building the first agent alongside your engineers and leaving them with the evals, guardrails and patterns to carry on.
MCP servers and clientsClaude tool useOAuth 2.1 connectorsClaude Code pluginsHeadless agents in CIVoice agentsLangGraphRAG and pgvectorEvalsGuardrailsPrompt-injection defenceSpend capsAudit trails
Ship KeysLaunch-readiness checks as a connector and a Claude Code plugin, with agents in customers’ CI
Private betaMini StepsGoal planning with owner-approved actions and end-to-end voice through MCP
Private betaMini NavigatorFamily records with strong privacy controls and model choice by query complexity
In developmentMini CallsVoice agent with MCP clients, secured actions and multi-agent hand-off
BuiltThe approach
Why the regulated discipline still helps you
You do not need a compliance manual. You do need AI that does not embarrass you, and that is the same engineering.
Nothing wrong goes out the door
A person approves anything customer-facing until the measurements say it is safe to loosen. Approval of a good draft is faster than writing from scratch, so this does not slow you down.
You can see what the AI did
Every action is logged with its inputs and outputs. When a customer asks why they got a particular email, you can answer in a minute rather than a week.
It is measured
Hours before and after, errors caught, calls answered, invoices chased. The pilot has a baseline and a decision date, so you know what it is worth before it grows.
You are not locked in
Standard tools, open protocols and code you own. If I disappeared tomorrow, another developer could pick it up, and you could stop support on a month’s notice.
The process
How it runs
Small steps, each one you can stop after. The aim is a measured pilot within six weeks.
1
Talk
30 minutesWhat is eating the week, and whether AI is the right lever for it.
2
Discover
2 to 4 daysI watch the work, count the hours and write up costed options.
3
Pilot
3 to 5 weeksOne workflow on live work, with a baseline and a decision date.
4
Run
MonthlyTuning, a short monthly report and someone to call. Stop on a month’s notice.
FAQ
Questions growing businesses ask
We are not regulated. Why would we want the regulated approach?
Because the same things that keep a regulator happy keep your customers happy: nothing wrong goes out the door, you can see what the AI did, and you are not locked into a tool that disappears. I am Steven Minnikin, and I bring that discipline from 11 years’ experience in regulated financial services. You get it without the paperwork you do not need.
What can AI realistically take off our plate?
The repetitive work with a clear pattern: first drafts of documents and emails, data moved between systems, phone enquiries out of hours, weekly reports, chasing and reminders. I measure the hours before and after so you know what it was worth.
Do we need new software?
Usually not much. Most of the value comes from connecting what you already have and putting AI on the steps in between. Where a new tool is needed, I recommend standard ones you could run without me.
Can you build an AI feature into our own product?
Yes. I build agents, assistants and MCP connectors into existing products, with the guardrails that make them safe to ship: evals, approval steps, spend caps and prompt-injection defence. See the work for examples.
How do we start?
A 30-minute call, then a short fixed-fee discovery if it makes sense. The aim is a measured pilot within six weeks of the first call.
Tell me what is eating the week.
Thirty minutes, no cost, and you will leave knowing which of it AI can take and roughly what that would be worth.
steve@minibit.co.uk · Newcastle upon Tyne, working across the UK