It reads.
Emails, web forms, invoices, order sheets, your own documents and your sales figures. It picks out what matters and files it in the right place.
I design and build connected AI systems for small and growing businesses. They read what comes in, know your documents, keep your records up to date, report back in plain English, and ask before doing anything that matters.
01 What this actually is
Most people meet AI as a chat window. Handy, but it can’t see your inbox, doesn’t know your price list and can’t update your records.
A business AI system is different: a few connected pieces, each with one clear job, working inside the tools you already use.
Emails, web forms, invoices, order sheets, your own documents and your sales figures. It picks out what matters and files it in the right place.
Updates your spreadsheet or customer list, drafts the reply, prepares the quote, builds the report. The repetitive steps you’d otherwise do by hand.
Anything important waits for your OK. Everything it did is written down, so you can see what happened, why, and what it cost.
02 The building blocks
You don’t need all of them. Most businesses start with one or two. Tap a piece to see what it does, in plain English.
Every piece reports to the control room. That’s where you approve things and see what happened.
Enquiry autopilot Demo planned
When a message, form or booking request arrives, it’s read, sorted by what the person wants, logged in your customer list, and answered with a draft reply that waits for your OK.
Copy the name, paste it in the spreadsheet, write the same reply again.
It’s already logged and the reply is drafted. You tap approve.
The nerd name Workflow automation + structured extraction
Watch the animated versionKnowledge assistant Demo planned
It reads your policies, price lists, manuals and FAQs, then answers staff or customers in seconds and shows exactly which page each answer came from. If it isn’t sure, it says so and passes the question to a person.
“Where’s that policy again?” Twelve times a day.
Instant answer, with a link to the source. New staff get up to speed faster.
The nerd name Retrieval-augmented generation (RAG) with citations
Paperwork reader Demo planned
Invoices, order forms, applications, even phone photos of paper. The important details are pulled out, checked against what you expected, and laid out on a review screen before anything is saved.
An afternoon of squinting at PDFs and typing numbers into boxes.
A tidy review screen. Mismatches are flagged in red. Glance, approve, done.
The nerd name Multimodal document extraction with human review
Connectors Part of every build
No need to move your business onto something new. The system connects to your email, spreadsheets, customer list, calendar or accounting software, and only gets the access each job actually needs.
Five apps that don’t talk to each other, and you as the go-between.
Information moves between them on its own, the same way every time.
The nerd name Integration layer with least-privilege tool access
Control room Flagship · In development
The heart of the system. A simple dashboard showing what each piece did today, what’s waiting for your approval, anything that went wrong, and what it all cost. Approve, edit or reject with one tap.
“Is the AI doing anything? Did that email actually go out?”
A clear history of every action, a queue of things to approve, and a spending total.
The nerd name Agent harness + operations console
Ask your numbers Working demo
Type “how did each location do last quarter?” and get a clear dashboard built from your real data, with charts, maps and a few key points. The numbers on screen come straight from your data, never from the AI’s imagination.
“I’ll get that report to you by Friday.”
The answer, with a chart, before your coffee cools.
The nerd name Tool-calling analytics agent + generative UI
Try Agent Workbench, the live demoCustom apps Demo planned
A booking page, a customer portal, a members’ area or an internal tool that replaces three spreadsheets. Sometimes AI is part of it, sometimes it isn’t. Good software doesn’t need AI to be useful.
A patchwork of spreadsheets, shared inboxes and sticky notes.
One tidy app, with logins, built around how you actually work.
The nerd name Full-stack web application
Marketing studio Demo planned
Describe a promotion or announcement once, and get matching drafts for email, social posts and your website, all in your brand’s voice. Nothing is published until a person has edited and approved it.
Writing the same news five different ways for five different places.
Consistent first drafts in minutes. You do the final polish.
The nerd name Multi-stage content pipeline with brand context
Research desk Demo planned
Ask it to look into a competitor, a supplier or a new market. It searches, reads, keeps track of where every fact came from, and hands back a short report you can check line by line.
Forty open tabs and a half-finished document.
A clear summary with every claim linked to its source.
The nerd name Research orchestration with source tracking
03 A day with the system
Here’s an ordinary Tuesday at a small wholesale business. Each piece does its bit, and they share what they know.
An illustrated example. Every business’s version is a little different.
“Could we get 24 cases of the oat milk and 10 of the cold brew for Friday?” The items, quantities and delivery date are picked out and matched to the customer’s account.
The customer’s agreed prices come from your price list. One item is low in stock, so that line is flagged rather than promised.
Order confirmation and invoice ready, with the stock note included. It sits in your approval queue. Nothing has been sent.
One tap. The email goes out, the order is logged, and your spreadsheet updates itself. Total time from you: about ten seconds.
Every line is read and compared with the purchase order. Line 4 is charged at the old price, so it’s highlighted for you instead of being filed.
“What do we do if a delivery arrives damaged?” Answered in seconds, with a link to the exact paragraph in your returns procedure.
A chart of sales by product and a few plain-English observations, built from your real figures, ready before you leave.
04 Ground rules
AI is powerful, and sometimes wrong. So every system is built around a few rules that don’t bend.
Sending emails, changing records, paying or ordering anything: those wait in your approval queue until you say yes.
Answers come with sources. Every action is logged, so you can always see what happened and why.
Each piece gets the smallest access that does the job. The assistant that answers questions can’t edit your accounts.
AI usage costs money, usually small amounts. You see the running total, and hard limits stop surprises.
Outputs are checked automatically, odd results are flagged, and errors are retried or brought to you, not quietly ignored.
Passwords and keys stay on the server, sensitive details are kept out of logs, and the AI only sees what a task requires.
05 How you’d get there
You don’t buy the whole thing up front. Each stage is useful on its own, quoted in advance, and the next one only happens if the last one earned its place.
We pick the most repetitive thing in your week and automate it, with you approving the important bits.
e.g. enquiries logged and replies drafted
Once the first is running smoothly, add another that shares what it knows. Now they start helping each other.
e.g. replies now use your real price list
The control room ties everything together: one queue of approvals, one history, one spending total.
e.g. your whole back office on one screen
Not sure where to begin? The AI consultation is a short session to find the realistic wins, and the ones that aren’t worth it.
06 Where things stand
I’m building a complete, working version of this system as a public demonstration, one piece at a time. It will live at its own address so you can try it yourself. Until then, here’s exactly what exists.
Demo status describes the public showcase. Each piece can be built for your business now, scoped and quoted up front.
07 Good questions
No, and most businesses never will. Start with the one that would save you the most time. Each piece is useful on its own; connecting them is a bonus, not a requirement.
It’s designed to take the repetitive parts off people’s plates, like copying details, chasing paperwork and answering the same question again. The judgement calls, relationships and interesting work stay with your team.
Usually the tools you already have: email, spreadsheets, your customer list, calendar or accounting software. If a tool has a way for other software to connect to it, it can normally be included. We check this together before anything is quoted.
There’s the one-off cost of building it, quoted up front, and then small ongoing costs for AI usage and hosting. The control room shows what’s being spent, and hard limits mean it can’t run away from you.
The system is built expecting that it sometimes will. Important actions wait for your approval, answers show their sources, numbers come straight from your data, and every step is logged so mistakes are easy to spot and fix.
Security is designed in from day one: access is limited to what each job needs, passwords and keys never sit in a web browser, sensitive details are kept out of logs, and anything consequential needs a human yes.
It depends on what we build first. A single focused piece is a much smaller project than a whole system, which is exactly why we start there. You get a clear scope and timeline before work begins.
08 Your move
Tell me about the task you’d most like to stop doing. I’ll tell you honestly whether a system would help, and what the first step would look like.