Service
AI Assistants Built Around Your Business
A generic chatbot knows the internet. A business assistant knows your prices, your policies, your customers and your systems — and can act in them.
This isn't a ChatGPT prompt
There's a large gap between typing a clever prompt into a chat window and having an assistant embedded in your operation. The first depends on a person remembering to use it and pasting in the right context. The second already has the context, runs where the work happens, and leaves a record of what it did.
A business AI assistant is built on top of your own material: your documents, your product and price information, your past correspondence, your processes. It's given a defined job and defined permissions, and it's connected to the systems it needs to read from and write to.
That's a software project rather than a subscription, and it's the difference between AI being an interesting experiment and AI being part of how the business runs.
What an assistant can do
Access company knowledge
Answer from your handbooks, policies, specifications and past jobs rather than from general internet knowledge.
Answer staff questions
New starters and busy staff get correct answers immediately instead of interrupting the person who knows.
Draft emails and documents
Quotes, replies, follow-ups and summaries prepared with your details filled in, ready for review.
Respond to customer enquiries
Handle the common questions properly and escalate the ones that genuinely need a person.
Search internal information
Find the right file, job record or historical decision without knowing where it was filed.
Process documents
Read incoming PDFs and attachments, extract what matters and route it onward.
Create tasks and records
Turn a conversation or an email into the CRM entry and follow-up task that would otherwise be forgotten.
Interact with your systems
Look up availability, check an order, update a record — through proper, permissioned API access.
A real example: Claude Email
Claude Email is an application I built and run myself. With explicit Google authorisation it reads incoming mail, identifies invoices and extracts their details, and prepares draft replies for review — including drafting from the correct business address, which required getting the Gmail OAuth configuration genuinely right rather than approximately right.
It's a useful example because it shows the shape of this work honestly. The interesting parts weren't the AI. They were the permission scopes, the review step, the handling of edge cases, and making sure a draft always lands somewhere a human sees it before a customer does.
I've written up the Gmail authentication work in full, including the part the documentation glosses over.
How I keep assistants trustworthy
- Grounded in your own documents and data, not general guesswork
- Narrow, explicit permissions on every system it can touch
- A human review step wherever the output reaches a customer
- Clear escalation rules for anything outside its remit
- Logging of what was asked, retrieved and produced
- An off switch that doesn't require a developer
On performance claims
I deliberately don't publish percentage figures for time saved or enquiries handled. Those numbers vary enormously between businesses, and quoting someone else's would tell you nothing useful about yours. During discovery I'll estimate the realistic saving for your specific volumes, and I'd rather that estimate be conservative and correct.
FAQ