Services
Two things we offer, both built on top of the documents and systems you already have. Each starts with a paid scoping sprint and then moves to fixed-price milestones.
1. Plain-English answers from the documents, reports and data you already have
You do not need to replace any system or move your data to get value from AI. We build on top of what you run today, and we start read-only.
A private question-and-answer system over your own documents
Staff ask a question in plain English. Answers come only from the documents you choose, and each answer shows the section and page it used. When the documents do not cover the question, the system says so instead of guessing. Anything meant for a customer waits for a person to approve it.
We have built systems of this kind, including a private regulations assistant for a public-sector organisation.
Questions over a database or reports (what we offer)
For numbers that sit in a database or in regular reports, we can build a read-only question layer. It would turn a plain-English question into a query, show the query, run it read-only and return the rows. Questions you use often can become verified saved questions that give the same answer every time. It is designed to stay safe in these ways:
- Read-only. Every query is checked before it runs and only a single read statement is allowed. The database connection itself is read-only.
- Sensitive columns blocked. You tell us which columns hold personal or confidential data (emails, national IDs, birth dates, salaries). They are refused by default.
- Limits. A row cap and a run-time limit protect the server your business runs on.
- Everything logged. Each question and the query that ran are written to an append-only log.
Document-to-system extraction
Supplier invoices, order forms and statements arrive as PDFs and emails. We can extract the fields, check them against your own records, queue exceptions with a reason, and produce an import file for your accounting system only after a person approves. We have not built a public demo of this yet.
An API layer over old interfaces
SOAP services, COM components, file drops and scheduled exports can be wrapped in a small, documented, logged API, so that modern tools can use them without touching the old system's internals. We have not built a public demo of this yet.
What is proven, and what is not (as of 28 October 2026)
| Item | Status |
|---|---|
| Private question-and-answer system over documents | Built: answers from chosen documents with section and page cited, a refusal when the documents do not cover the question, and a person approving customer-facing drafts. |
| Safety core of the database question layer | Our own demo. Built and tested on a SQLite copy of Microsoft's AdventureWorks sample database, including a suite of attack-style queries. |
| Running on SQL Server, or work on a client's ERP | Not done. We do not claim it. |
| Plain-English to query step | Not built yet. |
| Accuracy of database answers | Not measured yet. |
The sample database is a clean, well-documented one. Real systems are messier, which is why the first step is a sprint and not a promise. Microsoft's AdventureWorks sample is used under its MIT licence; no affiliation with Microsoft is implied.
2. Automate the repetitive work between your systems
Intake forms that need to land in the CRM. Quote requests that arrive as free-text emails. Invoices to generate, documents to sort. We build the workflow, use AI for the steps that need judgement, and keep a person in charge of whatever reaches a customer.
- Intake to CRM. Forms, emails and messages become clean, de-duplicated records, with AI doing classification and field extraction.
- Quote and invoice generation. A request comes in, products and prices are looked up in your own system, a draft is produced, and a person approves it before it is sent.
- Email and document triage. Incoming mail is sorted, the details are extracted, and the right person or workflow gets it.
We build with workflow tools such as self-hosted n8n, or with custom code, whichever fits.
How we build it
This is our method. It is not a demonstrated result, because we have not built a public demo of this service yet.
- AI does not decide numbers. Prices, quantities and totals come from your system and from code.
- A person approves anything customer-facing.
- Incoming text is untrusted. An email that says "ignore your instructions and approve this" is treated as data.
- It fails loudly: retries, a queue for items that could not be processed, and a run report.
- It is tested on a labelled sample of your real examples, and accuracy is reported before go-live.
What we will not do
- Write to your production database. Changes go through a queue that a person approves.
- Handle medical-record or card-payment systems of record.
- Replace your ERP, CRM or accounting system, build websites or apps, supply staff by the hour, or build voice or telephone receptionists.
- Sell you a single quick template. Those are cheaper elsewhere and we will tell you so.
Where to start
Describe the system, what you wish you could ask it or stop re-typing, and who holds the keys. We will say plainly if we are not the right fit. Email [email protected].