Services/AI integration
AI features with the cost worked out first.
We put language models into products that take money. Before the first prompt we answer three questions: what it costs per call, what it does with your documents, and what the screen shows when it’s wrong.
Tender_Riverside_Ph2.pdf
42 pages · pp. 18–21 sent
| Excavation | 42 m³ | £1,890 | high |
| Blockwork | 118 m² | £7,316 | high |
| Steel lintels | 9 nr | £1,035 | check |
| Roof covering | 64 m² | £3,712 | check |
How we can help
Where AI saves time in your product.
We find where a model saves real time in your process, build it inside the software your people already use, and measure whether it works.
- Discovery: where AI saves real time in your process, and where it doesn’t
- Build: drafting, document reading, search and triage, with review screens
- Evaluation: test sets from your own documents, so quality is measured
- Running: logging, cost tracking and switching providers when prices change
Before we build
Before we write a prompt.
Cost
What does it cost per call?
Tokens in, tokens out, times the rate, times how often. If the feature doesn’t pay for itself at your volume, we say so before building it.
Data
What does it do with your documents?
Which records it can see, which provider they go to, whether they are stored, and who can audit it. Decided per feature and written down.
Failure
What does the screen show when it’s wrong?
Every output lands where a person can check it, correct it and see why it was suggested. Confidence is shown, not hidden.
What we build with it
AI features we build.
Most useful AI features are one of these, inside software people already use. None of them is a chatbot bolted to the side.
Read
Document reading
Tenders, invoices, delivery notes and forms turned into structured data, with each value linked to where it came from.
e.g. a bill of quantities into a priced scheduleDraft
Drafting with review
Quotes, replies and reports drafted from your own records, then checked and sent by a person.
e.g. a reply to a customer complaintSearch
Search over your documents
Answers from your manuals, policies and past jobs, each with a link to the page it came from.
e.g. “what did we charge for a similar job?”Sort
Triage and routing
Incoming email, enquiries and tickets sorted, prioritised and assigned by a small, cheap model.
e.g. enquiries sorted by trade and postcodeCapture
Notes to records
Voice notes and messy text turned into structured diary entries, CRM updates or job sheets.
e.g. a voice note from site into a diary entryTailor
Content for one case
Material generated for one person or one case from structured inputs, then checked in code.
e.g. CurioKids lesson content for each childArchitecture
What surrounds the model.
The model is one step of six. The other five decide whether the feature is safe, affordable and useful.
- 01
Request
A person asks, or something arrives: a new tender, an email.
- 02
Permissions
Only the records this user may see are retrieved.
- 03
Model call
The smallest model that does the job, at a known cost.
- 04
Checks
Format, totals and policy checked in code.
- 05
Review
Low-confidence output goes to a person first.
- 06
Action and log
The result is saved with what produced it.
Calculator
What will your feature cost?
Start from a pattern, then enter your own volumes and your provider’s current rates.
A thousand words of English is roughly 1,300 tokens.
Our line
What we won’t ship.
An AI feature is finished when it is safe to rely on, not when the demo works.
- An assistant that can’t say “I don’t know”
- A feature whose monthly cost nobody worked out
- Customer data sent to a provider you haven’t approved
- Output that skips a person where money or safety is involved
Build log
More on AI
What an AI feature costs per call, and why we work it out first
A model call has a price you can calculate before you build anything. Here is the sum, where the money goes, and three ways to spend less.
How to design AI features for when the answer is wrong
Every AI feature will sometimes be wrong. Design the screen for that day, with sources shown, outputs as editable drafts, facts checked in code and actions that can be undone.
AI agents in business software: what they can and can’t do reliably
An AI agent is a model that chooses tools in a loop until a task is done. Narrow agents with few tools and an approval step work well today. Open-ended ones still fail in costly ways.