Drafting assistants
First drafts of the documents your team writes repeatedly, in your language and using your prior work.
AI intelligence layer
Your data, put to work, with judgment you can check.
For businesses that want AI in the workflow without handing over decisions.
AI nobody can audit is AI nobody uses twice.
A tool that produces an answer without showing its working gets tried once and quietly abandoned.
Trust decides adoption
People don’t stop using an assistant because it’s wrong occasionally. They stop because they can’t tell when.
The judgement is the bottleneck
The same call gets made dozens of times a week by whoever is available, with no record of why.
The data sits unused
Most businesses already hold what they’d need. It stays in systems nobody queries.
What you get
The things that will exist
Lead qualification and scoring
Enquiries rated against criteria you set, with the reasoning shown next to the score.
Summarisation
Calls, threads and documents reduced to what someone needs before they act.
Flagging for human review
Anything uncertain routed to a person by design, rather than guessed at confidently.
Knowledge base
Your own material made answerable, so the system cites your documents instead of inventing an answer.
Guardrails and review logs
A record of what was asked, what came back, and who accepted it. The part that makes the rest auditable.
Where we start
What we’d look at first
Before any package gets proposed. None of this requires you to have tidied anything up first.
The judgement made most often
The recurring decision your team makes dozens of times a week, and the criteria behind it.
What data you already hold
The documents, records and history that exist, and which of it is clean enough to rely on.
What must stay human
The decisions that will never be handed over. Agreed at the start so the boundary is a design input.
Sequence
Your first 30, 60 and 90 days
First 30 days
Pick one decision
Choose a single high-frequency task, write down how a good answer is judged, and assemble the material it needs.
Next 30
Build it with the human in
The assistant drafts, a person approves. Every output is logged with its reasoning so accuracy is measurable rather than felt.
Then
Widen only where it earned it
Tasks clearing the accuracy bar get more autonomy. The rest stay assisted. Nothing loses its review log.
A sequence, not a set of deadlines. What moves first depends on what we find in the first two weeks.
The split
What we handle and what we need from you
Published before the first call, so nobody discovers halfway through that the work needed more of your week than anyone said.
| Minchin handles |
|
|---|---|
| You provide |
|
| Shared |
|
How this gets judged
Four numbers, agreed before we build
Tasks assisted
How often the system is actually reached for.
Accepted without edit
The real quality measure: output used as-is.
Escalated to human
Correct behaviour, not failure. A rate of zero means the guardrails are too loose.
Time saved
Measured against the baseline, per task, not claimed in aggregate.
Where a buyer touches several channels before they enquire, attribution gets approximate. We’ll tell you when a number is an estimate rather than presenting it as fact.
Fit
Who this is for
This is for you if…
- You hold data you’ve never actually used
- Your team makes the same judgement call dozens of times a week
- You want AI in the workflow but need to see its reasoning
- Drafting and summarising take real hours
- You already have systems worth connecting
This is NOT for you if…
- You want decisions made without a human in the loop
- Your data is scattered and nobody owns it yet
- You’re looking for a chatbot on the website
Before you book
Questions we get asked first
What does this cost?
A fixed build fee, plus model usage billed to your own provider account at cost. Usage is the smaller number by a wide margin, and you can see it without going through us.
Where does our data go?
Into the provider account you own, under terms you can read. We don’t train anything on your data, and we’ll tell you exactly which documents a given feature can reach.
How do we know it isn’t making things up?
Answers cite the source document they came from, and anything the system can’t ground gets flagged rather than guessed. The review log shows what was asked and what came back.
What if it gets something wrong?
A person approves before anything acts. That’s the design, not a precaution. Errors that reach the log get corrected in the criteria, so the same mistake doesn’t recur.
Do we need clean data first?
Not perfect, but owned. If nobody can say what a record means or which copy is current, that gets fixed before any model sees it. That’s the exclusion above.
Which model do you use?
Whichever fits the task and your constraints. That’s a decision we make with you, and it’s revisitable. The system is built so the model can be changed without rebuilding around it.
Is this just a chatbot?
No. It’s judgement inside the workflow your team already uses, where the work happens. If a website chatbot is what you want, we’re not the right people for it.
Show us the judgement call your team makes fifty times a week.
One call, no deck. We’ll tell you what we’d build first and what it would take.