Drafts from data, not from nothing
AI drafting is genuinely useful when it transforms material you already have: an enquiry into a reply, records into a report, a specification into proposal copy.
It is much weaker asked to produce substance from nothing, and the result is fluent text with no information in it — which readers and search engines both recognise.
What we build
- Context assembly — pull the customer record, the history, the relevant policy
- Generation against your approved language and tone, not generic output
- Editing in place, so the person adjusts rather than rewrites
- One-click send from the same screen, with the record updated automatically
- Edit capture, so consistent corrections improve the prompt over time
The person sending it is accountable
Automate the assembly, never the accountability. Whoever presses send owns the content, and the interface should make that obvious rather than encouraging one-click approval of unread text.
For anything carrying professional liability — advice, clinical, legal, financial — we build in an explicit review step and say so in the specification.
Where it saves the most
- Support replies where the answer exists in documentation
- Client reports where the narrative changes but the structure does not
- Proposals assembled from previous ones and approved boilerplate
- Meeting summaries and follow-up actions from a transcript
- Internal updates that nobody enjoys writing
Measure the edit rate
Track how much of each draft survives. A high edit rate means the drafts are not useful; a zero edit rate probably means nobody is reading them, which is worse.
Somewhere in between is healthy, and the trend should improve as the system learns your corrections.