Generation applied where a draft saves real time.
Drafting, summarisation and classification embedded in the workflow that needed them — with a person reviewing before anything is committed.
Why generative pilots stay pilots
Most generative AI work stalls because it was deployed as a destination rather than a step. People must remember to visit a tool, paste context into it, and copy the result back. The saving is consumed by the ceremony, so usage decays to zero within a quarter.
- A licence was bought and adoption is a fraction of the seats.
- Output is useful but nobody trusts it enough to use unedited, and editing takes as long.
- The tool is separate from the system where the work actually happens.
- There is no agreement about what generated content is allowed to be used for.
What we build
Generation as a step inside an existing process — context supplied automatically, output landing where the work already lives, and a review gate before anything leaves the building.
- Generation embedded in the workflow rather than in a separate tool
- Context assembled automatically from the record being worked on
- Output constrained to a defined structure where downstream systems consume it
- A review step sized to the risk of the specific task
- Prompt and template versioning, so behaviour changes deliberately
- Usage and acceptance tracking, so you know whether drafts are actually kept
How it runs
The test for any candidate task: does a good first draft save meaningful time?
- 01Pick tasks that suit drafting
High-volume, low-variance writing and classification where a competent first draft is genuinely most of the work.
- 02Assemble the context
The record, its history and the relevant policy are gathered by the system rather than pasted by a person.
- 03Generate into structure
Where a downstream system consumes the output, generation is constrained to a schema rather than free prose.
- 04Review proportionately
An internal summary and an outbound customer letter get different review gates. Both get one.
- 05Measure acceptance
How often drafts are kept, edited or discarded is tracked — the honest measure of whether it is working.
What changes once it is running
What embedding generation in the workflow produces.
The blank page disappears
Work starts from a draft assembled from the actual record rather than from nothing.
Adoption stops depending on discipline
Because generation happens inside the existing process, nobody has to remember to use it.
Output stays reviewable
A person remains accountable for what is committed, with the review effort sized to the risk.
Value is measured honestly
Acceptance rate shows whether drafts are useful, rather than seat count showing whether licences were bought.
How an engagement is shaped
Narrow and embedded beats broad and optional.
Task selection
One to two weeks identifying which tasks genuinely benefit from drafting, and which are better served by a template or a rule.
Embed and pilot
Generation built into one workflow, with a defined review gate, measured on acceptance rate rather than usage.
Extend
Additional tasks added against the same pattern, with prompt and template governance in place.
Common questions
The things buyers ask before they commit. If yours is not here, it is a good first question for the assessment.
- Which model do you use?
- Whichever suits the task, the sensitivity of the data and your deployment constraints. Model choice is an implementation detail we expect to revisit; the workflow integration is the durable part.
- How do we stop generated content going out unchecked?
- A review gate is part of the build, not an optional add-on. Where content is customer-facing or contractual, generation produces a draft state that cannot be committed without a person.
- Is our data used to train someone else’s model?
- Not under the deployment patterns we use. Where a hosted provider is involved, this is a contractual and configuration question we settle explicitly during assessment rather than assuming.
Name the writing task your team repeats.
If a good first draft would save most of the effort, it is a candidate. If not, we will tell you.