Review an AI draft before it leaves your team
Follow one bounded AI task through drafting, human checks, exception cases and handover. Adjust the time assumptions without sending any data to an AI model.
Illustrative walkthrough. Sample data, not a customer result.
1. Bound the task
Who is involved: Together
Starting point: A repeated drafting task, approved reference material and examples of a useful response.
What happens: Agree one task, up to three action steps, two supported tools and one model or service. Define what the assistant may use and when it must stop.
What you receive: A clear pilot boundary and a human approval point before any external or consequential action.
Sample pilot boundary
Item · Agreed example
Task · Prepare an internal enquiry summary and reply draft
Allowed input · Submitted enquiry and approved service notes
Human decision · Check accuracy and approve any external reply
Stop condition · Missing evidence, unclear request or sensitive judgement
2. Inspect a sample draft
Who is involved: Labs
Starting point: A fictional request: “Can you connect our enquiry form to our existing customer system?”
What happens: Illustrate how the pilot would summarise the request and draft a response using approved material. Nothing on this page calls an AI service or sends a message.
What you receive: A draft for a person to check, with unresolved details called out instead of invented.
Illustrative draft review
Part · Example output · Reviewer check
Summary · Connect an existing enquiry form and customer system · Does this match the request?
Missing detail · Which customer system is in use? · Ask before promising compatibility
Draft next step · Confirm the system and discuss the required fields · Approve wording before sending
3. Try the edge cases
Who is involved: Together
Starting point: Up to ten agreed evaluation cases, including failures and escalation.
What happens: Inspect the draft and the human checkpoint for each case. Review unresolved cases before deciding whether to launch; this example is a plan, not an accuracy benchmark.
What you receive: A documented view of where the pilot is useful, where it needs revision and when a person must take over.
Sample evaluation cases
Case · Expected behaviour · Human decision
Clear request · Prepare a grounded draft · Approve or edit
Missing system name · Ask for clarification · Request the missing detail
Unsupported claim · Avoid inventing an answer · Check the source or decline
Tool unavailable · Notify and stop · Handle the enquiry manually
Sensitive decision · Escalate without acting · Use the appropriate human process
4. Launch with controls
Who is involved: Together
Starting point: The reviewed pilot and agreed changes.
What happens: Set usage limits and failure notifications, apply the included revision and hand over the workflow. Review it two weeks after launch.
What you receive: One bounded pilot with human review. The delivery target is one to two working weeks once inputs and access are ready.
Handover
You receive the pilot configuration, evaluation notes and an operating handover, followed by a review two weeks after launch. Model usage and third-party subscriptions are separate. This is not an autonomous decision system or a custom-trained model.
Compare drafting and review time
Illustrative assumptions: 40 Drafts per week; 8 minutes currently and 5 minutes in the proposed process per item.
5.3 hours currently; 3.3 hours in the proposed process per week.
Invented planning inputs include reading, drafting and the human review in both processes. Setup, evaluation, ongoing monitoring and tool costs are excluded. Poor drafts may take longer: increase the proposed minutes to explore that case.
This comparison is an estimate, not a promised saving.