01 · Decision guide
Find the operating problem before choosing the tool
Ask the team to show you the work: where does information arrive, who checks it, which systems are used, and what happens when something is missing? An inbox full of requests may need clearer ownership before it needs an agent. A stable rules-based process may be better served by ordinary automation. AI becomes worth testing when interpreting varied documents, retrieving knowledge or drafting responses is a meaningful part of the work.
- Operations: assemble an exception brief from approved sources so a manager can investigate.
- Customer support: draft responses grounded in an approved knowledge base for an agent to review.
- HR: organise policy questions and prepare answers for HR review, without automating employment decisions.
02 · Decision guide
Build a baseline the business can recognise
Record volume, handling time, turnaround, correction rates and review effort for a representative sample. Include difficult cases, not just clean examples. Agree who judges correctness and what an unacceptable error looks like. The baseline is the reference for deciding whether the change is worth keeping.
- Define one primary outcome, such as reduced handling effort without a rise in correction rate.
- Set a quality floor and a clear route for escalation to a person.
- Include integration, training, review, licensing and ongoing support in the cost estimate.
03 · Decision guide
Run a bounded pilot
Limit the pilot to a named team, a defined workflow and approved data. Keep write access and customer-facing actions behind explicit review until the acceptance tests justify more autonomy. Record failures and corrections so the team can see what the system actually does.
- Agree the inputs, outputs, users and excluded actions before building.
- Test outdated information, conflicting records and questions the system cannot answer.
- Give users a way to flag problems and return to the existing process.
04 · Decision guide
Decide whether to scale
Compare the pilot with the baseline across quality, time and total effort. A faster draft is not a saving if checking it takes longer. Scale only when there is a process owner, support plan and evidence that users can operate it reliably. Pause or redesign if the value depends on assumptions the pilot did not prove.
- Illustrative calculation: 400 tasks a month × 10 minutes saved = about 67 hours of gross capacity. Deduct extra review and support time; this is not automatically cash saved.
- For an India delivery team, agree ownership, handover, access boundaries and quality measures as part of the same transformation.
Bring this to the conversation
A brief that makes discovery useful.
Prepare the following at a high level. No confidential records are needed for an initial discussion.
- Workflow and monthly volume
- Current tools and approved data sources
- Current handling time and quality problems
- Business owner and pilot users
- Success measure, constraints and desired timing
Experience and product fit
Explore the work behind the approach.
Explore ACV’s workflow engineering and delivery experience in the Gradient customer story. For GCC leadership workflows, explore Conduit. An AI engagement begins with its own baseline, scope and acceptance measures.