RickyPark LtdRickyPark Ltd · 20 September 2026
A useful AI pilot answers a specific business question: can this task be performed better at an acceptable level of quality, cost, and risk? Agree on the question and evaluation method before building, so the result supports a real decision.
1. Choose one workflow with an accountable owner
Start with repeatable work with reasonably clear inputs and outputs, such as classifying a defined enquiry type or extracting fields from one document category. Identify who uses the output, who handles exceptions, and who can accept the work. Avoid changing an entire department in the first pilot.
2. Establish a baseline before making comparisons
Record current handling time, volume, error types, rework, and staff effort. If reliable figures do not exist, observe a representative period first. Separate queueing delays from active handling time so you do not attribute every delay to the task you are automating.
3. Write down scope and boundaries
Specify the data, systems, languages, and actions requiring approval. Deliverables should include a usable prototype, evaluation results, and known limitations. Sending external messages, making payments, or updating consequential records should remain subject to authorised review.
4. Use representative evaluation examples
Include common cases, missing information, unexpected formats, duplicates, and questions the system should decline to answer. Keep final evaluation samples separate from development examples. Check both answers and sources for a knowledge assistant; evaluate document extraction field by field.
5. Calculate the full operating cost
Include model usage, platforms, human review, content updates, and support alongside development fees. If every output needs extensive revision, time savings may be limited. State workload and frequency assumptions explicitly, and evaluate higher-volume scenarios.
6. Agree when to expand, revise, or stop
Define quality thresholds, exception handling, owners, and a decision date before acceptance. Results may support expansion or show that data quality and scope need work. Keep a fallback process and establish who monitors quality, handles incidents, and approves changes after launch.
A preparation checklist
- One defined workflow and business owner
- A current baseline and representative test set
- Data boundaries, permissions, and review requirements
- Development and ongoing operating costs
- Written acceptance, stopping, and recovery criteria
HKT’s 2026 research identifies cost and return on investment as important deployment considerations. This guide is our practical planning approach informed by that finding, not a guarantee of results for an individual project. Read the research source
Explore the related service