Use this pack with the person who runs a recurring process, one operator, and the sponsor who can change it. Allow an initial working session, then verify the baseline from records. A workshop estimate alone is insufficient for an investment forecast.
The output is a decision about what to investigate or change next. AI is one candidate intervention. The workflow selection guide explains the reasoning.
1. Define the work before comparing solutions
| Field | Record |
|---|---|
| Business problem | A specific delay, cost, error, service constraint or unmet demand |
| Unit of work | One receipt, invoice, application or other defined case; explain groups and duplicates |
| Start and finish | Observable intake event and accepted completion event |
| Owner and users | Process owner, operators, benefit owner and affected downstream team |
| Boundary | Entity, systems, record types, currencies, languages and excluded decisions |
| Consequence | What happens when a case is late, wrong or incorrectly discarded |
| Current alternative | Existing software, manual work, outsourcing or a planned process change |
Ask the operator to walk through one ordinary case and one recent difficult case using the records. Note where the written process differs from practice. Separate time spent actively working from elapsed waiting time: removing five minutes of typing may not fix a three-day approval delay.
2. Establish a baseline that accounts for every case
Choose a stated measurement window covering the relevant operating cycle. Month-end, seasonal peaks and unusual currencies may need separate observations. Record how cases were selected and why the window is useful; do not quietly exclude unfinished work.
Use this conservation check, with mutually exclusive categories:
Opening backlog + new unique cases = completed cases + authorised withdrawals + closing backlog.
Record duplicates at intake separately; they are not new unique work. A rejected case that still needs correction remains unresolved work, even if it has left the AI tool's queue. Reopened cases need a stated counting rule so one case is not counted twice as successful output.
| Observation | How to collect it |
|---|---|
| Intake and final disposition | Reconcile the case register to source-system totals |
| Active handling | Observe or time work across ordinary and difficult case types |
| Waiting and backlog age | Use timestamps and report the tail as well as a typical case |
| Review and rework | Include follow-up, corrections, disputed cases and work transferred elsewhere |
| Quality | Define material errors and check completed outputs independently |
| Peak demand | Record the busiest relevant periods and available trained capacity |
| Operating burden | Support, access administration, tool failures and manual recovery |
Use a stable case identifier, avoid unnecessary personal data, and preserve the source of each measurement. If timing requires self-report, label it as such and check a sample by observation.
3. Compare interventions against the same outcome
| Option | Evidence to obtain before preferring it |
|---|---|
| Repair the process | Whether better inputs, ownership or fewer approvals remove the main constraint |
| Configure an existing product | Results on the same cases, incremental licence cost and actual configuration limits |
| Buy a specialist service | What the supplier completes, what remains with staff and how exceptions are priced |
| Add bounded AI assistance | Whether interpreting variable inputs reduces total effort at acceptable final quality |
| Commission a custom workflow | A material unmet requirement, delivery estimate and someone able to maintain it |
| Defer | The condition that would change the decision, such as sufficient volume or data access |
A comparison fails if one option ends at extracted fields and another ends at an approved system-of-record entry. Write down the common completion standard before requesting estimates.
4. Resolve the next uncertainty
Record the assumption most likely to reverse the business case, its evidence source, a bounded way to test it, the person responsible, and the decision date. Examples include whether exceptions really take eight minutes, whether a vendor permits the required integration, or whether a delayed approval is the actual bottleneck.
End with one of four outputs: repair the process; run a specified comparison; proceed to a costed delivery proposal; or defer with a reconsideration condition. Attach the baseline and rejected alternatives. Carry a promising case into the economics workbook, including the unresolved work and reserve capacity found here.
Read our editorial principles for evidence, contribution and correction standards.