Updated October 14, 2026
An AI shortlist becomes useful when it supports a resource decision. Leaders need to know which problem merits investment, what the evidence supports and what would change the recommendation. A long list of attractive ideas provides little guidance if every entry assumes rapid adoption, clean data and inexpensive integration.
Start with a common comparison period and a consistent definition of value. Separate observed facts from estimates. Then compare the proposed workflow with the current process and at least one simpler alternative.
Establish the baseline before the business case
For each candidate, document the eligible workload, current effort, error or rework rate, service outcome and existing cost. Measure enough cases to expose meaningful variation. Keep peak periods and unusual cases visible instead of burying them in an average.
Identify the business owner and the operational change required to realise a benefit. Faster drafting has limited value if another step remains the bottleneck. Time released has limited financial meaning until someone explains how it will be used or which expenditure can actually change.
The following examples are hypothetical. Their figures illustrate a comparison method; they are not client results, benchmarks or forecasts.
Candidate 1: drafting routine customer replies
Assume a support team handles 1,200 eligible requests per month. A sample indicates nine minutes of active work per request, including checks and corrections. An assistant would retrieve approved information and prepare a response for a person to review and send.
The testable hypothesis is that total active work could fall to six to eight minutes while meeting the agreed quality standard. That represents a hypothetical 20 to 60 hours of monthly capacity before extra support and maintenance. It is not yet evidence of cash savings.
Cost the knowledge-base cleanup, retrieval setup, review time, software usage and ongoing content ownership. Evaluate factual accuracy, missing information and inappropriate disclosures alongside speed. Compare the assistant with better templates and search.
Recommendation: consider a bounded pilot if the knowledge sources and review process are ready. The next evidence should be timed, representative cases, including requests for which the assistant should decline to draft.
Candidate 2: extracting information from supplier documents
Assume 700 documents arrive monthly, with six minutes of manual entry per document. The proposed workflow extracts selected fields into a review queue. In this example, the main uncertainty is how much time reviewers would spend correcting different document formats.
A headline estimate based on eliminating all entry time would be misleading. Measure field-level errors, complete-record acceptance, correction effort and downstream rework. Include duplicate documents, poor scans and missing fields in the test set.
Integration and exception handling may cost more than extraction itself. A standard supplier template or an existing system feature could be sufficient for part of the workload.
Recommendation: run a small offline comparison before committing to integration. Defer a live pilot if the expected benefit disappears once correction time is included. Record what would justify reconsideration, such as a narrower document set or better source quality.
Candidate 3: forecasting customer churn
Assume a commercial team wants to identify accounts needing attention. The available records do not yet consistently define churn, and no team owns a standard retention intervention.
The potential revenue benefit is uncertain because a prediction alone does not establish that an intervention changes the outcome. The business case needs a clear outcome definition, usable historical data, an agreed intervention and a credible evaluation design. It also needs the cost of acting on false alarms.
Recommendation: defer model development. First agree the business definition, assess the data and test whether the proposed intervention is operationally viable. Set a review date and an owner for those prerequisites so “defer” has a practical meaning.
Make uncertainty part of the decision
Use ranges for value and cost, with the assumptions shown. Distinguish measured evidence, small-sample observations and expert estimates. Check whether the ranking changes when adoption is slower, review takes longer or integration becomes more expensive.
A narrow, well-supported opportunity may be preferable to a larger speculative one. Equally, a discovery exercise can be justified when it resolves an important uncertainty at modest cost. State which decision the research will enable before funding it.
Keep a one-page decision record
For every candidate, capture:
Problem, affected users and accountable owner.
Baseline, data source and confidence in the evidence.
Proposed workflow and simpler alternatives.
Benefit range, realisation plan and full cost range.
Main risks, controls and unresolved dependencies.
Next experiment, decision date and spending boundary.
Recommendation: proceed, investigate, defer or stop.
Revisit the record when evidence changes. The goal is a portfolio the organisation can explain and deliver, with clear reasons for what it chooses to leave alone.
Explore Daly Ventures’ AI Strategy & Readiness approach to build an evidence-led opportunity shortlist and a practical next step.
