Choose One Business Process Before Funding the Next AI Pilot

An AI pilot can shorten one task while the customer’s problem still waits in three other queues. Measure the whole journey before deciding the pilot worked.

In the September 21 WIRED–McKinsey interview, Dan Swan puts the starting question plainly: “What is the business problem I need to solve, and how can AI help solve that business problem?” His recommendation is to concentrate effort on a small number of end-to-end business domains.

Read the interview

For a leadership team, that means naming the outcome, the owner and the boundaries before choosing a model. “Supplier onboarding” is still too loose.

“From a complete supplier application to an approved, rejected or escalated decision” gives the team something it can trace.

Consider a hypothetical procurement team. An assistant drafts a supplier summary in minutes. Legal still waits for missing contract terms. Security repeats a questionnaire because it cannot see the earlier evidence. Procurement chases signatures. A faster summary has not necessarily shortened onboarding.

Map one real case from arrival to decision. Record each handoff, the evidence it needs, who can approve it and how often work comes back. Put waiting time beside processing time. That makes it harder to mistake a better demonstration for a better service.

There is a useful reporting example at DBS. Its 2025 annual-report letter says data analytics and AI/ML delivered approximately SGD 1 billion in economic value, across more than 430 use cases and over 2,000 models. That is a company-reported portfolio figure, not an independently verified return on a single generative-AI project. DBS chairman and CEO letter.

The practical lesson is to connect a use case to an economic measure. Track completed decisions, rework, exceptions and total cost. Login counts can explain adoption; they cannot establish business value.

Hypothetical decision for the next operating meeting: “Should the COO approve a six-week, read-only supplier-onboarding pilot by Friday, provided procurement supplies a baseline, security approves the data boundary and legal retains every exception decision?”

Write down what would change the answer. Missing source permissions should delay the start. Rising rework should block expansion. A lower average completion time should not excuse worse performance on complex or high-risk cases.

Use a one-page scorecard with five rows: median and 90th-percentile completion time; reviewer minutes per case; returned or reopened cases; missed material risks; and cost per completed case. Keep the case types comparable and publish the denominator for every rate.

Calculate capacity value separately from cash savings. If a hypothetical team handles 200 cases a month and saves 30 minutes each, that releases 100 hours. At an illustrative loaded cost of $60 an hour, the capacity equivalent is $6,000. It becomes cash savings only if spending actually falls; otherwise identify what useful work those hours fund.

Deduct integration, model usage, human review, training and ongoing controls. Avoid counting the same benefit twice as both labor savings and additional output.

A local pilot has not yet proved durability, safety on rare cases or performance at peak volume. The business owner retains the expansion decision, with security and legal able to stop activity within their remit.

Run a 30/60/90-day sequence. By day 30, map the process, establish the baseline and agree the stop rules. By day 60, test a bounded workflow with a comparison group where feasible, or a documented before-and-after comparison with its limitations. By day 90, decide whether to expand, redesign or stop.

Choose the success thresholds before seeing results. Appoint one operating owner who can resolve cross-team handoffs. If nobody owns the full journey, settle that before buying another tool. This is a proposed management method, not a guarantee of a 90-day return.

Which customer or employee outcome remains slow even though several of its individual tasks are already “AI-enabled”?

WisdomTwin builds sovereign, governed AI Judgment Twins that let regulated enterprises make high-consequence decisions without waiting for the next meeting.

Bring one delayed decision and its current approval path to a WisdomTwin demonstration. Use the scorecard above to judge fit; no result in this article establishes a WisdomTwin customer outcome.

Founder disclosure: Roman Bodnarchuk is Co-Founder and CEO of WisdomTwin.ai. Operational examples are hypothetical; cited research and company reports are identified separately. Demonstrations use synthetic scenarios.

Book a 20-minute call: https://calendly.com/romanbodnarchuk/20min