A ranking is a policy decision wearing a number
In banking, relationship profitability is one of the most requested and most contested analyses I have seen. A private banking client may hold deposits, a mortgage, a credit line and an investment account across several entities. Whether that relationship looks profitable depends on how you group the accounts, which costs you allocate and over what period. Change those choices and the ranking changes. Ask an AI assistant to rank customers without settling them and you will get a confident list built on assumptions nobody approved.
Start with the decision. Customer profitability can inform account planning, pricing or service design, and each may need a different view. A list built to find support-intensive accounts should not quietly become a list of customers to exit. Name the decision and the person authorized to make it.
Agree what profitability means for this decision
Gross margin, contribution margin and fully allocated profit are different measures, and dashboard labels rarely reveal which costs are inside. An assistant should not infer policy from a column name or pick whichever calculation is easiest.
Fix the customer boundary and the period. Parents, subsidiaries, households and billing accounts all create different groupings, and an account that is profitable at one level can look unprofitable at another. This is a master data question as much as a finance one, and the business owner should approve the hierarchy the pilot uses.
Ask finance which allocation methodology governs the intended decision and whether every required input is available. Relationship managers may know context that the model cannot see, especially when several accounts belong to one relationship. Keep their judgment distinct from the calculation. The assistant should explain the approved view and its gaps so the account team can decide how much weight to give it.
Make cost allocation explicit
Direct costs are usually straightforward. Shared support, servicing, delivery and overhead require allocation rules that can materially change the ranking. Record the approved basis, its limits and the owner who can revise it. The assistant applies and explains policy; it never invents it.
Use two customers who consume shared services differently, and show how the result moves under alternative legitimate allocations. Reviewers will see that much of the disagreement is about policy, not arithmetic. Label estimates as estimates: service effort inferred from incomplete time records can support a discussion but should not appear as exact fact.
Handle timing, one-time items and access
Revenue and related costs are often recorded at different times, so a short window can make a customer look unusually good or bad. Ask finance whether a rolling view or specific adjustments fit the decision.
Separate one-time implementation work, credits and exceptional service incidents from the recurring pattern instead of burying them in an average. A helpful response points to the evidence for each.
Profitability summaries reveal pricing, contract terms and service costs. Evaluate access against the actual account responsibilities of pilot users. A ranking is not a record export, but it still needs authorization.
Before rollout, put the cost policy itself in front of the CFO for explicit approval. In my experience that single signature changes the conversation: when a business leader disputes a ranking, the discussion moves to whether the policy should change, which is a legitimate management question, rather than whether the AI got the math wrong. That is a far healthier place for an executive team to argue, and it keeps the assistant's role clear.
Test the edges before anyone acts
Have reviewers inspect customers near decision boundaries, where small assumption changes move rankings substantially. That sensitivity discussion is often more valuable than the ranking itself. Record disagreements by cause: identity, timing, missing data, allocation or interpretation. Those categories guide remediation far better than a single pass rate.
Include the account owner before any pricing or service-level decision. They know contractual commitments and relationship history the calculation cannot capture. Publish the approved cost policy beside the analysis so nobody has to reverse-engineer it from the ranking.
When I present this kind of analysis to a CFO or a board, I frame the claim narrowly: this is a defined profitability view, prepared and explained consistently. It is not an objective value for every customer. Commercial decisions still need strategy and accountable judgment, and good AI makes that judgment better informed, not optional.