On 24 September 2026, the Blue Cross Blue Shield Association published an analysis linking more intensive hospital coding to an estimated $942 million in additional costs for its member plans. Its concern about AI-assisted billing poses a broader management question: whose definition of success is an automation project serving?

What happened

BCBSA said more patients were being recorded as medically complex without corresponding evidence of changed care in its claims analysis. It associated that pattern with the spread of AI coding tools. This is an insurer association’s assessment of billing trends, not proof that every additional dollar was caused by AI or that individual hospitals acted improperly.

The distinction matters. A tool can make documentation more complete while the financial effects remain disputed. It is possible for different participants to evaluate the same improvement differently. The useful business discussion begins by identifying those participants, their incentives and the evidence needed to judge the outcome.

Why it matters

Consider a hypothetical sales team that introduces AI to produce more detailed proposals. The team celebrates a shorter preparation time and a higher quoted contract value. Delivery managers then discover that the proposals contain complex custom commitments, while customers struggle to compare the options. Local productivity has improved; the commercial process may have become harder to operate.

The mistake would be to measure only the activity the tool makes easier. Proposal volume tells you little about profitable work delivered. Similarly, a reduction in handling time can hide an increase in repeat contacts, escalations or disputed invoices. Those downstream effects belong in the original business case, even when another department carries them.

For an AI pilot, I would ask the receiving team to help define success before selecting a performance target. What will arrive in its queue? What constitutes an acceptable handover? Which costs could move across the departmental boundary? This gives the project a shared outcome rather than a collection of impressive local dashboards.

The bigger shift

AI makes it easier to optimise a measurable objective. That increases the importance of choosing the objective carefully. A target that rewards more recommendations, higher classifications or faster closures can steer a process towards activity that looks valuable in isolation.

The safeguard is not to abandon efficiency. It is to pair an efficiency measure with an outcome and a countermeasure. A customer service experiment might track resolution time, whether the issue stays resolved, and the number of cases reopened. A proposal experiment might track preparation effort, delivery margin and the customer’s understanding of the offer. These are examples, not universal scorecards.

This is why AI strategy must start with the business outcome. The decision about what to reward should precede the decision about which task to automate. Otherwise, a capable system can accelerate behaviour the organisation later has to undo.

My take

The strongest lesson here is about incentives, not a verdict on medical coding. A business should not assume that a more efficient process automatically creates more value for everyone affected by it.

Before expanding an AI pilot, review one completed case from beginning to end. Include the person who receives the output and, where appropriate, the customer who experiences the result. Compare what the project saved with what it required elsewhere: review, correction, negotiation or additional work.

Then decide whether the result deserves to scale. If the gain disappears when the whole process is counted, change the target before buying more capacity. Faster output is useful. A better outcome is the reason to invest.