Why AI Costs More Before It Pays

Most AI business cases show benefits beginning soon after implementation and growing steadily from there. In my experience with large financial services programs, returns rarely follow that pattern. Costs usually rise first, and financial benefits tend to appear two to three years after the main investment. Understanding why helps leaders set realistic expectations and judge programs on the right evidence.

Why costs rise first

During the early phase of an AI program, an organization typically runs the old process and the new one in parallel. Existing staff remain in place while new technology costs begin. Existing controls stay in force while new ones are designed and approved. Integration work adds systems and interfaces, and new governance adds meetings and reporting. Much of the necessary work, such as cleaning data, building shared platforms, designing controls and training managers, produces no financial benefit on its own.

This does not mean the program is failing. It does mean that a business case promising net savings in the first year is likely to disappoint, and that leaders should plan for a period in which costs exceed measurable benefits.

How returns typically develop

Across the programs I have led or reviewed, returns have tended to develop in four stages.

1. Foundation. Investment goes into data, platforms, controls and training, and little value is visible outside the program team. Programs are most often cut at this stage.

2. Early operating signals. Pilots move into production and workflows begin to change. Non-financial measures improve first: more customers use digital channels, more transactions are processed without manual intervention, exception rates fall and satisfaction rises in the areas affected. Financial results generally remain flat, and a second group of programs is cut here.

3. Financial results. Usually two to three years after the main investment, operating improvements begin to show up in the financials. Customers who now use self-service contact the call center less, which lowers the cost to serve them, better service improves retention, and growth is absorbed without a proportional increase in cost.

4. Compounding. Shared platforms and data built in the first stage support additional uses at much lower cost, and returns from earlier work help fund new investment.

The transitions between these stages are often where the largest gains occur, but they are difficult to see in quarterly reporting.

One program I worked on illustrates the pattern. Leadership committed to a multi-year investment in digital services, including mobile apps, self-service and AI-supported servicing. Financial results were weak in the first year and only slightly better in the second, and at the eighteen-month review the financial data alone would have supported stopping. Leaders continued because the operating indicators were moving in the right direction: digital adoption and self-service completion were rising, and satisfaction was improving most in the customer segments with the highest adoption. About two years after the investment, call volumes fell as more customers resolved issues online, cost to serve declined and retention improved. By the third year, these improvements were reinforcing one another.

Judging programs on the right evidence

This pattern is not an argument for unlimited patience. Some programs do fail, and the argument that transformation takes time can be used to protect them. The task for leaders is to distinguish programs that are progressing through the early stages from those that are not progressing at all. Three practices help with this. First, specify in advance which operating indicators should move at each stage, and roughly when. Second, agree on stop criteria when the program is approved, before sunk costs make the decision harder. Third, report results against commitments plainly, showing what was promised, what was delivered and the size of the gap.

Two factors business cases often miss

Two further factors affect the economics of AI and are often left out of business cases.

The first is where the value appears. Early in one large program, the corporate center set cost reduction as the main objective for AI. A business unit in a fast-growing Asian market argued that this was the wrong objective for its situation, because it was still building market presence and cost was not its main constraint. After several months of discussion, the unit applied AI to growth instead, through better customer engagement, faster servicing and more relevant product recommendations. The resulting revenue gains substantially exceeded what the cost-reduction target would have delivered. In growing markets, one of the largest benefits of AI is the ability to absorb growth without adding cost, and standard business cases, which tend to count only reductions, rarely capture it.

The second is the cost of delay. Long approval cycles and repeated reviews have a real cost, but it is rarely calculated. In March 2020, I was leading a large transformation program at a regulated North American bank. In the first weeks of the pandemic, decisions that normally required six weeks of cross-functional review were made in two days, approvals that usually needed three signatures were given by a single person on a video call, and customer changes that had been in the backlog for a year were delivered in three weeks. Much of the delay we had treated as unavoidable turned out to depend on how decisions were organized. Estimating the monthly cost of waiting for a decision, and comparing it with the risk that a further review is meant to reduce, gives leaders a practical basis for deciding how much review is warranted.

Conclusion

AI programs tend to cost more before they pay, and their returns arrive in stages rather than in a straight line. Leaders who plan for this, define the indicators that should move at each stage, count all of the ways value is created and treat delay as a cost are better placed to keep strong programs funded and to stop weak ones early.