Every AI use case wants more money. Which ones deserve it?

AI investment is increasing quickly. In September 2026, Gartner reported that 85% of functional leaders in its survey, planned to increase AI spending during the year. The research, covered 1,303 respondents from organisations with at least US$50 million in annual revenue. Roughly 11% did not know what their function had spent on AI in 2025.

That creates an obvious need for better financial visibility. Organisations need to know what they are spending, which activities are consuming resources and who is accountable for that expenditure. FinOps provides much of that evidence base by making spend, ownership and efficiency visible. The portfolio allocation decision remains a leadership judgement. A consolidated AI budget can tell a CFO how much the enterprise expects to spend.

It cannot tell them whether the next increment of funding is going to the right place. Consider two AI initiatives. The first automates a high-volume administrative process. Its costs are relatively predictable, and the business has evidence that it is improving operating efficiency.

The second supports a more complex customer-facing process. It is more expensive to run, but further investment could improve retention, increase profitable revenue or remove a significant service constraint. Both may have positive business cases. That still does not tell management which one should receive the next R5 million.

The organisation needs to understand what additional contribution each investment could make, how confident it is in that outcome, what resources will be required and what it gives up by choosing one over the other. That is the part a project-by-project business case can easily miss.

When an AI capability is first proposed, management has limited evidence. The business case is largely a hypothesis. It contains assumptions about adoption, cost, performance, demand and the value the capability might create. Some of those assumptions will prove correct.

Others will not. Once the capability is operating, the organisation knows more. It can see what customers and employees actually do with it. It can observe real consumption.

It can measure the quality of the outcomes. It can see where human intervention is still required, how the operating costs behave and whether the anticipated benefits are appearing. That evidence should change the next funding decision. Imagine an AI agent supporting a contact centre.

The initial deployment handles straightforward customer enquiries effectively. It reduces pressure on employees, and the business has evidence that the economics of those interactions are improving. The sponsor now wants to extend the agent to more complex requests.

Those requests require more reasoning, additional integrations, greater human oversight and stronger operational controls. The next group of transactions may therefore have very different economics from the first. A solution can perform extremely well across the easiest part of a workload and become progressively less attractive as it expands into more difficult work.

AI scale is not necessarily economically linear. That means successful pilot economics should not simply be extrapolated across the entire addressable process. The next investment needs to earn its own case. What additional value will it create?

What additional cost and risk does it introduce? And is that contribution still attractive compared with what else the organisation could do with the same resources?

1. Make every new tranche earn its way

An initiative that performed well last year does not automatically deserve more money this year. Equally, an initiative that is underperforming should not continue receiving resources simply because the organisation has already invested heavily in it. Past expenditure matters when evaluating whether the original investment succeeded.

It should not create an entitlement to future funding. Suppose an organisation has spent R20 million building an AI capability and now needs another R5 million to expand it. The relevant decision is not how to protect the R20 million already spent. It is what the business can reasonably expect to gain from committing the next R5 million from this point forward.

That sounds obvious. In practice, it can be difficult. A sponsor has invested time and credibility. Employees may already have adopted the solution.

The technology team has built integrations. Stopping or reducing investment can feel like admitting that the original decision was wrong. But those are two different questions. An initiative may have been completely sensible when it was approved and still become a poor candidate for further investment once better information becomes available.

A good decision made six months ago does not have to remain the right decision today. The reverse is also true. A cautious pilot may reveal much stronger economics than expected and justify substantially more investment. The purpose of learning is to change what management does next.

2. Compare contribution, not AI activity

Different AI investments create value in very different ways. A fraud capability may reduce expected losses. A customer-service agent may lower cost-to-serve, improve customer retention or release operational capacity. A sales capability may create incremental revenue, although that revenue still needs to be considered alongside the cost of acquiring and serving the additional business.

A shared platform may reduce duplicated engineering and make it easier for multiple business units to deploy new capabilities. A governance or control investment may reduce material operational or regulatory risk. These cannot be compared credibly using tokens consumed, users onboarded, prompts submitted or tasks automated.

Even ROI needs care when the timing, certainty and nature of the benefits are different. The point is not to force unlike investments into one artificial measure. It is to make their economic contribution and underlying assumptions visible enough for management to make a real trade-off.

That means distinguishing between what has already been realised and what is still expected. It means understanding what the capability costs to operate and what further investment is required to scale it. And it means being clear about uncertainty. R10 million of value that has already been demonstrated is not economically identical to R10 million forecast from an experimental capability whose adoption, cost and performance are still uncertain.

The uncertain initiative may still deserve investment. Its potential may be considerable. But management should know what it is funding. Sometimes the next tranche buys value.

Sometimes it buys evidence. That distinction matters. If an initiative has significant potential but important assumptions remain unresolved, the right decision may be to fund the smallest sensible next step that answers the most important questions. Not everything needs to be either fully scaled or stopped.

But each additional investment should buy something meaningful: more value, or better information about whether the value is real.

Sometimes the next tranche buys value. Sometimes it buys evidence. Each additional investment should buy something meaningful.

3. Make reallocation normal

The resources AI initiatives compete for extend well beyond money. They may require scarce engineers, data specialists, architecture capacity, risk and governance expertise, business SMEs, change capacity and executive attention. Sometimes the binding constraint on an AI portfolio will not be the budget.

It will be the organisation's ability to execute. That means an attractive initiative can still be the wrong investment. Imagine two projects each require R5 million and access to the same scarce engineering team. The first has a credible case for creating R7 million of value.

Looked at by itself, it appears entirely sensible. The second could address a more important constraint and create materially greater value from the same resources. If funding the first means the second cannot proceed, management has to consider the opportunity cost of that decision.

Positive ROI is therefore not enough. The question is whether this is the best available use of the resources being committed, given the organisation's priorities and the other credible opportunities competing for them. This is where portfolio discipline becomes important.

Management needs to become comfortable redirecting investment when the evidence changes. That should not be treated as failure. Sometimes an experiment has done exactly what it was supposed to do by showing that further investment is not justified. The organisation paid to learn.

Stopping at that point can be evidence of good management. Likewise, evidence of unexpectedly strong performance should allow capital and capacity to move towards an opportunity more quickly.

Gartner’s September 2026 research found that organisations it characterised as “high performers”, those that consistently tracked ROI on AI initiatives, treated AI as a portfolio of value, and regularly assessed performance and reallocated or discontinued underperforming initiatives, reported positive returns in 81% of their AI initiatives. This is an association within the surveyed group; it does not
establish that portfolio discipline alone caused the higher returns.

But the management principle is sound. New evidence should be allowed to change where resources go.

There is one complication. Some of the most important AI investments may look unattractive if they are judged in isolation. Consider a shared data capability, an enterprise control layer or common AI infrastructure. Its own direct revenue may be negligible.

Its value may come from allowing ten other initiatives to deploy faster, operate more safely or avoid building the same capability repeatedly. Treating it purely as a cost centre would miss that contribution. The same applies to investments required for regulatory compliance, operational resilience or material risk reduction.

They do not need an invented revenue number simply so they can compete with a sales use case. The organisation needs to understand what they enable or protect. That does not mean every platform or foundational investment is automatically strategic. Shared capabilities can also become expensive infrastructure that few people use.

They still need evidence. The question is different: What portfolio value does this investment enable that would otherwise be slower, riskier, more expensive or impossible? A good AI portfolio therefore cannot simply rank every initiative by forecast return and allocate all available resources from the top down.

That would favour what is easiest to measure today and could systematically starve experimentation, shared capabilities and longer-term strategic options. Portfolio discipline is not the same as continually funding today's highest-return use case. It is about being explicit about why different investments exist and holding each one to the type of contribution it is expected to make.

A sensible starting point is a consolidated view of the organisation's material AI investments. Not another enormous reporting layer. Just enough information to make the decisions visible. For each material initiative, management should understand its purpose, accountable owner, expected and realised outcomes, current operating economics, future resource requirements and the important assumptions that remain uncertain.

Then ask a small number of difficult questions. Which capabilities have demonstrated value and have a credible opportunity to scale? Which could become attractive but need their economics or operating model improved first? Which are sufficiently promising to justify another limited investment in learning?

Which shared capabilities enable enough portfolio value to warrant continued investment? And which initiatives should now receive fewer resources because better opportunities exist elsewhere? This also requires the enterprise to reconcile value across the portfolio.

If one business unit claims a productivity improvement created by an AI capability and the shared platform team claims the same improvement as part of its own value case, the organisation has not created that value twice. Ultimately, the portfolio has to reconcile back to the economics and operational performance of the business itself.

The point is not to make AI investment perfectly measurable. It rarely will be. The point is to make the trade-offs visible: What could this investment contribute?

How confident is the organisation? What will it require? What does it enable? What opportunity is being given up by funding it?

Those are much better questions than asking whether every project has a positive business case.

As AI becomes more deeply embedded in enterprise operations, the investment conversation will have to move beyond deciding which use cases should be approved. Executives will be making continuing choices about consumption, operating capacity, shared infrastructure, risk and the allocation of scarce resources across competing opportunities.

Some AI capabilities will deserve more investment even as their operating costs increase. Others will need to be redesigned, scaled back or stopped. Some uncertain opportunities will deserve another small investment because the information gained could unlock something much larger.

And some of the most important investments will be the shared capabilities that make the rest of the portfolio possible. A growing AI portfolio is not, by itself, evidence of a successful AI investment strategy. Neither is a portfolio in which every initiative can point to a positive original business case.

The real test is whether management is prepared to keep changing where resources go as better evidence emerges. Because the question is no longer whether an AI use case once deserved funding. It is whether, in light of the evidence now available, it deserves the next rand.

The question is no longer whether an AI use case once deserved funding. It is whether, in light of the evidence now available, it deserves the next rand.

The discipline of a mature AI portfolio is herefore not simply approval. It is the willingness to keep reallocating capital, scarce capability and executive attention as better evidence emerges.