Beyond the Budget: The AI Decisions That Only Humans Can Make

Earlier this month I spent time with a group of senior executives discussing the economics of AI: what it actually costs, where the value is and is not materialising, and what the organisations that are getting returns are doing differently from the ones that are not.

That conversation encapsulates why this series is called Beyond the Budget. Not because cost does not matter. It does. But because the budget is where the consequences show up. The decisions that determine those consequences were made much earlier: about what to trust, who to invest in, and what to build first.

This final piece is about the third of those decisions, and it has the most direct line to the cost conversation keeping boards awake right now.

When we polled the group on their primary AI cost concern, Tokens came out as the clear leader at 58%, with Queries second at 33% and storage accounting for most of the remainder. That is understandable. But the conversation that followed was more revealing than the poll result.

The executives who had made the most progress on AI cost had not done it by finding cheaper tokens. They had done it by making the AI work less hard. And the reason their AI was working less hard was that they had made better decisions earlier, about what to build first, and what to build it on.

That is what this final piece addresses. Not how to spend less on AI, but how to make decisions that mean the costs do not spiral in the first place.

Many companies started 2026 with an annual AI budget and are discovering that it may last only a few months. Agentic AI uses technology differently. It runs more often, draws on more data, and makes more decisions without waiting for a human prompt.

The first reaction is often to slow deployment, negotiate harder with vendors, or move to cheaper models. Those things may help, but they do not solve the underlying problem.

AI becomes expensive when organisations try to do too much, too quickly, without being clear about the problem they are solving or the foundations they will need.

The silo trap and how to avoid it

This is where a lot of AI cost advice appears to contradict itself. You will hear two things that appear to be in tension: start with one specific problem rather than a big roadmap, and avoid building siloed projects that cannot share context or scale. Both of those are correct. The question is how to do the first without creating the second. The answer begins with a question that should be asked before the pilot starts.

The organisations making real progress are asking: what is one important problem we can solve now, and what will solving it teach us? Rather than trying to design the perfect AI strategy from the start, they focus on producing a measurable result. They learn what works, what does not, and what capabilities they will need next. In a market moving this quickly, the ability to learn and respond is more valuable than the appearance of having a complete plan.

But there is another question that separates a deliberate first use case from a silo: what does this use case need from the data layer that the next use case will also need? It might be a trusted customer record, a common business definition, or a reliable way to connect to a core system. Whatever the answer, that becomes the shared foundation. The first use case sits on top of it. The foundation is then reused by the next use case, and the one after that. Organisations that skip this question may build a successful pilot, but they also create a stranded asset. Those that ask it build a successful pilot and a reusable foundation. Over time, the difference in cost is significant. The difference in AI capability is structural.

Poor design makes AI more expensive

When teams build separately, AI systems repeatedly search for information, reconstruct context, and query the same data because nothing has been designed for reuse. That increases cost while often producing less consistent answers. The answer is not simply to find a cheaper model. It is to make the AI work less hard.

The organisations getting this right are not just making their data available to AI. They are making it intelligent. They are ensuring the business logic, the definitions, the relationships between things, are already understood before the agent asks its first question. When that is true, AI does not have to figure out the business every time it runs. It can get straight to the answer. A follow-up question does not trigger a fresh query because the context from the previous question is already there. That is cheaper, faster, and more consistent.

Here is what that looks like in practice. A leading retailer facing escalating AI costs deployed an intelligence layer on top of their existing data warehouse rather than replacing it. The result was a 75% reduction in warehouse query costs, faster response times, and richer output across hundreds of active users. The model did not change. The architecture underneath it did.

Three decisions leaders need to make

The organisations generating real returns from AI tend to make three decisions well.

1. Decide what outcome needs to change

Do not begin with the technology or ask where AI might add value. Start with a specific business problem and the result you want to improve, whether that is faster resolution times, fewer errors, lower costs, or better forecasting, and then decide whether AI is the right way to achieve it. A clearly defined problem produces a measurable pilot. A broad ambition to use more AI produces an expensive experiment.

2. Decide how and when you will judge it

Agree on the baseline, the measures, and the decision date before the pilot begins. At the end of that period, make a clear call: scale it, stop it, or change direction. A pilot that continues without a decision is not producing learning, it is simply consuming budget. The purpose of a pilot is not to prove that the original idea was right. It is to produce enough evidence to make the next decision well.

3. Decide what the next use case can reuse

Every pilot should leave the organisation with more than a one-off result. Ask what it needs from the data layer that another use case will also need, and build that capability once, then use it again. When the shared foundation carries intelligence, when the business logic and definitions are already embedded, every subsequent use case inherits that advantage without having to earn it again. That is how isolated experiments become a compounding capability, and it is how token costs stay predictable rather than multiplying with every new deployment.

Beyond the budget

AI cost is a real and growing problem and leaders and boards are right to take it seriously. But the organisations that are managing it well are not doing so by negotiating harder or spending less. They are addressing the decisions that determine whether AI investment creates value or waste in the first place.

That is how the three pieces in this series connect, and why each one has a direct impact on the budget conversation.

• Organisations with strong discernment spend less time correcting unreliable output, and correction costs money.

• Organisations that give people genuine permission to develop AI fluency get more value from the tools they already have, because fluent users ask better questions and better questions produce more useful answers from fewer tokens.

• Organisations that choose problems carefully, ask the forcing question before every pilot, and build intelligent data foundations are the ones whose AI costs are predictable rather than alarming, not because they spent less, but because they built the conditions for AI to work well.

The organisations that will lead on AI in three years will not be those that outspent their competitors or found the cheapest path. They will be the ones that made better human decisions about what to trust, who to invest in, and which race to stop running.

The budget is where the consequences show up. The decisions are where the outcome is determined. That is the conversation every leadership team needs to be having right now, not just about AI spend, but about the decisions that sit underpin it.