Budget season has a predictable effect on AI conversations. A list of tools appears. Someone asks for licences. Someone else proposes a chatbot. A figure is put against “AI” and the discussion moves on.
That is not an AI budget. It is a software purchase plan.
The more useful budget question is what business outcome the organisation wants to improve in 2027, and what needs to change around the technology for that outcome to become real. The answer may include licences. It will also include people, process work, data preparation, review and the leadership attention to make decisions stick.
Start with three operating problems
Choose no more than three recurring problems where better information, better drafting, faster handling or more consistent decisions could make a genuine difference. Make them specific enough to observe. “Improve productivity” is too broad. “Reduce the time between a client query arriving and a reliable first response” is a workable starting point.
For each problem, write the current cost. It might be staff time, rework, missed follow-up, slow turnaround, customer dissatisfaction or a decision that is repeatedly delayed. If the cost cannot be described, the projected benefit will be difficult to describe too.
This is where many technology budgets should become smaller. A tool without a defined operating problem does not become strategic because it has a place in next year’s spreadsheet.
Budget for the work around the tool
An honest AI budget usually has five lines.
Tool and access. Licence costs, appropriate user accounts, integration and any supplier support.
Data and process preparation. Cleaning a shared document set, mapping a workflow, defining a knowledge source, or deciding which information cannot be used.
Capability. Short, practical learning tied to the task, guidance for normal users and time for people to practise.
Governance and assurance. A simple policy, review arrangements, security checks where they are needed and a named owner for the use case.
Change capacity. The management time required to make decisions, handle exceptions and review whether the work has actually improved.
The last item is frequently zero in the proposal and substantial in reality. Treating it as visible makes the decision more honest.
Fund a discovery phase
For a small firm, the first quarter of the year may be better spent on diagnosis than on a large rollout. Fund a limited discovery phase for one or two selected problems. Map the work, establish a baseline, test a suitable approach and decide whether the expected benefit survives contact with the actual process.
This is not a delay tactic. It is a way to avoid committing a full-year budget on the basis of a demonstration. A modest, well-run discovery phase produces information that a vendor presentation cannot.
Set a value threshold before you buy
Decide in advance what would justify continuing. The threshold could be a reduction in turnaround time without a decline in quality, a measurable drop in routine errors, greater client capacity, or a decision that becomes more consistent and easier to audit.
Avoid setting “staff enjoyed using it” as the only test. User acceptance matters, but it is not the same as business value. Equally, do not demand a dramatic financial return from every small experiment. Some early projects are worth doing because they build a necessary capability. The point is to name which kind of return is expected.
Keep high-consequence decisions separate
A budget for administrative support and a budget for systems that influence people are not interchangeable. If a proposed use concerns hiring, customer eligibility, credit, health, legal rights or another high-consequence decision, it needs a different level of scrutiny. The business should understand the data, the potential impact, the human role and the accountability before it gets near a production budget.
NIST’s AI Risk Management Framework is a useful reference for thinking about risk proportionately. It is not a template to copy word for word. It is a reminder that governance should match the consequence of the decision.
The strongest 2027 AI budgets will not look the most futuristic. They will show a small number of named operating problems, a credible path from trial to use, and enough capacity to do the ordinary management work that makes new technology worth having.