How Is the Budget for an Enterprise AI Project Determined?

In short: the budget for an enterprise AI project is built from four line items: setup, running, internal team time and maintenance. Deciding on the basis of a single quoted figure is the most common reason for surprises in the project's second year. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, and lists escalating costs among the leading reasons. Getting the budget right is the most concrete way to keep a project from being cancelled.

Why a single figure misleads

The quote received for an AI project usually covers only the setup cost. Model usage costs, integration licences and the maintenance share that appear after go-live all sit outside that figure.

Gartner's June 2025 prediction points at this directly: over 40% of agentic AI projects will be cancelled by the end of 2027, driven primarily by escalating costs and unclear business value. In other words, projects mostly close not because they failed to work but because their cost was never forecast.

That is why a budget is built on four line items rather than on a quoted figure.

The four budget line items

1. Setup. Diagnosis, design, development and go-live. It is a one-off item and it is usually the part visible in the quote. Its scope is driven by the number of systems to integrate with and the complexity of the process.

2. Running. Model usage fees, server or cloud cost, third-party service charges. Monthly and usage-based. It grows with volume, so an annual budget is never built on the figure seen during a pilot.

3. Internal team time. The hours the process owner, IT and end users spend on the project. Because it is never invoiced it is usually left out of the budget, yet it is a real cost and typically lands somewhere between a quarter and half of the setup fee.

4. Maintenance and improvement. The project does not end when go-live does. Model output has to be monitored, defects corrected, and the system updated as the process changes. Setting aside 15% to 25% of the setup fee per year is a realistic starting point.

Three factors that grow the budget

What separates the cost of otherwise similar work is rarely the technology choice. It is usually these three factors:

Data order. If documents are scattered, if the same fact is written three different ways in three places, or if access rights are undefined, the first third of the project goes into gathering data. This pulls the cost noticeably above what an organization with tidy data would pay.

Number of integrations. There is a serious difference between a solution connected to one system and one connected to CRM, ERP and email at the same time. Every integration brings its own authorization, error handling and testing load.

Approval and compliance requirements. Processes that handle personal data, fall under audit, or are bound by sector regulation add record keeping, traceability and approval steps. These are not technically hard, but they take time.

The right way to keep a budget small

The worst way to cut a budget is to keep the scope and compress the schedule. The right way is to narrow the scope.

Approach	Result
Start with one process	The highest-volume process is chosen, the result is measured in 90 days, the next step is decided from data
Five processes at once	None finishes, nothing can be measured, the budget runs out
Try an off-the-shelf tool first	If the process is covered, no development happens; if it is not, the gap is clearly visible
Go straight to custom development	Full price is paid for a system that may not have been needed

The two approaches in the first column buy more learning for the same money.

Questions to ask when evaluating a quote

• Which items does this figure cover, and which does it not?
• What is the estimated monthly running cost after go-live, and on what assumption does it rest?
• How many people from our team do you expect, for how many hours?
• How is maintenance priced after the first year?
• How will a request outside the agreed scope be priced?

A quote without written answers to these five questions is not a comparable quote.

Frequently asked questions

Is there a minimum budget for an AI project? The minimum condition for a meaningful result is scope, not budget. A single measurable process can be built on a limited budget. Keeping the scope wide and the budget small does not produce a result.

Do model usage costs rise over time? They rise as volume rises, though unit prices are trending down. Budgeting for volume growth is safer than relying on unit-price decline.

Why set aside a maintenance share? Processes change, systems update, users bring new needs. Projects without a maintenance share become unusable in their second year.

Who should approve the budget? The business unit that owns the process, together with IT. Projects owned only by IT go unused by the business; projects run only by the business get stuck at integration.

Sources

• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025) — escalating costs, unclear business value and inadequate risk controls are named among the reasons for cancellation.
• McKinsey & Company, The State of AI: Agents, innovation, and transformation (2025) — 88% of organizations use AI, only 39% can show an impact on earnings.

Closing

An AI budget is not a figure but a structure of four line items. When the items are written separately, quotes become comparable and the second year holds no surprises.

See our enterprise AI consulting service or look at our engagement models.