AI Budgets for 2027 Must Be Calculated In-House

Every budget season, a vendor presentation lands on the committee’s table, complete with a chart projecting AI capability growth over the coming year. Processing speeds have increased, the number of tasks handled by agents has grown, and more business processes are covered. Yet none of these figures answers the committee’s central question: how much of that promised growth will translate into additional profit this year?

In 2025, MIT examined the results of generative AI pilots at hundreds of companies. In 95% of them, no measurable impact on profits was recorded. That figure is worth keeping in view at every budget committee meeting. Various reasons are cited for these failures, from process integration to data quality. In my work with budget committees, however, I see a simpler and more common cause. Decisions are tied to the expiration date of a vendor’s commercial offer rather than the company’s budgeting cycle, and the discipline of financial analysis gives way to urgency. The committee ends up approving a budget based on a vendor’s promise, bypassing the steps normally required for a capital investment.

I will not consider a proposal without a baseline. First, I document the current process: its timelines, costs, and number of transactions per employee. Only then do I calculate what the process will look like after AI is introduced. A proposal that fails to describe the “before” state is just hope dressed up in an Excel spreadsheet, not a budget item.

Next, I separate the time horizons and assign them to different reporting lines. Some of the impact shows up in the current year’s profit: lower application-processing costs, shorter approval times, and savings on routine tasks. These gains feed into the profit and loss statement (P&L) for the coming quarter and are checked at the first review milestone. The other component is future cash flow. This includes entering a new market segment, retaining customers, and the cumulative benefits of scaling up. These returns are assessed using net present value (NPV) or customer lifetime value (CLTV) and recorded as a separate budget item with a longer reporting horizon. A committee that combines both figures into one number will struggle to explain to the board why an initiative looks underwhelming in its first year but implausibly generous three years later.

There is a third practice that is essential to keeping the numbers honest: calculate payback for each initiative, rather than assigning a single figure to the entire company. An average conceals the fact that one AI deployment may pay for itself in a quarter, while another may require three years of continuous oversight of operating and maintenance costs. Without this breakdown, a profitable initiative’s reputation can end up sustaining funding for an unprofitable one.

There is also a separate budget line that committees almost never bring up for discussion in advance: the cost of keeping AI manageable and under control. Governance and compliance now account for 8–12% of AI project budgets, up from 3–5% two years ago, making them the fastest-growing cost category. Every decision an AI agent makes without human involvement carries accountability requirements. Companies need to know who approves the result, who is responsible when something goes wrong, and how auditing and oversight work. Based on my experience with these deployments, this is precisely the cost category most often left out of total cost of ownership (TCO) calculations during the pilot phase. A year later, it returns as unplanned cost growth, forcing the committee into crisis-driven budget revisions instead of routine oversight.

These five questions should be applied to every AI budget proposal for 2027, with a separate interim review date assigned to each one. Has the process baseline been documented before implementation? Is the expected impact assigned to the right reporting line: quarterly profit or cash flow over several years? Is payback broken down by initiative, or hidden in a single average? Does TCO include an allowance for governance and compliance? And finally, who in the company is accountable to the committee if the promised results have not materialized by the review date?

A proposal that cannot withstand these five questions funds a vendor’s presentation at the company’s expense, rather than an initiative with measurable results. The companies that achieve real AI-driven margin growth in 2027 will have one thing in common. It will not simply be that they adopted the technology, because adoption alone guarantees nothing. It will be that their budget committees know how to reject a proposal that cannot provide these five answers.

Next budget season, another presentation will land on the same table, this time with even bigger numbers. AI capabilities are advancing faster than companies’ ability to keep them under control. The 2027 budget must answer the committee’s question: how much profit will those capabilities generate under its oversight?

Stanislav Yezhov, AI Development Director, PJSC Astra Group