Spend
Enterprise AI Budget: How to Plan It and Keep It Under Control
Editorial team Delphis · Published on · 6 min read
An enterprise AI budget is the forecast of spending on artificial intelligence over a fiscal year, broken down by cost item, function, process and technology. It serves two purposes: deciding in advance how much to invest and where, and checking every month that actual spend stays consistent with those choices.
In many companies AI spend has grown by layering: licenses bought by different departments, usage-based services enabled for pilots, internal developments funded from innovation budgets. The result is real spend that nobody sees in full. This article explains how to build a complete AI budget, how to read it every month and how to manage variances.
What should the AI budget include?
A budget limited to licenses underestimates spend. There are six items to plan.
- Licenses: seats for tools such as Microsoft Copilot, ChatGPT and Claude, with their contract tiers.
- Usage-based spend: model API calls, consumption-based services, infrastructure for local models.
- Projects and agents built: development, integration with company systems, maintenance.
- Training: AI literacy and specialist training, including for the purposes of article 4 of the EU AI Act.
- Governance and management: the AI team's time, security and data protection work, control tools.
- Reserve: an unallocated share for new use cases that emerge during the year.
To estimate the less visible items correctly, see the overview of hidden costs in AI TCO.
How is the AI budget structured?
The cost item tells you how much is spent, not why. To govern the investment, the budget must also be read on three other axes.
- By function: how much each area spends (sales, operations, legal, HR). It makes shares and responsibilities visible.
- By process: which processes the money goes into. This is the axis that links spend to return.
- By technology: how much goes to each vendor or model. This is the axis that reveals concentration and overlaps.
The same euro appears on all three axes. A license assigned to a financial controller, for example, is spend of the finance function, of the reporting process and of the chosen technology. Keeping the three axes consistent requires every cost to be mapped to a user or a project, not just to a contract.
How do you set the AI budget, step by step?
The method below works both for a first budget and for the following ones.
- Inventory current spend. Contracts, invoices and usage for the last twelve months, by vendor and cost center.
- Map spend to usage. For each license, check whether it is used and by whom, as described in unused AI seats: how to find them.
- Start from use cases. For each active or planned use case, estimate users, tools required and full cost.
- Separate recurring from one-off. Licenses and usage repeat, projects have a build phase and a maintenance phase.
- Set targets next to costs. For each relevant item, an adoption target and a return target, with the baseline to start from.
- Assign owners. One overall owner and a contact for each function.
- Plan the review cadence. Monthly actuals, quarterly forecast review, annual contract check.
Step 5 is what separates an AI budget from a simple list of costs. Without targets, at year end you can say how much was spent, but not whether it was worth it. The method for building them is in the AI business case.
How do you read actuals every month?
Monthly actuals compare, for each item, planned spend for the period with actual spend. An example with hypothetical values shows the method.
A company has an annual AI budget of €1,200,000, split as follows: licenses €600,000, usage-based spend €180,000, projects €300,000, training €60,000, governance and management €60,000. At the end of September, after nine months, the linear planned share is 75% of each item, that is €900,000 in total.
Actual spend is €816,000, an overall variance of €84,000 under budget (84,000 / 900,000 = 9.3%). Read on its own, the total suggests everything is under control. The individual items say otherwise.
- Licenses: planned €450,000, spent €438,000. Variance of €12,000 under budget.
- Usage-based spend: planned €135,000, spent €162,000. Variance of €27,000 over budget (+20%).
- Projects: planned €225,000, spent €150,000. Variance of €75,000 under budget.
- Training: planned €45,000, spent €24,000. Variance of €21,000 under budget.
- Governance and management: planned €45,000, spent €42,000. Variance of €3,000 under budget.
The sum of variances (−12,000 + 27,000 − 75,000 − 21,000 − 3,000) comes back to −€84,000. The real picture is usage growing faster than planned while projects and training lag behind: an apparent saving that hides two execution problems.
How do you manage variances?
Every significant variance should be explained before it is corrected. The questions are always the same.
- Which item and which function cause it? A variance on the total is not actionable.
- Is it a volume effect or a price effect? More users or more calls, or a changed unit cost.
- Is it temporary or structural? A peak linked to a project or a trend that will continue.
- What did it produce? Spending over budget on a use case that delivers may be the right choice.
There are three possible decisions: correct spend (for example by setting usage limits, as described in AI FinOps), reallocate budget between items or functions, or update the forecast. Which one to choose depends on the item's return, not just its cost.
Which mistakes should you avoid?
- Checking only the total. Variances of opposite sign offset each other and hide problems.
- Ignoring usage-based spend. It is the most variable item and the one that grows without a formal purchase.
- Confusing assigned and used licenses. The budget must be based on actual usage.
- Renewing contracts without actuals. Renewal is when the budget is really decided, as the AI license renewal checklist shows.
- Separating budget from return. A budget met on use cases that do not deliver is still a poor investment.
Who should own the AI budget?
The AI budget cuts across functions and needs a single owner. In most cases this is the Chief AI Officer, or the CIO where that role does not exist, with the CFO for control and function heads for their shares. The owner is accountable for the overall budget, the contacts for the spend and results of their area. The role is covered in Chief AI Officer and AI team: responsibilities and KPIs.
Where Delphis fits
Delphis is a platform to govern AI investment, independent of vendors. It lets you set the AI budget by function, process and technology, compares it every month with spend reconciled against real usage and flags variances item by item, together with the adoption and return of use cases. See the platform overview.
Frequently asked questions
- What should a company's AI budget include?
- Licenses, usage-based spend, projects and agents built, integrations, training and the internal cost of management. Limiting it to licenses underestimates real spend.
- How often should the AI budget be reviewed?
- Every month, with actuals by item and by function. Usage-based spend can change quickly and a quarterly review often comes too late.
- Who is responsible for the AI budget?
- A single owner works best, usually the Chief AI Officer or the CIO, with the CFO for control and function heads for their own shares.
- How do you manage a variance on the AI budget?
- Identify the item and function that cause it, separate volume effects from price effects, then decide whether to correct spend, reallocate budget or update the forecast.
Your company's AI budget, spend, adoption and return, in a single view.
A walkthrough on your tools and use cases, with our team.
Request a demo