There is no universal correct percentage, and any specific figure circulating in industry surveys should be treated as a rough benchmark rather than a target, since the right AI budget depends on how many validated, high-value use cases a company actually has, not on matching a peer average reported in a press release. A more useful approach than picking a percentage is working backward from identified use cases: total the cost of running discovery, piloting, and then scaling the two or three highest-value opportunities found through a proper prioritization exercise, and let that sum become the AI line item rather than reserving an arbitrary share of IT spend upfront and searching for projects to fill it. Companies earlier in AI adoption often underspend on data readiness and governance relative to model or tool spend, which causes pilots to stall later at higher cost than if the groundwork had been funded from the start. As adoption matures past the first few production systems, ongoing costs shift from one-time build spend toward continuous items like GPU or API usage, monitoring and maintenance, which should be budgeted as recurring operating cost rather than one-time project cost. Nanobase AI helps clients build this bottom-up budget from actual identified use cases rather than starting from an industry percentage that may not fit their situation.
Why a percentage benchmark is the wrong starting point
Industry surveys reporting an average AI share of IT budget circulate widely, and treating that figure as a target rather than a data point leads companies to either reserve a budget and search for projects to fill it, or feel behind for spending less than a peer average that may not reflect their actual use case portfolio at all. The right AI budget is whatever it costs to run discovery, pilot, and then scale the specific use cases a company has actually identified and validated, not a percentage borrowed from an industry survey.
Building the budget bottom-up, by line item
Summing the actual cost of each category for the specific use cases a company has prioritized produces a defensible number for board approval, unlike an arbitrary percentage of IT spend.
| Budget category | When it's incurred | What it covers |
|---|---|---|
| Discovery and use case scoping | Upfront, one-time | Interviews, technical assessment, prioritization |
| Pilot build | Early, one-time per use case | Development, integration, initial evaluation |
| Infrastructure or API usage | Ongoing, recurring | GPU capacity or hosted API token spend at production volume |
| Governance and compliance | Ongoing, recurring | Committee time, policy maintenance, regulatory tracking |
| Training and change management | Recurring, per rollout | Role-specific training, champion support, refresh cycles |
| Monitoring and maintenance | Ongoing, recurring | Drift monitoring, periodic re-evaluation, support |
This bottom-up total also lines up with typical enterprise AI project cost ranges seen at each phase, which is a useful cross-check once the line items are filled in.
The underspending pattern that causes later problems
Companies earlier in AI adoption commonly underspend on data readiness and governance relative to the more visible line items, model or tool licensing and initial development, which causes pilots to stall later at higher cost than if the groundwork had been funded from the start. Data cleanup discovered as a blocker mid-pilot costs more in both money and momentum than the same work budgeted upfront, and a governance function stood up reactively after an incident costs more in remediation than the same structure built proactively alongside the first deployment.
How the budget mix shifts as adoption matures
Early-stage AI spend concentrates in one-time costs: discovery, pilot builds, initial infrastructure. As adoption moves past the first few production systems, spend shifts toward recurring operating costs, GPU or API usage at scale, ongoing monitoring, periodic re-evaluation against newer model options, and this shift should be reflected in how the budget is planned rather than treating AI spend as permanently one-time and project-based. Companies that keep budgeting AI purely as project cost, without a recurring operating line, tend to underfund the maintenance that keeps a production system performing well over time.
A practical exercise for the next budget cycle
- List every AI use case currently identified, whether piloted, in production, or still under consideration.
- Estimate cost for each against the line-item categories above, distinguishing one-time from recurring spend.
- Sum across use cases to produce the actual bottom-up total.
- Compare that total to current IT budget as a sanity check, not as the starting point for the calculation.
- Flag any category, especially governance and data readiness, that has historically been underfunded relative to build spend.
Frequently asked questions
Is there any reasonable percentage benchmark to reference at all?
Published industry figures exist and can serve as a rough sanity check, but they average across companies with very different use case portfolios and maturity levels, which makes them unreliable as a target. Use them only to flag a bottom-up total that looks unusually high or low, not as the basis for the number itself.
Should AI budget be separate from the general IT budget?
Either approach can work, but AI spend should be tracked distinctly enough to see the breakdown between one-time build cost and recurring operating cost, regardless of which overall budget line it sits under. Losing this distinction makes it hard to plan for the following year's recurring costs.
How do we budget for AI when we don't yet know which use cases will scale?
Budget discovery and a small number of pilots first, with scaling costs estimated only once a pilot has validated real usage and value. Committing scale-phase budget before a pilot has proven out tends to lock in assumptions that turn out wrong once real usage data arrives.
How Nanobase AI helps
Nanobase AI helps clients build this bottom-up budget from actual identified use cases rather than starting from an industry percentage that may not fit their specific situation, so the resulting number holds up under board scrutiny because it traces back to real, scoped work.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.