When leadership teams first budget for an AI project, the instinct is usually to focus almost entirely on the technical build — licensing costs, model usage fees, engineering time to build the system. That instinct produces a budget that’s consistently, predictably wrong, because the AI model itself is typically the smallest line item in a realistic total cost, not the largest. The real cost — and the real risk of underfunding — lives in three other areas that get far less attention in the initial budgeting conversation.
Where the budget usually needs to go
Data preparation, which is almost always more work than expected. AI systems are only as good as the data they’re built on, and most organizations significantly underestimate how much work is required to get their existing data into a state that’s actually usable — cleaning inconsistent records, consolidating data scattered across multiple systems, filling gaps that were never a problem when a human was making judgment calls with incomplete information but become a real problem when an AI system needs consistent, structured input. This work is unglamorous, rarely gets its own line item in an initial budget proposal, and routinely ends up consuming more time and cost than the AI system itself.
Integration with existing systems, which is where a lot of technical complexity actually lives. An AI system that works well in isolation still needs to connect to whatever existing systems it’s meant to inform or act within — a CRM, an ERP, an internal tool nobody’s touched in years. This integration work is often more technically demanding than building or fine-tuning the AI component itself, particularly when the existing systems weren’t designed with this kind of integration in mind, which connects to the API-readiness challenges we discuss elsewhere in our Insights.
Change management for the people who’ll actually use the AI’s output. This is the most consistently underbudgeted category, and it connects directly to the core argument in AI Adoption Is a Decision Architecture Problem — an AI system’s output only creates value if the people receiving it actually change how they work because of it, and that requires real, deliberate investment: training, updated workflows, a genuine effort to build trust in the new system’s output rather than assuming trust will develop automatically just because the system works technically.
A rough allocation that reflects reality better than most initial budgets
For a typical mid-sized AI project, a more realistic allocation often looks something like this: a smaller share — sometimes surprisingly small — for the AI model or platform itself, a meaningfully larger share for data preparation and cleanup, a comparable or larger share for integration work connecting the system to existing infrastructure, and a genuinely significant share — often underbudgeted to near zero in initial proposals — for training, workflow redesign, and the ongoing work of building trust in the new system among the people expected to use it. The exact proportions vary by project, but the pattern holds broadly: the model is rarely where most of the money should go, even though it’s usually where most of the initial budgeting attention goes.
A note on contingency budget specifically for data surprises
Beyond the core allocation, it’s worth building in a specific contingency line for data-related surprises, since this is where realistic estimates most often prove optimistic. Data problems tend to reveal themselves gradually as a project progresses, rather than all at once during initial scoping, and a dedicated buffer for this category tends to get used far more often than a general project contingency would predict.
Why underbudgeting change management specifically tends to sink otherwise sound projects
A technically excellent AI system with an underbudgeted change management plan tends to fail in a very specific, avoidable way: the system gets built successfully, performs well in testing, and then sits underused in production because the people who were supposed to incorporate its output into their daily work were never given the training, workflow changes, or trust-building time needed to actually do so. From a budget review perspective, this looks like a technical success and a business failure simultaneously — which is confusing and frustrating for leadership, because by every technical measure, the project delivered exactly what was promised.
A practical way to build a more realistic budget upfront
Before finalizing a budget proposal, it’s worth explicitly asking, for each major AI initiative: what will the data cleanup actually require, realistically, not optimistically? What existing systems will this need to connect to, and how complex is that integration likely to be? And critically — who will actually need to change how they work because of this system’s output, and what will it genuinely take, in time and effort, to get them there? Answering these three questions honestly before finalizing a budget tends to produce a number that’s larger and better distributed than the model-centric budget most initial proposals start with — and considerably more likely to reflect what the project will actually cost by the time it’s genuinely delivering value.
What this looked like for one of our clients
A manufacturing company came to us with an initial AI project budget allocated almost entirely to model development and licensing, with a token line item for “training.” Working through a realistic cost breakdown, it became clear that data preparation across their fragmented legacy systems would require a substantially larger investment than initially planned, and that the change management effort for their operations team — who would need to genuinely trust and act on the AI’s output — deserved a much larger allocation than the initial token line item reflected. Rebuilding the budget around this more realistic distribution avoided a mid-project funding crisis that would have otherwise hit right as the technical build was completing but before the harder change management work had even begun. You can read more in our manufacturing AI budget planning case study.
The bottom line
If your AI project budget is weighted heavily toward the model or platform itself, it’s very likely underfunding the parts of the project most responsible for whether it actually succeeds. Data preparation, systems integration, and change management deserve a larger share of the budget — and a more serious place in the initial planning conversation — than most first-draft AI budgets give them.