Every AI vendor has a pricing page. Every AI business case has a budget. And almost every AI project ends up costing 2-3x the original estimate.

This isn’t because people are bad at budgeting. It’s because there’s a whole category of costs that don’t appear in vendor quotes or standard project templates. After years of watching AI projects go over budget, I’ve catalogued the recurring offenders.

The Obvious Costs Everyone Budgets For

Let’s start with what does appear in most business cases:

• Software licenses
• Cloud compute for training and inference
• External consultants or implementation partners (firms like PwC, KPMG, or specialist AI consultancies)
• Some training for end users

If your budget only covers these, you’re looking at roughly 40% of your actual spend. Maybe less.

The Hidden Costs That Kill Budgets

Data Preparation (The Big One)

I put this first because it’s consistently the largest hidden cost. The typical estimate I see for data preparation is 10-15% of project budget. The reality is closer to 40-50%.

Why? Because most organisations don’t actually know what state their data is in until they try to use it for AI. Common discoveries include:

• Critical fields that are 30% empty
• Inconsistent formats across different systems
• Historical data that was never migrated properly
• Labelling requirements nobody anticipated
• Data quality issues that require process changes to fix

One financial services client budgeted $200,000 for data preparation. Final spend was $780,000, and they had to reduce scope significantly.

The Integration Tax

Your AI system needs to talk to other systems. This sounds straightforward until you realise that:

• Your ERP system is running a version from 2016 that doesn’t have modern APIs
• Your CRM has custom fields that don’t map to anything standard
• Your data warehouse was built by contractors who left years ago
• Each integration requires its own security review, testing, and sign-off

Budget for each major integration to take 4-8 weeks and cost $50,000-$150,000. Multiply by however many systems you need to connect.

Computing Costs That Escalate

Cloud compute looks cheap until it isn’t. Common surprises:

• Inference costs that scale faster than expected as usage grows
• Retraining costs when model performance degrades
• Storage costs for training data that never get deleted
• Development environment costs that were supposed to be “temporary”
• Egress charges for moving data between cloud providers

One retail client’s monthly compute costs went from $8,000 during pilot to $67,000 at scale. The AI still delivered positive ROI, but the margin was much thinner than projected.

Change Management (Actually Doing It Properly)

Every project plan includes a line item for change management. Most dramatically underestimate it.

Real change management for AI involves:

• Training that goes beyond “how to click buttons” to “how to think differently about your work”
• Communication campaigns that address legitimate concerns about job impact
• Workflow redesign that actually integrates the AI into daily processes
• Champions programs to drive adoption in each business unit
• Ongoing support for users who struggle with the transition

Done properly, this is a 6-12 month effort requiring dedicated resources. Budget 15-20% of your total project cost.

Governance and Compliance

AI brings new risks that require new governance. This means:

• Legal review of AI-generated outputs and liability implications
• Privacy assessments for any personal data used in training
• Bias audits if the AI makes decisions affecting people
• Documentation for regulatory compliance
• Ongoing monitoring and audit capabilities

For regulated industries (financial services, healthcare, insurance), add another 20-30% to your budget for compliance alone.

Internal Resource Costs

The people who “just need to spend a few hours a week” on the AI project:

• Business SMEs for requirements, testing, and validation
• IT resources for infrastructure and integration support
• Data owners for access approvals and quality issues
• Security team for reviews and approvals
• Legal for contract review and risk assessment

Add up all these “few hours a week” and you’ll find several FTEs worth of effort that never appeared in the budget. Some organisations track internal labour; most don’t. Those that don’t consistently underestimate project costs.

The Ongoing Tail

AI systems don’t just run themselves. Ongoing costs include:

• Model monitoring and performance tracking
• Regular retraining as data patterns change
• Prompt tuning and optimisation
• Version updates and security patches
• Support for users and escalated issues
• Continuous improvement based on feedback

Plan for 20-30% of initial project cost as annual ongoing expense. More if you’re using cutting-edge technology that evolves rapidly.

A More Realistic Budget Framework

Based on patterns I’ve seen, here’s a framework for realistic AI budgeting:

Category	Typical Estimate	Realistic Estimate
Core technology	30%	20%
Data preparation	15%	35%
Integration	10%	15%
Change management	5%	15%
Contingency	10%	15%
Total	70%	100%

Notice that “core technology” – the thing everyone focuses on – is actually the minority of spend.

What This Means for AI Decisions

I’m not trying to discourage AI investment. These projects can deliver significant value. But that value calculation needs to account for true costs.

The questions to ask:

1. Have we fully scoped the data preparation effort with people who’ve actually seen the data?
2. Are internal resource costs tracked and included?
3. Have we included realistic contingency (at least 15-20%)?
4. Does the business case still work if the project costs 2x the estimate?

If the answer to that last question is no, the project is higher risk than it appears. Either find ways to reduce scope or be very confident in your estimates.

Final Thought

Hidden costs aren’t hidden because people are being deceptive. They’re hidden because AI projects touch more parts of the organisation than traditional IT projects. Every touchpoint has a cost.

The organisations that do AI well aren’t the ones with the biggest budgets. They’re the ones who understand what things actually cost and plan accordingly.