- 01Prepare the data
- 02Build + evaluate
- 03Operate + improve
The most common thing we hear from clients who've tried to build AI before: "It cost three times what we expected." Here's why — and how to build a more accurate budget.
The Costs Everyone Forgets
Data Preparation
Most AI project budgets estimate "data costs" as the cost of storing or querying data. The real cost is the labor to make data usable.
In a typical enterprise AI project, data preparation takes 40–60% of total project time. This includes:
- Auditing existing data quality
- Writing parsers for inconsistent formats
- Cleaning, deduplicating, and normalizing records
- Building the ingestion pipeline from source systems
- Creating or procuring labels (if supervised learning is involved)
If you have a 12-week project budget, 5–7 weeks of it is data work. Plan accordingly.
Evaluation Infrastructure
How do you know if your AI system is working? This question requires investment that's almost never in the initial budget.
Evaluation infrastructure includes:
- A golden test set of inputs with known-good outputs (requires human expert time to create)
- An automated evaluation pipeline that runs after every change
- Metrics dashboards that show quality over time
- Human review workflows for cases the automated metrics don't cover
Without this, you don't know if your model is getting better or worse as you iterate. In practice, this means you keep iterating past the point of diminishing returns, or you ship a regression without knowing it.
Budget: 10–15% of the total project cost for evaluation infrastructure on any ML project.
Iteration Cycles
The first version of an AI system is rarely the right one. Real projects include multiple rounds of:
- Prompt refinement (for LLM-based systems)
- Hyperparameter tuning
- Model architecture changes based on evaluation results
- Dataset expansion when the model fails on edge cases
If your budget only covers one pass, you will exceed it.
Post-Launch Monitoring and Retraining
AI systems degrade over time as the world changes and they don't. A model trained in 2024 may perform worse in 2026 because:
- The distribution of inputs has shifted (data drift)
- The desired behavior has changed (concept drift)
- The underlying model it calls has been updated by the provider
Budget for:
- Monitoring infrastructure to detect degradation
- Periodic evaluation runs against the golden test set
- Model retraining or fine-tuning cycles (annually at minimum)
- Human review queues for flagged low-confidence outputs
Prompt Costs at Scale
LLM API costs are deceiving. A prompt that costs $0.002 per call seems trivial until you run 100,000 calls per day.
$0.002 × 100,000 calls/day × 365 days = $73,000/year in API costs alone.
Build a cost model before you start. Estimate:
- Number of requests per day/month
- Average input tokens per request
- Average output tokens per request
- Cost per 1000 tokens for your chosen model
Then add 30% for overhead. Then plan for the model provider to update pricing.
If the cost model doesn't work at scale, you need a different architecture (smaller model, caching, batching, self-hosted inference) before you build.
Change Management
Enterprise AI projects have a human cost that software projects don't always have: getting people to actually change how they work.
A new AI tool that no one uses is worthless. Getting adoption requires:
- Training sessions for end users
- Documentation that explains how to use outputs correctly
- A feedback mechanism for users to flag problems
- A champion in each team who drives adoption
- Iteration cycles based on user feedback
Budget for at least 2–3 person-weeks of change management for any enterprise AI deployment.
A Realistic AI Project Budget Template
For a mid-size enterprise AI project (3–6 month timeline, one primary AI feature):
| Category | % of Total Budget |
|---|---|
| Discovery and architecture | 10% |
| Data preparation and pipelines | 30% |
| Model development and integration | 25% |
| Evaluation infrastructure | 10% |
| Deployment and DevOps | 10% |
| Change management and training | 10% |
| Ongoing monitoring and iteration (first year) | Budget separately |
The "ongoing" costs are often omitted from project budgets entirely. They shouldn't be. AI systems require maintenance; plan for it.
Questions to Ask Before Signing a Statement of Work
- Does the data cost estimate include data preparation labor, or just storage/transfer costs?
- Is evaluation infrastructure included in scope?
- How many iteration cycles are included in the timeline?
- What does post-launch support and monitoring look like, and who pays for it?
- What is the API cost model at expected scale?
- Is change management and user training included, or extra?
If the answer to any of these is "that's out of scope," you now know where your budget overrun will come from.
The Shortcut That Costs More
The most common budget mistake: underscoping the project to get approval, with the intention of asking for more budget later when the project is already in flight.
This is a bad strategy for several reasons:
- Projects that run over budget lose stakeholder trust and get cancelled
- Descoping mid-project to hit a budget creates technical debt that costs more to fix later
- Teams that work under a false budget constraint make worse architectural decisions
A realistic budget that gets approved is worth more than an unrealistic budget that looks good until it doesn't.
When we scope AI projects, we always build a detailed cost model that includes the hidden costs. If you want a realistic estimate for your AI initiative, reach out.