What AI implementation actually costs

Two numbers explain the strange mood around AI budgets this year. Gartner projects global AI spending will reach $2.59 trillion in 2026, up 47% from last year. MIT's Project NANDA found that 95% of generative AI pilots at large companies show no measurable P&L impact. Record spend on one side, invisible returns on the other. The gap isn't a model problem. It's a budgeting problem, and it starts with what the pilot invoice leaves out.

Why do most AI pilots fail to reach production?

A pilot proves a model can do a task. Production requires everything around the task: pipelines that feed the model clean data, access controls, an evaluation suite that catches regressions before your customers do, error handling for the cases that used to be a human's judgment call, and people who trust the output enough to act on it. None of that shows up in the pilot budget, so when the demo works, the real invoice arrives as a surprise.

Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, and inadequate risk controls. Read that list again. Not one of the three is "the model wasn't smart enough." All three are scope failures. The model just takes the blame.

The line items the demo hides

When we scope AI work for mid-sized companies, the model is rarely the expensive part. The budget that survives contact with reality has five lines the pilot never showed:

• Data preparation. Budget analyses across 2026 consistently put data prep at 30 to 50% of total implementation cost, and it's the line teams underestimate most. If your product data lives in four systems with three definitions of "active customer," the model inherits the confusion.
• Integration. The agent that drafts a quote is a demo. The agent that drafts a quote inside your CRM, with your price book, logged against the right deal, is a project.
• Evaluation. You need a repeatable way to know the system still works after every prompt change and every model upgrade. Teams that skip this ship regressions and lose internal trust, which is harder to rebuild than the pipeline.
• Security and compliance review. Access scoping, audit trails, and vendor terms. If you sell to regulated buyers, their procurement team will ask. We wrote about that side separately in our AI governance guide.
• Change management. The cheapest model in the world produces zero ROI if the team routes around it. Training and workflow redesign are real line items, not soft costs.

The second bill: what AI costs after launch

The build is not the budget. Total cost of ownership analyses through 2026 put ongoing run cost at 20 to 40% of the original build cost, every year. Inference bills that scale with adoption, monitoring, re-running evaluations, and the migrations nobody plans for: model deprecations now arrive on 12-to-18-month cycles, and each one means re-testing everything that touches the model.

Treat that as a product budget with an owner, not a project remainder. The companies in that MIT 95% mostly didn't fail at launch. They failed at month nine, when the pilot's champion moved on and nobody owned the run cost.

How should a mid-sized company budget for AI?

The framework we use with clients has five rules:

1. Name the P&L metric before the pilot. Hours saved, cost per ticket, conversion rate, days-to-close. If you can't name the number, you're funding a science fair.
2. Budget the 24-month TCO, not the pilot. A useful rough multiple: whatever the pilot costs, the production system costs 3 to 5 times that over two years, run cost included.
3. Fund data readiness first. It's the biggest line and the longest lead time. We've seen enough stalled projects to give this its own playbook in the data readiness gap.
4. Set kill criteria in writing. A pilot without a defined failure condition never dies. It lingers, and lingering pilots are where the 95% lives.
5. Assign the run budget an owner. Someone's name next to the monthly number, reviewed quarterly like any other product investment.

None of this makes AI cheaper. It makes the spend legible, which is what lets you scale the two use cases that work instead of feeding ten that don't.

This is the arithmetic our AI practice runs in the first weeks of an engagement, usually alongside Data, because the readiness work is where the budget truth lives. The DGTL Readiness Index puts a number on that AI dimension, next to the other seven.

Related: From pilot to production → · RAG systems in production → · The data readiness gap →