The AI ROI problem: what AI really costs

Budget for what it takes to make an AI tool useful, not just what it costs to buy: integration, data preparation, usage, evaluation and training. Then measure the return across the whole workflow the tool supports, not the single task it speeds up. Our report The AI ROI Problem has the full argument and a self-assessment.

Estimating what an AI tool costs to purchase is easy. Most vendors publish subscription prices, and usage-based services publish rates for processing requests. Estimating what it will cost to make that technology useful inside an organization takes far more work.

An AI assistant might cost little to operate, but connecting it to internal systems, preparing the data, managing permissions and supporting employees can require a substantial investment. Some of those costs occur during implementation. Others continue for as long as the system is in use.

The potential return deserves the same scrutiny. Saving an employee ten minutes on a task doesn't save the organization ten minutes of paid labor if the surrounding workflow stays the same. We think both sides of that equation need more attention when organizations evaluate AI investments.

Why so many projects struggle to reach production

Companies are investing heavily in AI, but turning those investments into reliable business systems remains difficult.

In a 2025 survey by S&P Global Market Intelligence, 42% of organizations said they had abandoned most of their AI initiatives, up from 17% the year before, and the average organization had scrapped 46% of its AI proofs of concept before they reached production.¹

Those figures cover projects with different objectives, technologies and levels of complexity, but they show how often experimentation fails to become an operational capability.

A proof of concept might demonstrate that a model can answer questions or perform a task. Putting that capability into daily use introduces requirements that weren't part of the demonstration: reliable information, access controls, integration with existing software, ongoing support, and people willing to use it. Each of those affects the cost and the likelihood of a return.

Where the costs actually come from

A complete AI budget accounts for more than the subscription or model usage charges.

• Integration and development. Connecting AI to an organization's applications, databases and workflows is engineering work. Some systems have well-documented interfaces. Others rely on older software, custom configurations, or processes that were never designed for integration.
• Data preparation. Documents and records may need to be cleaned, organized, deduplicated, or made accessible through new interfaces. Existing permissions also need to be understood and enforced.
• Infrastructure and usage. Running models, retrieving information, storing data and processing requests all cost money to operate. How much depends on the system's architecture, the models selected, and how often people use it.
• Evaluation and maintenance. Models and the underlying information change. The organization needs a way to evaluate performance, investigate incorrect results, maintain integrations, and respond when something fails.
• Training and operational support. Employees need to understand how the system fits into their work, which tasks it can support, and when its output needs checking. Managers and technical teams also need time to oversee adoption and address issues.

The weight of each category varies. A standalone productivity tool may need very little integration, while a custom system that works across several internal platforms can need far more engineering than it spends on models.

We recommend estimating these costs over the expected life of the system rather than treating implementation and ongoing operations as separate decisions.

Usage affects both cost and return

As more people use an AI system, its operating costs rise. More requests consume more computing resources, and broader use brings more need for support, monitoring and infrastructure.

That growth isn't a problem in itself. A system that delivers measurable value can justify higher operating costs, as long as the benefit grows with them.

Consider an internal assistant used to retrieve engineering records. During testing, a small group might submit a few hundred questions each week. A company-wide deployment could generate thousands of requests across several departments. The larger deployment needs more infrastructure, more frequent data updates and more oversight. It also has the potential to return far more employee time. The budget should account for both.

This is also why adoption needs attention throughout the project. Employees avoid a system that doesn't fit their workflow, produces unreliable answers, or takes more effort than their existing process. Few have been shown how: in a 2025 SurveyMonkey study, only 13% of American workers said they had participated in AI training.² Training alone won't ensure adoption, but employees need enough guidance to understand where a tool can help.

We prefer to introduce AI within a defined first production workflow, with an accountable owner and a way to measure whether employees are using it effectively.

Measure the workflow, not just the individual task

AI performance is often described by how much faster a particular task can be completed. That is useful, but it isn't the return to the business.

Imagine a process that takes an employee 30 minutes to complete. An AI tool reduces one 10-minute portion to two minutes. That task is now five times faster. The overall process, if everything else stays the same, improves from 30 minutes to 22.

Time saved also doesn't always turn into lower costs. Employees may use the recovered time for other responsibilities, or the process may still depend on approvals, reviews and handoffs that haven't changed.

Those outcomes can still be worthwhile. Greater capacity, faster turnaround, better accuracy and less rework all have value, provided the organization decides what it intends to improve and measures the result. That is why we evaluate AI at the level of the workflow it supports.

Budget before, during and after deployment

We organize AI investment decisions around three stages.

1. 01

   Before implementation

   Establish what the current workflow costs in time, labor, errors or delays. Identify the expected improvement and estimate the full cost of building and operating the system, using a range of usage assumptions rather than a single forecast.

2. 02

   During implementation

   Track development spending alongside answer quality, system performance and early user feedback. As the system moves into production, watch usage and operating costs together, so you can tell whether a change in spending comes from adoption, inefficient design or unexpected technical requirements.

3. 03

   After deployment

   Compare actual performance against the original baseline. Review whether the system is being used, what it costs to maintain, and whether the intended improvements happened. Then decide whether to expand, improve, change approach or stop.

Having those measurements makes whichever decision you reach easier to defend.

Starting small

A modest application that removes recurring work often provides more dependable value than an ambitious system with unclear ownership or uncertain operating costs.

If you're evaluating an AI initiative, or trying to understand why an existing project hasn't delivered what was expected, a free assessment is a place to start.

Sources

1. S&P Global Market Intelligence survey of more than 1,000 organizations in North America and Europe (2025), as reported by CIO Dive. ↩

2. SurveyMonkey, AI Sentiment Study, Q3 2025. Online panel of 10,558 US adults, fielded July to September 2025. ↩