What Digital Accessibility Teaches About Effective AI Use

The end of the cheap-token era has changed the ROI equation.

AI is not priced by seat. It’s priced by consumption. Early on, as the frontier-model companies were trying to drive adoption and further develop their models, the price did not reflect the actual cost. This led to widespread adoption, in which effectiveness and ROI were secondary to determining whether the problem could be solved at all with AI. This is now changing faster than most organizations have had time to fully reckon with.​

In one highly visible example, Uber exhausted its entire annual AI coding budget before summer. COO Andrew Macdonald said publicly that the “link is not there yet” between token consumption and useful products shipped. And while the company hasn’t been named, there’s a figure making the rounds about an enterprise spending $500 million with a single AI provider in a single month because nobody set usage limits.​

According to the FinOps Foundation, an industry group focused on cloud and AI financial operations, AI cost management was a concern for roughly one-third of financial operations practitioners in 2024. By 2026, it concerns nearly all of them. Per IDC, as reported by CIO, “Global 1,000 companies will underestimate their AI infrastructure costs by 30% through 2027.” IBM’s Institute for Business Value found that unaccounted technical debt can turn a projected 39% ROI into an actual negative 14% return on the same project.​

These are early signals of a broader adjustment that every organization deploying AI at scale will need to make. The reality-priced-token era introduces a new set of trade-offs that include new capabilities, new cost structures and new risks. Giving your employees unbounded token availability, especially if they have no insight into or experience with cost-effective use, can result in out-of-control costs.​

Getting the most from AI in this new era means developing an analytical framework built around consistent measurement of cost, speed and quality. Our work in digital accessibility has given us a practical view of what that framework looks like.​

What does digital accessibility teach us about AI and ROI?

Digital accessibility is a domain with codified standards and measurable outcomes. That makes it one of the earliest places where the impact of AI-assisted development becomes visible and quantifiable, and where the distinction between what AI does well and what it doesn’t is clear.

Here’s what we’ve observed. AI excels at tasks where equivalence of meaning across modalities is important, judgment is required and variability is acceptable. Large language models (LLMs) are trained to be helpful and generative; this means they’re not optimized for exhaustive validation.​

For example, when asked to review a codebase, they will not read every line of code; instead, they will use searches to find areas of interest and then evaluate the surrounding context. They have a bias toward action and must be explicitly told or configured to not just return the first best answer, but instead, do more thorough work—iteration that, while useful, is expensive. They are built to produce output, not to exhaustively audit it.​

In accessibility work, that distinction has direct and measurable consequences. The tasks that require exhaustive, consistent verification—checking every element against a defined standard, every time, without exception—are tasks where deterministic, rules-based tools outperform AI on every dimension that matters: speed, cost, consistency and completeness. AI is essential for handling the judgment calls, synthesis and explanations of what the results mean and how to address them. Deterministic algorithms ensure consistency and completeness. Each approach does what it does best.​

This is, in fact, what has taken LLM-based chat from a purely statistical generation machine into a more usable system—tools have added the determinism that the model inherently lacks. If you don’t know this, you might assume that all the advances made since the earliest LLMs have been made in the model itself, whereas a lot of the advances have actually been made in the harness, through deterministic tool-calling.​

Many companies think that they can just ask their AI to solve any problem—for example, code and test for accessibility. The market data reflects what happens as a result. WebAIM’s 2026 annual report, which scans 1 million websites, found accessibility errors up 10.1% year over year, with page complexity up 22.5% in a single year. CodeRabbit’s analysis of 470 real-world pull requests found AI-generated code carries nearly three times as many cross-site scripting vulnerabilities as human-written code. In Deque’s 2026 survey of 200 enterprise engineering leaders, 64% of respondents reported accessibility as the top driver of rework—despite explicitly prompting for accessibility in AI agents.​

The solution isn’t less AI. It is the appropriate application of AI. Agentic loops (in which AI agents autonomously prompt themselves rather than waiting for more input) are gaining traction for exactly this reason. By repeating the same instructions until the model converges on a solution, they explicitly account for AI’s structural strengths as well as its weaknesses. The downside is that this process costs a lot of tokens.​

In part two, I’ll share three levers for efficient, ROI-positive AI use—levers that are practical, actionable and applicable to any organization deploying AI at scale.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?