Enterprise AI must move from token-maxxing to ROI-focused strategies

The economics of AI integration have changed dramatically as the reality of token cost has set in. Tokenmaxxing may have been fun while it lasted, but innovation without ROI is not sustainable. Now that we’re beyond the cheap-token era, AI must justify itself with real, measurable business outcomes across the enterprise. That means no more throwing AI at every challenge. Organizations need to be far more strategic, using AI for what it’s good for, and deploying other approaches when there’s a better solution.

The Hidden Cost of AI Velocity

Increases in development velocity from AI tools are worth celebrating. But if the code is riddled with issues, the real progress is negligible. The consequences are steep: lost time and money, legal risk, and poor customer experiences.

The field of digital accessibility can teach us a great deal about how to use AI strategically. It’s an arena where codified standards and strict compliance requirements are everyday realities, and there’s no room for inaccuracy or inefficiency. With digital accessibility, the goal isn’t to get better at fixing issues; it’s to prevent those issues from happening in the first place. That goal should be the same for any application of AI. Otherwise, technical debt accumulates fast.

According to IBM, ignoring technical debt can result in an ROI decline of 18 – 29%. Results like that can wipe out any velocity gains from AI. In Deque’s 2026 survey of 200 enterprise engineering leaders, 64% named accessibility as the top driver of post-production rework, even though these same teams explicitly prompted their AI agents to write accessible code.

Accessibility debt, just like technical debt, is the accumulation of unresolved accessibility issues across an organization’s digital properties. It’s debt that compounds over time—issues that go unaddressed in design and development become costlier to fix later. Recent research shows that it’s 30x more expensive to fix an accessibility issue caught in production than at the design stage. That gives organizations a real financial incentive to catch issues early, especially with today’s code production volumes.

Strategic AI vs. Deterministic Tools

AI isn’t always the answer. In many cases, rules-based, deterministic tools deliver more consistent results, faster and cheaper. To make the right decision about what to use when, teams should begin by breaking the workflow into subtasks and assessing how best to tackle each challenge. For example, tasks that require consistent verification—checking every element against a defined standard—are ideal candidates for deterministic, rules-based tools.

AI is suited to judgment, synthesis, and generating options. It is not suited to validation: checking everything the same way, every time.

Some engineering teams handle this by running the same review repeatedly and comparing results. It works, but it isn’t free. In Deque’s own experiments, a single code review pass consumed roughly 60% of the tokens spent on a task. Coding itself took about 13%. Writing the tests took a comparable share. And one pass is rarely enough. The same review often has to run three to 10 times against the same codebase before it converges on the full list of real issues. Each pass is a fresh search, not a cumulative one, so nothing carries over from the last run.

Whether that’s worth it depends on cost tolerance. An organization can spawn enough agents, let them iterate and check each other’s work, and converge on a good answer. But doing so means paying for it in tokens and cycle time. The alternative is simpler: a deterministic check returns the same result every time, with no repeated passes.

Harnessing Context and Human-in-the-Loop

Another approach described in a recent case study involved combining automation and agentic AI with human review. In the organization’s new workflow, accessibility findings were provided to an AI agent, which used a remediation tool to apply expected HTML fixes directly to the source code and then automatically created and documented pull requests. Engineers then reviewed the AI-generated changes, approved pull requests, and maintained governance over quality and outcomes. The result: 253 engineering hours returned, with remediation 98% faster overall. The workflow was estimated to save more than $25,000 in engineering costs.

Getting good results from an AI agent depends on how it’s set up, not just on good prompts. Engineers call this harness engineering and context engineering: building the specific tools, checks, and available information that let an agent do good work in a particular environment, rather than leaving the model to figure it out on its own. An agent reviewing code doesn’t need the whole codebase loaded into its context window. Let it search for the relevant files and pull in only the surrounding code it needs, and it will usually do better work for less money than one handed everything at once. The same logic applies to reusing context across calls instead of rebuilding it from scratch each time.

Stepping back, remember that the reason any of this matters is that the volume of AI-generated code is growing at a genuinely frenetic pace. In that context, rising costs seem inevitable: either from having to fix all the issues unvalidated AI code produces, or from checking ever-increasing amounts of code before it hits production.

But rising costs aren’t inevitable. And digital accessibility offers a solution—one that involves several balancing acts, between AI and rules-based, deterministic tools; between AI-powered automation and human validation; and between scalability and cost-efficiency. Getting these balances right is about organizational discipline. The sooner organizations develop the discipline to use AI intentionally, efficiently, and appropriately, the sooner they can expect to see positive, measurable returns on their AI investments.