For the past several years, the executive conversation around artificial intelligence has been fairly predictable:
“We need an AI strategy.”
“What are our competitors doing with AI?”
“Why aren’t we doing more with AI?”
And, inevitably:
“Can we add AI to the presentation?”
Apparently, no corporate initiative is complete until a small glowing robot has been placed somewhere on Slide 4.
But the conversation is changing. The enthusiasm is still there, the investment is certainly still there, and the PowerPoint robots remain gainfully employed. What has changed is the question coming from CFOs, boards and private equity sponsors:
Where is the return?
That question is arriving at an awkward time.
Gartner forecast worldwide generative AI spending would reach $644 billion in 2025, an increase of more than 76% from 2024. Meanwhile, the 2025 MIT NANDA report found that only 5% of task-specific enterprise GenAI tools in its research sample reached successful implementation. The report described this as a 95% failure rate for enterprise AI solutions. It also noted that the figures were directionally based on interviews and differing organizational definitions, rather than uniform company reporting. Still, “directionally concerning” is not generally the phrase executives hope to hear after approving a seven-figure AI budget.
This does not mean AI is a bad investment. It means indiscriminate AI purchasing is a bad investment.
There is a big difference.
The next phase of enterprise AI will not be about buying less AI simply to contain costs. It will be about buying the right AI, for the right employees, tied to the right workflows, under commercial terms that do not require a team of quantum physicists to forecast next year’s bill.
Welcome to the AI ROI reckoning. Or, as procurement teams may come to know it, the AI tax.
Software shelfware is not new. Enterprises have spent decades purchasing more licenses, modules and capabilities than employees actually use. AI has simply given this proud tradition a more exciting name.
The classic rollout looks something like this:
Success!
Several Microsoft-focused licensing and security advisers report that organizations deploying Copilot broadly, without a phased adoption strategy, can find approximately 30% to 40% of licenses unused within the first 90 days. That figure should be treated as a practitioner estimate, not a universal Microsoft benchmark. Microsoft has not published a single market-wide statistic confirming that percentage. The underlying risk, however, is real enough that Microsoft’s own Copilot Dashboard distinguishes between assigned licenses and active users, recommends reassigning unused licenses, and supports reporting across 7-, 30-, 90- and 180-day windows.
In other words, Microsoft has thoughtfully provided the tools required to determine how much Microsoft software you purchased but are not using.
Much appreciated.
The problem usually begins with the assumption that every knowledge worker is equally positioned to benefit from an AI license. They are not.
A salesperson preparing proposals, summarizing meetings and drafting follow-up correspondence may generate meaningful value from an AI assistant. A financial analyst working through recurring reporting tasks may also have measurable opportunities. An executive assistant coordinating meetings, documents and communications could become a highly active user.
A worker who spends most of the day in a specialized operational system and opens Word twice a month probably does not need the same license on Day One.
AI licensing should therefore be based on roles, workflows and measurable opportunities, not merely employee counts.
Before purchasing 10,000 licenses, identify the 1,500 employees with high-frequency use cases. Establish a baseline. Measure adoption and outcomes. Then expand.
This may sound less dramatic than announcing an enterprise-wide transformation at the next town hall, but it has the advantage of being economically rational.
Traditional enterprise software pricing was rarely simple, but at least the units were familiar. Users, devices, processors, cores, transactions or capacity could usually be placed into a spreadsheet and argued about for several months.
AI pricing adds some exciting new possibilities:
Soon, the standard enterprise order form may come with its own graduate-level statistics course.
The market is already moving away from purely flat, per-user pricing toward hybrid and consumption-based structures.
GitHub moved Copilot plans to usage-based billing on June 1, 2026. Under the new model, usage consumes GitHub AI Credits calculated through token consumption, including input, output and cached tokens, with different models potentially carrying different economics.
Microsoft prices Copilot Studio through Copilot Credits. The number of credits consumed can depend on the type of agent, the knowledge sources used and the complexity of the response or action. Microsoft also offers pay-as-you-go access for certain Microsoft 365 Copilot services, allowing usage-based billing without a full upfront license commitment.
Salesforce offers Agentforce through a combination of consumption-based Flex Credits, conversations and per-user licensing. Its published Flex Credits rate card also states that credits must be used before the order end date, that rollover is not permitted under those terms, and that usage types and multipliers may be updated from time to time.
This creates a forecasting problem.
Under a flat license, the buyer knows the annual cost, even when the software is not used. Under consumption pricing, the buyer pays closer to actual usage, which sounds wonderfully efficient until adoption accelerates, agents begin performing more complex tasks, model multipliers change, or previously inexpensive workflows begin consuming credits like a teenager with access to a parent’s DoorDash account.
The commercial risk is no longer limited to buying too much. Buyers must now manage two opposite problems:
Underconsumption: You prepay for credits or capacity that expire unused.
Overconsumption: Adoption succeeds, usage expands, and the resulting bill celebrates your success more enthusiastically than you do.
The answer is not to stop buying AI. That would be the strategic equivalent of solving cloud cost problems by unplugging the data center.
The answer is to apply more discipline to what is purchased, how it is deployed and how vendors are held accountable.
Divide users into practical cohorts based on likely value:
Deploy first to the roles where AI can remove a measurable bottleneck, reduce external cost, improve throughput or affect revenue. Do not begin with “everyone.” Everyone is not a use case.
The MIT research found that successful organizations tended to focus on narrow, high-value workflows, integrate tools into actual operations and measure outcomes rather than simply admiring the technology. It also found that back-office functions such as operations and finance frequently produced clearer savings than more visible front-office experiments.
AI licenses should not remain assigned indefinitely because someone used the tool enthusiastically during launch week.
Review adoption at 30, 60 and 90 days. Measure more than logins. Look at recurring use, functions used, workflows affected and output generated. Reclaim licenses from inactive users and redeploy them to employees demonstrating demand.
This is not punitive. It is portfolio management.
An unused AI license is not innovation. It is shelfware wearing futuristic glasses.
Buyers should treat AI contracts as evolving economic structures rather than ordinary SaaS renewals.
Depending on the supplier and deal, negotiate for:
Not every supplier will agree to every protection. That is why they are called negotiations rather than order-entry assistance.
The most dangerous commitment is a large, noncancelable consumption floor based on projected adoption that has not yet occurred. A three-year forecast built on “employees will probably use it” is not a forecast. It is optimism with cell borders.
The time to decide how AI value will be measured is before the contract is executed, not eleven months later when Finance asks whether the renewal should be approved.
Each use case should have an agreed baseline and measurable outcome. Depending on the application, that could include:
“Employees saved time” may be useful, but it is not automatically a P&L result. If the saved hours do not reduce cost, increase throughput, avoid hiring or improve revenue, they may represent a benefit without representing a financial return.
There is nothing wrong with acknowledging softer benefits. Just do not put a dollar sign in front of them because the steering committee enjoys large numbers.
The AI ROI reckoning should not be viewed as a retreat from innovation. It is the point at which innovation becomes accountable.
Enterprises should continue experimenting, but they should stop confusing experimentation with transformation. They should continue investing, but they should not commit broadly before proving where the technology fits. They should pursue adoption, but under contracts designed for uncertain usage and rapidly changing pricing models.
NET(net)’s approach to technology value has long centered on right-buying, right-licensing and right-pricing, followed by direct optimization and negotiation. That discipline becomes even more important when supplier economics are changing faster than most annual budgeting processes.
The objective is not to squeeze AI spending until nothing innovative remains. It is to ensure that every license, credit and commitment has a reasonable path to measurable value.
Because the real AI tax is not the price of artificial intelligence. It's the cost of buying it without intelligence.
About NET(net)
At NET(net), we don't just optimize IT investments, we weaponize them for competitive advantage. As the world's leading technology investment optimization firm, we've spent over two decades perfecting the art and science of extracting maximum value from technology supply chains while neutralizing vendor pricing manipulation.
Our battle - hardened methodology has influenced trillions of dollars in technology investments, captured hundreds of billions in documented value, and transformed how enterprises approach every facet of IT spend - from emerging technology such as AI, ML, IoT, RPA, Quantum, and Blockchain, to IaaS, PaaS, and SaaS, to enterprise hardware and software solutions, and professional services arrangements including strategic outsourcing relationships.
We're not consultants who theorize about optimization, we're the specialists who help you devise and execute your strategy. Our proven frameworks turn vendor pricing chaos into strategic opportunity, licensing complexity into competitive advantage, and cost centers into value engines. Whether you're facing an aggressive vendor audit, navigating a forced migration, or simply refusing to accept runaway IT costs, NET(net) delivers the expertise, experience, and execution you need to dominate rather than merely survive.
Founded in 2002, NET(net) has established itself as the essential strategic partner for enterprises and technology providers who demand performance, not promises. We've mastered every major area of IT optimization because we understand that in today's vendor-hostile environment, half-measures guarantee defeat.
Experience the NET(net) advantage. Contact us at info@netnetweb.com, visit www.netnetweb.com, or call +1 (616) 546-3100 to discover how we can transform your technology investments from cost burden to strategic weapon.
Legal Disclaimer: NET(net)'s website, blogs, articles, and other content are subject to NET(net)'s legal terms and are offered for general information purposes only, and do not constitute legal advice. While NET(net) may offer views and opinions regarding the subject matter, such views and opinions are those of the content authors, are not necessarily reflective of the views of the company, and are not intended to malign or disparage any other company or other individual or group. Visit our legal notice page for more information.