The New Economics of Enterprise AI

Most organizations still think about AI as another software purchase. You buy licenses, give people access, and expect costs to grow more or less with the number of users. That model worked reasonably well when AI use meant opening a chatbot a few times a day. It becomes much less useful as employees begin relying on AI throughout the workday and organizations start building assistants, automations, and agents that consume computing resources every time they run.

That shift is already showing up in corporate budgets. According to the research, enterprise spending on large language models tripled over a 12-month period by the end of 2025. In McKinsey's May 2026 Enterprise AI FinOps survey, 93% of qualified respondents said they had exceeded their AI budgets. Roughly one in five respondents in a State of AI research study from Stanford said operating costs had constrained AI use. Those numbers are especially interesting because the underlying cost of AI continues to fall dramatically. Stanford found that the inference price required to achieve roughly GPT 3.5-level performance fell from about $20 per million tokens in late 2022 to $0.07 by October 2024, a decline of more than 280 times.

So, the problem isn't simply that AI is expensive; it’s that we are using much more of it.

“The problem isn’t simply that AI is expensive; it’s that we are using much more of it.”

That distinction matters because it changes the budgeting conversation. Rather than asking which AI platform everyone should use, organizations need to start thinking about where AI creates value, how much capacity different types of work require, and where additional investment is justified. Legal departments already make these kinds of decisions with outside counsel, software, staffing, training, and specialized technology. AI is becoming another business resource that needs to be allocated intentionally.

AI Does Not Scale Like Traditional Software

Traditional software licensing gives management a relatively clean relationship between headcount and cost. Buy 500 licenses and, barring a contract change, you have a reasonable sense of what 500 users will cost.

AI complicates that equation because two employees with the same enterprise subscription can consume very different amounts of capacity. One may use AI a few times a day to draft emails or summarize documents. Another may spend hours working with uploaded files, custom assistants, deep research, or agentic features. The license may be the same, but the underlying computing demand is not.

Agents widen that gap considerably. Gartner estimates that agentic systems can require five to 30 times more tokens per task than a standard generative AI chatbot. The reason is fairly straightforward. An agent may make repeated model calls, carry context forward, invoke tools, generate reasoning, check its own work, and retry failed steps. BCG describes one consequence particularly well: under some agent structures, a session that feels twice as long can cost roughly four times as much because growing context is processed again and again.

This helps explain the apparent contradiction in AI economics. Token prices are falling, but total AI spending is rising. The research puts it succinctly: token prices are deflationary while AI consumption is hyperinflationary.

That makes AI less of a licensing problem and more of a resource allocation problem.

Revisiting the Bell Curve

If you've followed my work, you've probably seen me use a simple bell curve to talk about AI adoption. I think the same general framework is useful for thinking about AI investment.

The percentages are not meant to be precise measurements. They are a planning tool. Most organizations have a relatively small group of light users, a large middle group using AI regularly for professional work, and another smaller group whose work requires significantly greater AI resources. Those groups create value differently, and they should not all be provisioned the same way.

I think of them as Everyday Users, Productivity Users, and AI Power Users.

The goal is not to give everyone the same AI. The goal is to give people enough capability to do their work well without paying for capacity they do not need.

Everyday Users: Bottom 10%

Every organization has people who simply want AI to make routine work a little easier. They summarize a document, improve an email, or use an AI feature built into software they already know. They are not creating custom assistants or experimenting with agents, and there is little reason to push them in that direction.

For this group, simplicity is an advantage.

Use the AI already included in the applications the organization has licensed. Microsoft 365, Google Workspace, Adobe Acrobat, Zoom, and many other enterprise platforms now include AI features that are more than capable of handling occasional administrative tasks.

From a cost perspective, organizations should rarely be paying separate usage-based token costs for this group. There is little business case for building custom infrastructure around someone who needs AI a few times a day. In many cases, the better strategy is simply to make sure employees know which approved AI features are already available and how to use them safely.

The business objective at this level is modest but useful: remove small points of friction without adding unnecessary technology, training, or cost.

Productivity Users: Middle 80%

This is where I think most legal organizations should concentrate their AI strategy.

These are the professionals who are integrating AI into daily work. They are using it to work through and automate administrative tasks that previously consumed a surprising amount of time.

The models have also become good enough that I don't think most organizations need to obsess over finding the single "best" one. For mainstream legal and business work, several mature enterprise models can perform extremely well. The research shows why chasing the frontier can become expensive quickly. BCG estimates that simple models and frontier models can differ by roughly 5 to 25 times in token cost, while the broader research repeatedly concludes that the most economical choice is usually the cheapest sufficient model, not automatically the most capable one.

For legal departments, there is another reason to be cautious about living on the frontier. Models that have been available for a while are generally better understood, easier to evaluate, and less novel from a governance perspective. The newest release may be more impressive on a benchmark, but that does not automatically make it the better enterprise choice. Tested models bring less novel risk into the organization.

I have also seen a pattern emerge with clients that has changed how I think about standardization. More than one organization has moved from ChatGPT to Claude with the intention of consolidating on a single platform, only to reactivate ChatGPT later and use both. It was not because Claude had failed. As employees became more sophisticated users and began relying on features such as Projects for sustained work, some regularly encountered usage limits. What initially looked like unnecessary duplication turned out to provide useful capacity and flexibility.

I think there is a broader lesson in that experience. Choosing a primary enterprise platform still makes sense, but organizations should be cautious about turning standardization into dependence on a single provider. Standardize governance, security expectations, training, and acceptable use policies. For regular AI users, access to two approved enterprise platforms can provide valuable flexibility and help keep work moving when capacity, performance, or availability becomes an issue.

There is also a practical cost advantage. Rather than immediately putting this group onto metered API usage, let them work within the capacity already included in enterprise or business subscriptions. Teach them to manage that capacity. If they hit a limit in one platform, they should know how to move to the other without losing half a day of work.

Think of it as basic business continuity for AI. These model providers are still operating in an unusually fast development environment. Features change, service interruptions happen, usage limits shift, and a model that performs beautifully one week can occasionally behave differently the next. Depending entirely on one provider creates a single point of failure that is increasingly difficult to justify for people who rely on AI every day.

The focus for this middle group should remain practical: automate administrative pain points surrounding professional work. The objective is not to automate legal judgment. It is to reduce the amount of professional time lost to organizing, searching, drafting, summarizing, formatting, and other work that AI can increasingly handle well.

AI Power Users: Top 10%

The final group is smaller, but its economic impact can be much larger.

This group includes people who naturally push technology further. In legal organizations they may work in legal operations, innovation, knowledge management, IT, or within a practice group. Some will code, but many will not. They may build custom assistants, connect AI to workflows, experiment with agents, automate repetitive processes, or work with technical teams to create custom applications.

But building technology is not what defines this group. It also includes professionals whose work is inherently AI-intensive, even if they are not building custom tools. Deep legal research, large-scale document analysis, investigations, due diligence, and similar work can require AI to process enormous amounts of information. Some work also depends on specialized legal or proprietary data that sits behind a paywall and is not available through a general-purpose LLM. In those cases, the right investment may be a professional legal AI platform that combines strong AI capabilities with access to the underlying content.

What puts these users in the top tier is not technical sophistication. It is the level of AI capability, data access, and computing resources required to do their work effectively.

For those building custom AI solutions, the economics change again. The organization is no longer simply buying access to an AI product. Custom assistants, agents, automations, and applications may rely on APIs where the organization pays directly for the tokens and computing resources consumed every time the solution runs. That cost needs to be part of the project plan from the beginning.

Gartner's estimate that agentic tasks can consume 5 to 30 times more tokens than traditional chat provides some perspective on how quickly usage can grow. The same principle applies directly to legal agents. A contract workflow used ten times a month has very different economics from one used thousands of times across an enterprise. An agent that completes a task in three model calls has different economics from one that requires twenty calls, repeated retrieval, retries, and long outputs.

This is where organizations need to stop thinking in terms of token price alone. One of the strongest lines in the research comes from McKinsey: “Tokens are not value; tokens are the bill.” The better business measure is cost per successful outcome. What did it cost to review the contract, resolve the request, complete the analysis, or execute the workflow? With that in mind, legal organizations should ask the same questions they would ask about any other investment before building a custom AI tool: What problem are we solving? How many people will use it? How often will it run? What is the likely operating cost? What happens if usage grows tenfold? And does the value created justify the ongoing expense?

“Tokens are not value; tokens are the bill.”

Token forecasting does not need to become an accounting exercise for every lawyer. But someone should understand the expected cost of operating the tool before it becomes embedded in a business process.

A Few Principles I Would Carry Forward

Taken together, the bell curve suggests a fairly simple AI investment strategy. Use included AI for light users. Give regular professional users access to two mature enterprise platforms and let them work primarily within subscription capacity. Reserve higher-cost AI investments for the smaller group whose work requires them, whether that means specialized legal AI tools for deep research and access to proprietary data, or metered API spending and more sophisticated infrastructure for those building applications, agents, and automations. In either case, the additional investment should be tied to a clear business need and expected return.

Flexibility also has economic value. Model prices, capabilities, and usage limits are changing quickly. The research recommends maintaining at least two viable model or provider paths for economically important workloads because systems tied too tightly to one model can become technically or economically outdated within months. That does not mean organizations need a sprawling collection of AI tools. It means avoiding dependencies that make it unnecessarily difficult to adapt as the technology and economics change.

The place to standardize is governance. Organizations should establish clear rules around how AI is used, how data is protected, which tools are approved, how those tools are evaluated, and how business value is measured. They should be much more cautious about standardizing every employee and every workflow onto one provider simply for the sake of uniformity.

That is why I think the next phase of enterprise AI will be less about selecting models and more about managing resources. Organizations that understand where AI creates value, give different users the right level of access, maintain enough flexibility to adapt as the market changes, and spend heavily only where the return justifies it.

Technology may be getting cheaper, but using it well is becoming a management discipline.