The $11.5 Million Question: Why AI Spending Keeps Climbing While ROI Stays Invisible

Enterprises spent an average of $11.5 million on AI in 2026, yet most can't prove a single dollar came back. The move from chatbots to autonomous agents is the main cost driver: token usage is projected to grow 24-fold in four years and 55-fold by 2040. A simple $0.04 chat can balloon into a $1.20 orchestration. To survive, companies are metering tokens, using smaller models, and tying spend to budgets.

I've watched this pattern repeat across a dozen AI budgets now, and the uncomfortable truth is simple. The bill scales faster than the model providers cut prices, and faster than anyone can show a return. Let me walk through the numbers, because they tell a sharper story than the vendor decks do.

The Real Cost Driver Is Agents, Not Chat

The first wave of enterprise AI was cheap because it was dumb. You typed a question, a model answered, and the meter barely moved. Those days are ending.

The shift now is from single chatbots to autonomous agents that plan, retrieve tools, and spawn subagents to finish a task. Each of those steps burns tokens, and the steps multiply.

Here's the figure that should reset your forecasting. A $0.04 chat can become a $1.20 orchestration once it requires tool retrieval, planning, and subagents. That's a 30x jump for what looks, to the end user, like the same request.

Now multiply that across an organization. 18% of organizations are now orchestrating multiple agents across workflows, up from 9% in the prior period. Adoption doubled. The per-task cost went up by an order of magnitude. You can see where the budget line is heading.

Why price cuts won't save you

The common rebuttal is that token prices keep falling, so volume growth washes out. It doesn't.

Usage is projected to rise 24-fold within four years and 55-fold by 2040. Provider price reductions are real, but they don't move at that speed. When consumption climbs 24x and prices drop maybe 2x or 3x over the same window, the net direction of your invoice is up. Way up.

The intensive compute demands of agents are the reason. A reasoning loop that calls three tools and verifies its own output isn't a slightly bigger chat. It's a different cost class entirely.

The Coding Bill Is About to Pass a Human Salary

If you want one prediction to put in front of your CFO, use this one. AI coding costs are expected to surpass the average developer salary by 2028.

Sit with that. The tooling sold as a way to make engineers cheaper may, on a per-seat basis, cost more than the engineer. Token consumption in code generation is brutal because the models read large contexts, draft, test, and rewrite, often several times per task.

This doesn't mean AI coding is a bad investment. It means the "it's basically free" assumption that justified the pilot is dead, and the business case has to be rebuilt on real throughput numbers, not vibes.

The $11.5 Million You Can't Account For

Now the part nobody wants on the quarterly slide. Enterprises spent an average of $11.5 million on AI in 2026, and most can't demonstrate a clear return.

Not a vague return. A clear one. The disconnect between spend and measurable outcome is the defining problem of this cycle.

A few things are true at once here:

• Some companies have optimized specific workflows. Payment processing and research are the two I see cited most, and the gains there are genuine.
• Most initiatives lack a quantified ROI number anyone outside the AI team trusts.
• The macroeconomic data still hasn't shown the productivity surge the spending implied. If everyone got this much more productive, it should be visible by now. It isn't.

The result is predictable. Investors have stopped accepting narrative and started demanding evidence. "We're seeing strong adoption" no longer clears the bar. They want a dollar figure that came back in.

Why the ROI is hard to find

Part of the problem is measurement, and part of it is real. On the measurement side, most teams never set a baseline before deploying, so they have nothing to compare against. You can't prove a 20% improvement if you never recorded the starting point.

On the real side, a lot of agent deployments automate work that wasn't expensive to begin with, while the agents themselves are expensive. You can spend $1.20 to save 90 seconds of a task no one was paying much for. The orchestration is impressive. The economics are upside down.

What the Disciplined Companies Are Doing

The firms keeping this under control share one habit: they treat tokens like cloud spend, with the same rigor FinOps brought to AWS bills a decade ago. Monitoring token usage is now essential, not optional.

Here is what the better operators are actually doing, with names attached.

1. Real-time dashboards. Priceline runs live visibility into token consumption so cost spikes get caught in hours, not at month-end.
2. Automated alerts. Smartsheet sets thresholds and fires automated warnings as usage approaches a limit, before the overage lands.
3. Chargeback models. Qualcomm and OpenText link AI usage back to departmental budgets, so the team spending the tokens owns the bill.
4. Smaller models by default. Instead of routing everything to the most capable (and most expensive) model, they reserve the big models for tasks that genuinely need them.
5. Deployment tied to business goals. Every agent has to map to an outcome someone will defend, which kills the vanity pilots early.

The through-line is transparency and accountability. When a department sees its own AI invoice, behavior changes fast. Chargeback is less a finance trick than a forcing function for discipline.

The smaller-model lever is underused

Of these, model right-sizing is the one most teams skip. There's a reflex to use the strongest model "to be safe," and it's expensive insurance.

Most production tasks (classification, extraction, routine drafting) run fine on smaller, cheaper models. Routing logic that sends only the hard 20% of requests to the frontier model can cut a bill substantially without touching the quality your users actually notice.

How to Read This If You Own the Budget

I'm not in the camp that says enterprise AI is a bubble with nothing inside it. The optimized workflows are real, and the agent capabilities are genuinely new. But the spending pattern right now is a setup for a hard correction.

My practical take:

• Instrument before you scale. If you can't see token usage per team and per workflow, you're flying blind into a 24x growth curve.
• Set a baseline first. No metric before deployment means no ROI story after. This is the cheapest fix available and almost nobody does it.
• Assume the per-task cost will rise, not fall. Plan budgets against agent economics, not chatbot economics.
• Kill pilots that can't name an outcome. If a deployment can't point to revenue, cost saved, or a defended quality gain, it's a science project, not an investment.

The companies that come out of this cycle ahead won't be the ones who spent the most. They'll be the ones who could prove where the money went and what came back.

FAQ

Why is AI ROI so hard to prove despite heavy spending?

Enterprises spent an average of $11.5 million on AI in 2026, but most never set a baseline before deploying, so they have nothing to measure improvement against. Many initiatives also automate low-value tasks at high per-task cost, and the broader macroeconomic data has not yet shown the expected productivity surge.

How much are AI costs expected to grow?

Token consumption is projected to rise 24-fold within four years and 55-fold by 2040. This growth is driven by autonomous agents and outpaces the price reductions offered by model providers, so most businesses should expect rising overall AI expenditures.

Why do AI agents cost so much more than chatbots?

A single chatbot reply burns few tokens, but an agent plans, retrieves tools, and spawns subagents, with each step consuming more tokens. A request that costs $0.04 as a chat can cost $1.20 as a full orchestration, roughly a 30x increase for the same user-facing task.

What is a chargeback model for AI costs?

A chargeback model links AI token usage back to the budget of the department that generated it. Qualcomm and OpenText use this approach so teams own their own AI bills, which improves transparency and accountability and discourages wasteful usage.

Will AI replace developers if coding costs are rising?

Not straightforwardly. AI coding costs are expected to surpass the average developer salary by 2028, which undercuts the assumption that these tools are nearly free. The business case has to be rebuilt on measured throughput gains rather than the assumption of cheap automation.