Opinion
AI agents are changing the software business model. Enterprise finance must catch up
Gartner estimates AI coding costs could surpass average developer salaries by 2028, with some companies already seeing more than $2,000 per developer per month.
Most model launches are treated as another round in the AI arms race. Enterprises should read this one differently: the work is moving. The Gartner numbers are the warning sign. They show what happens when AI moves from answering questions to performing work, and when software shifts from something companies license by seat to work they consume one action at a time.
Everyone keeps telling me AI cost is impossible to control because it moves too fast. I don’t really buy it. Cloud moved fast too. A runaway cluster could eat a weekend budget years before anyone called a workflow an agent. Speed makes the bill arrive faster, but it does not explain what the bill means.
What is actually different now is where the work lands. One agent, doing one task, can spend in two places at once. It calls a model, so it shows up on the AI bill. It also touches cloud systems to do the real work: compute, storage, data, logs and calls to other tools. The engineering team sees one job. The model provider sees tokens. The cloud provider sees infrastructure. Nothing connects those pieces by default. In practice, the surprise often comes from unchecked agent runs, oversized context windows and retry loops that look small until they are added up.
That is why the first surprise often arrives as a bill, but the harder problem is not the bill itself. It is the missing link between work, ownership and outcome. A coding agent may feel fast inside a sprint. A support workflow may look successful because usage is growing. A product feature may increase engagement. None of that tells the company which team created the cost, which product or customer it served, or whether the result was worth it.
This is where systems built for the last decade of enterprise software start answering the wrong question. License trackers, procurement calendars and quarterly true-ups were built for seats, contracts and predictable renewals. Agents do not behave that way. They run on demand. They call tools. They retry. They carry context. They move through several systems before a single task is done. Stretching old software-license logic over agentic work is not enough, because the old logic was built to answer a different question.
Cloud already taught companies that fast-moving infrastructure can still be governed. At first, companies stared at AWS bills as if the total was the answer. Over time, they learned that the useful questions were messier. Shared services had to be split. Untagged resources had to find an owner. A cost could be justified for one product and pure waste for another. The bill was not the story. The work behind the bill was.
AI adds a second cost surface on top of that problem, and it is more opaque than the first. The market is still spending too much attention on model price. Model price matters, but it is not the unit the business buys. The business buys a resolved ticket, a merged code change, a cleaner forecast, a fraud review, or an internal workflow that used to take people hours.
A cheap model can become expensive if it needs five retries and carries half the codebase in context. An expensive model can be cheaper overall if it finishes cleanly, uses fewer steps and avoids rework. Looking only at token price is like judging cloud efficiency by the cost of an instance without asking what workload it served.
Anthropic’s Claude apps gateway is one sign of the same shift. It turns Claude Code from a tool developers run individually into infrastructure an organization can govern. That is useful, but it also shows the limit of tool-level controls: a company still has to connect AI usage to the workflow, team, product and outcome behind it.
The same pattern will repeat across the AI stack. Each vendor will expose a different dashboard, billing language and control surface. The business will still need one answer: which work created the cost, who owns it and whether the result justified the spend.
The next management question is not how many people used AI. Adoption was a useful signal during the pilot phase. It told companies that employees were curious and tools were spreading. In production, adoption is incomplete. It does not tell you whether the work improved output, reduced support load, increased quality or quietly turned a good product margin into a bad one.
This is already visible in software products that add AI features. Usage can look great for two quarters, then someone runs the gross margin math and the room goes quiet. The average user may be fine while the heaviest users consume enough inference, storage or agent time to erase the economics of the feature. Seeing the cost is only the starting point. The harder questions are pricing, packaging and ownership. If the AI work is bundled into the base subscription, someone eventually has to answer whether the price reflects what the feature really costs to deliver.
The same issue is coming to internal productivity. A development team may feel faster with coding agents, but finance cannot manage “feels faster.” It needs to know which teams are consuming the budget, which workflows changed, which projects benefited and which usage patterns are just expensive loops. A support team may automate more interactions, but the useful unit is not the number of AI conversations. It is cost per resolved case, time to resolution and quality of the answer. The work unit matters more than the tool name.
This is where the discipline that grew up around cloud cost moves next. It started by turning cloud invoices into accountability: attribution, budgets, forecasting, anomaly detection and unit economics. AI does not replace that muscle. It makes the muscle more important because the cost moves faster, crosses more systems and lands closer to the product itself. Tokens are part of the story, but tokens alone do not answer the business question. The answer has to connect model usage, cloud usage, team ownership, product context and outcome.
The companies that manage AI well will be the ones that see beyond the best model or the lowest token rate. They will know which agent runs deserve more investment, which ones need a different model, which ones require policy, and which ones should be stopped because the economics do not work. They will make AI easier to use where it creates leverage and easier to stop where it adds cost without changing the outcome.
The first wave of enterprise AI was about access. The second wave is about accountability. The winners will be the companies that can look past the model invoice and say, with confidence, what each AI task cost and what it gave back.
Roi Ravhon is Co-founder and CEO of Finout and serves on the Governing Board of the FinOps Foundation.
