Why do AI tool costs explode after a company wide rollout, and how do you control them?
Costs grow when every connector is on, every agent reads everything and nobody measures the cost of a piece of work. Control them per piece of work, not per seat.
AI costs grow fastest after a rollout because much of the spend is usage based: it grows with every run, not with every seat. Every connector that is switched on, every document an agent reads again and every routine step sent to a model adds to the bill, and nobody sees the total per piece of work. Control the cost the same way you control quality: measure what each piece of work costs, give each agent only the tools and context its job needs, and move fixed steps out of the model.
Why the question comes up
The first months of a rollout look cheap. Licences are a known line, people try things, and the usage bill is small. Then adoption works. People connect more tools, build their own workflows and leave them running. On one call, a mid market software company said it plainly: its costs were exploding after it adopted every connector, and individual workflows were burning credits without anyone noticing.
It is not one company's problem. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and escalating costs are the first reason it lists, next to unclear business value and weak risk controls.
Where the money goes
Four things drive most of the growth.
Everything is connected. Each connector and MCP server adds tool descriptions that a model has to read before it starts. Anthropic describes agents that process hundreds of thousands of tokens of tool definitions before they read the request, and an example where loading only the tools a task needs cut usage from 150,000 tokens to 2,000.
The same context is read again and again. Without a shared, maintained source, every agent and every run searches the same documents from scratch. Large documents pass through the model more than once when an agent copies them from one system to another.
Fixed steps run through a model. A step that runs the same way every time, such as moving a field or sending a reminder on a schedule, costs money every time it goes through a model and is cheaper and more reliable as ordinary software.
Nobody owns the cost of a piece of work. The bill arrives per tool or per account. Nobody can say what one quote, one invoice check or one support answer now costs, so nobody can tell which workflow is worth it.
How to control AI costs
- Measure cost per piece of work. Pick the unit (one quote, one invoice, one answer) and track what it costs to get it done to the standard, including review time. The method is in how to measure AI agent ROI.
- Cut the connector list. Keep a short approved set and give each agent only the tools its job needs. Fewer tools in front of a model means less to read on every run. See how to govern MCP servers and connectors.
- Give agents the smallest view that is still enough. One maintained source that knows which facts are current costs less than letting every agent search everything. More context makes agents worse, and it also makes them more expensive.
- Move fixed steps out of the model. Keep the workflow automations that already work and use an agent only where a step needs judgement. More in keeping workflow automations when you add AI agents.
- Test models against the work. The most capable model is not always needed. Run the same real examples through cheaper models and keep the one that still meets the standard.
- Give every workflow an owner. Someone decides when a workflow runs, what it may cost and when it gets switched off. Every run should be recorded so the owner can see it.
| What drives the cost | What controls it |
|---|---|
| Every connector switched on | A short approved list, scoped tools per agent |
| Agents rereading the same documents | One maintained source with the current facts |
| Fixed steps sent to a model | Ordinary workflow automation for those steps |
| The biggest model for every task | Models tested on real examples, cheapest that passes |
| No cost per piece of work | Cost tracked per task next to quality, time and effort |
| Workflows nobody owns | An owner per workflow, every run recorded |
How Ortelian approaches it
We treat cost as one of the four things every piece of work is measured on, next to quality, time and human effort. Evaluations are built from real examples of your work before launch and keep running after it, so a cost increase shows up per piece of work, not only on a monthly bill.
The platform separates agents, which handle work that needs judgement, from workflows, which run the fixed steps the same way every time. Each agent gets only the tools its job needs, and every run, action and approval is recorded. The world model keeps the facts from your systems with their source, and what one agent learns, the next can use, so the next agent starts from what is already known instead of searching everything again. Ortelian works across AI models and tests them against your work. More on the platform and how we work.
Questions people ask
Why did our AI costs go up after people started using it more?
Much AI spend is usage based. More users, more connectors and more automated workflows mean more runs, and each run reads more context when everything is connected. Seat licences stay flat while usage grows.
Is it cheaper to use one big AI agent for everything?
Usually not. One agent with every tool reads more on every run and is harder to test. Smaller agents with scoped tools, plus ordinary automation for fixed steps, are cheaper to run and easier to check.
Should we limit who can build AI workflows?
Limit what a workflow can reach and make sure each one has an owner, rather than banning people from building. Record every run so the owner can see what each workflow costs and does.
How do you know if an AI workflow is worth what it costs?
Compare the cost per piece of work, including review time, with what the same work cost before. If quality meets the standard and the cost is lower over several cycles, keep it. If not, fix it or switch it off.
Sources
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025
- Anthropic, Code execution with MCP: building more efficient agents, 4 November 2025
- Ortelian, Platform, Trust and How we work
