The first AI invoice is usually small. The second one is confusing. The third one arrives after a workflow has become part of daily operations, and nobody can explain why a task that cost a few cents in a demo now costs more than the human process it replaced. This is not a pricing problem. It is a measurement problem.
Model cost is only one line in the real unit cost. A production workflow also pays for retrieval, tool calls, evaluation, observability, failed attempts, human review and the support team that fixes edge cases. If those costs are not in the model, the business is making a procurement decision on incomplete information.
Where things stand in 2026
AI spending is rising faster than most operating budgets. Gartner expects worldwide IT spending to reach $6.37 trillion in 2026, up 14.2 percent, with AI infrastructure and software among the fastest-growing segments. McKinsey's 2026 survey finds widespread AI use but much weaker evidence of bottom-line impact; cost pressure is one reason. When a workflow is cheap to try but expensive to run, it often stays a demo.
The practical response is to treat an AI workflow like any other production service. Give it an owner, a budget, a unit-cost metric and a review cadence. The vendor's price per million tokens is an input, not the output.
2027: route, cache and set guardrails
2027 is the year to stop sending every task to the largest model. Build a routing layer: simple classification and extraction tasks go to a smaller model, complex reasoning goes to a larger one, and repeatable answers come from a cache or a deterministic rule. This is not just a cost tactic; it usually improves latency and reliability.
Set budgets at the workflow level, not the company level. A support-triage workflow should have a monthly ceiling and an alert at 70 percent. A document-extraction workflow should have a per-document ceiling. If a workflow cannot be bounded, it is not ready for production.
- Measure tokens per completed task, not per request. Retries and failed tasks are part of the cost.
- Log tool calls. A search API, a map lookup or a payment check can cost more than the model call.
- Cache safely. Cache stable answers and retrieved documents, but respect permissions and freshness.
- Review prompts like code. A prompt change can double token use without changing the output quality.
2028: agentic workflows multiply the cost surface
By 2028, more workflows will use agents that plan, call tools and retry. That is useful, but it multiplies the cost surface: a single user request can trigger several model calls, a retrieval step, two tool calls and a verification step. The bill is no longer proportional to the number of users. It is proportional to the number of decisions.
This is where chargeback or showback becomes useful. Give each department a monthly view of its AI consumption and the outcomes it bought. Do not punish teams for using the service; make the cost visible next to the value. A sales team that spends more but shortens quote turnaround by a day may be making a good trade. A support team that spends more and escalates more is not.
Commitments and discounts start to matter, but only after usage is stable. Committing to a volume you cannot forecast is how finance loses trust in the program. Keep a portion of usage on flexible pricing until the seasonal pattern is clear.
2029: standardize, then negotiate
2029 is a consolidation year. Most SMEs will have too many AI vendors, too many model providers and too many overlapping tools. Standardize on a small number of approved models and a common gateway for logging, permissions and cost control. Then negotiate from evidence: actual volume, growth forecast, latency requirements and switching cost.
Open-weight and smaller models become more credible for bounded tasks. They are not automatically cheaper; hosting, operations and evaluation have costs. But they can reduce dependency on a single provider and make data residency easier. Run a side-by-side evaluation on your own cases before moving a production workflow.
Cost per resolved outcome is the metric to defend. For support, that means cost per resolved ticket. For finance, cost per matched invoice. For sales, cost per qualified opportunity. The model bill divided by tasks is interesting; the model bill divided by outcomes is the business case.
2030: price the outcome, not the seat
By 2030, buyers should expect more outcome-based pricing, especially for narrow, measurable workflows. That can align incentives, but it needs a clear definition of "resolved", a dispute process and a floor for accountability. A vendor that is paid only when a ticket is resolved may close tickets too quickly. A vendor that is paid per seat may resist automation. The contract has to measure the right thing.
Internally, AI cost becomes part of technology value management. FinOps teams already manage cloud, SaaS and licensing; AI adds a usage-based layer that behaves more like cloud than like software. The mature operating model reviews cost, risk and value together, monthly, with the business owner in the room.
What to do in the next 90 days
- Pick one production workflow and calculate its full unit cost, including retrieval, tools, review and failures.
- Add cost tags to every model and tool call: workflow, owner, environment and outcome.
- Set a monthly budget and a 70 percent alert for that workflow.
- Test a smaller model or a cached answer on 100 real cases and compare quality and cost.
- Write a one-page cost review that shows spend, volume, unit cost and outcome quality.
How to measure progress
- Cost per successful task: total cost divided by completed, accepted tasks.
- Tokens and tool calls per task: the main drivers of variable cost.
- Retry and failure rate: hidden cost that shows up as a worse customer experience.
- Forecast accuracy: actual spend versus the monthly budget.
- Gross margin by workflow: the number that tells you whether to scale, fix or stop.
Cost control is not about buying the cheapest model. It is about knowing which work is worth doing, which model is good enough and which failures are expensive. That discipline will matter more in 2030 than any single price cut.
Where this fits in our work
If you want help applying this to your business, start with Business Automation. We publish scope and starting prices on the pricing page. For the next step in this series, read AI adoption for SMEs: 2027, 2028, 2029, 2030.
Sources
- Gartner, "Worldwide IT Spending to Grow 14.2% in 2026", 27 July 2026.
- Gartner, "Worldwide IT Spending to Grow 15% in 2027", 21 October 2026.
- McKinsey, "The State of AI: Global Survey 2026", 2026.
- Flexera, "2026 State of the Cloud Report", 18 March 2026.
- FinOps Foundation, accessed 16 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.