Tokenomics before automation: how to decide which AI use cases deserve budget
AI automation should not start with "which model should we use?" It should start with a harder question: what operational result are we buying with every token?
In 2026, that question became practical. Uber reportedly capped monthly token spend for agentic coding tools at USD 1,500 per employee per tool after burning through its 2026 AI budget in four months. That is not a retreat from AI. It is a finance discipline around a tool employees clearly wanted to use.
The lesson is simple: AI can increase productivity, accelerate delivery and unlock new products. It can also turn an unclear process into a token-burning machine.
At Enacment we have seen this in live projects. In some cases, AI usage rose up to 670% above the expected budget. The technology was not the only problem. The work entered the system too open: repeated context, unclear output contracts, too much validation churn and weak cost controls. We had to rethink prompts, data architecture, caching, evaluations and scope to bring the work back into budget.
The answer is not "use less AI." The answer is: measure ROI before automation and treat tokens as operational capacity, not cheap magic.
!AI tokenomics, ROI and operational efficiency control room
AI pays off, but not in any order
The right operating sequence is:
1. Well-defined processes. What happens, who decides, what exceptions exist and what "done" means.
2. Governed data models. Where trusted data lives, which entities matter, which fields are reliable and which permissions apply.
3. Stable operating structure. Roles, SLAs, support paths, backlog, acceptance criteria and decision rhythm.
4. Automation with measurable ROI. Only then should the team decide whether to use rules, traditional software, RPA, generative AI, agents, analytics or a combination.
Skip straight to automation and the AI absorbs ambiguity. Ambiguity shows up later as tokens, rework and operational risk.
What the data says
The 2025-2026 story is not "AI does not work." It is more specific: AI works best when connected to real workflows.
The practical read: AI has value. It just punishes poor use-case selection.
The 2x2: operational value vs token complexity
!2x2 matrix to prioritize AI use cases by operational value and token complexity
Automate first: high value, low complexity. Repetitive processes, clear data, low risk and outputs that are easy to validate.
Design carefully: high value, high complexity. Agents, copilots and decision workflows can pay off, but need output contracts, evaluations, governed data, monitoring and support.
Keep manual or semi-automated: low value, high complexity. These cases become expensive through exceptions, poor data, permissions, low frequency or marginal impact.
Experiment cheaply: low value, low complexity. Good for learning, templates and low-risk pilots. Do not sell it as transformation.
The brief before the PRD
Every AI backlog item should enter with a one-page brief:
1. Operational pain.
2. User and accountable owner.
3. Inputs, systems, data and permissions.
4. Expected output.
5. Quality contract.
6. Target cost per run, user or month.
7. Current baseline.
8. Risks.
9. Control model.
10. Go/no-go metric.
If the brief cannot explain ROI in one page, the PRD will hide uncertainty in more words.
Output contracts
Automation output should feel like a contract, not a suggestion. Define exact format, required fields, permitted sources, confidence level, rejection rules, exception handling, traceability, approval owner and versioning.
Without a contract, the model improvises. With a contract, the system can validate.
The budget is more than tokens
Separate AI budget by layer:
When the budget only includes tokens, the project looks cheap. When it includes the full operating system, the right decisions become visible.
Quick guide
1. Define the process and exceptions.
2. Establish an economic baseline.
3. Govern the required data.
4. Calculate target cost per run.
5. Prioritize by value vs complexity.
6. Write briefs before PRDs.
7. Define output contracts.
8. Decide where humans stay in the loop.
9. Separate model, data, app, evaluation and operating costs.
10. Kill weak cases early.
AI does not need to go everywhere. It needs to go where the process, data and economics can defend it.
That is operational efficiency: not doing more things with AI, but doing better the things that deserve AI.