OpenAI Outlines Five Steps for Managing AI Investments in the Agentic Era
OpenAI has released a practical framework to help enterprise leaders manage artificial intelligence investments as organisations move from conversational assistants towards agents capable of completing longer, multi-step tasks.
The guidance reflects a significant change in how businesses need to evaluate AI expenditure. When employees use AI primarily for occasional questions or document drafting, costs can often be managed through straightforward subscriptions and usage limits. Agentic workflows are different: they may run for longer, use multiple tools, access connected systems and perform work on behalf of users.
This creates the potential for greater business value—but also higher and less predictable expenditure.
OpenAI’s framework recommends that organisations improve visibility into usage, measure the cost of successful outcomes, introduce governance before workflows scale, fund repeatable use cases and match computing capacity to proven demand.
Token prices do not reveal the full cost
The cost of accessing leading AI models has fallen substantially. OpenAI says the price per million tokens declined by 97% between GPT-4 and GPT-5.4.
The company also claims GPT-5.6 achieves stronger coding-agent performance while using 54% fewer output tokens and completing tasks in 57% less time.
However, a lower token price does not automatically translate into a lower business cost.
A cheaper model may require several attempts to complete a task, produce work that needs extensive correction or fail frequently enough that employees stop using it. A more capable model may carry a higher price per token but deliver an acceptable result faster and with less human review.
OpenAI argues that organisations should therefore measure “useful work per dollar”. Relevant measures could include:
- Tasks successfully completed
- Employee time saved
- Decisions improved
- Customer cases resolved
- Software changes accepted
- Revenue protected or created
- Risks avoided
- Additional organisational capacity generated
The important metric is not how cheaply an AI system produced an answer. It is how much the organisation spent to achieve an acceptable outcome.
1. Improve visibility into AI usage and expenditure
Enterprise leaders first need a clear picture of how AI is being used across the organisation.
This includes understanding which employees and teams are using AI, what products and models they select, how much capacity they consume and what type of work the usage supports.
A rising AI bill is not necessarily a problem. It could indicate inefficient experimentation, but it might also reveal a valuable workflow becoming central to business operations. Without information about the work behind the expenditure, administrators cannot distinguish between the two.
OpenAI has updated the ChatGPT Admin Console with usage analytics and spending controls that allow administrators to examine adoption, credit consumption and expenditure by user, product and model.
OpenAI suggests analysing this information at several levels:
- Workspace level: Are adoption and expenditure increasing together?
- Team and user level: Which areas are experiencing growing demand?
- Product and model level: Where are more expensive models being used, and is the demand sustained?
- Workflow level: Does the activity represent experimentation, individual power use or a repeatable business process?
This visibility can help leaders decide where to increase capacity, provide training or introduce limits.
2. Measure model efficiency by outcome
Comparing AI models solely by their input and output prices can create misleading results.
OpenAI recommends developing evaluations based on real organisational tasks, including edge cases and the situations most likely to cause failure. The organisation should define what constitutes an acceptable result before testing begins.
The total cost of achieving that result should then include:
- Model and tool usage
- Number of attempts
- Completion rate
- Processing time
- Human review
- Corrections and rework
- Operational failures
For a customer-service workflow, the most useful metric may be the cost of resolving a case successfully. For a software-development agent, it might be the cost of producing a tested change that passes human review.
This approach can produce a different model-selection decision than simply choosing the lowest advertised token rate.
OpenAI also notes that model choice is only one part of efficiency. Clear instructions, focused tools, reusable context and explicit stopping conditions can reduce unnecessary loops and prevent agents from continuing to consume resources after they have completed their objective.
Smaller and faster models may be sufficient for structured, predictable work. Frontier models should be reserved for tasks involving greater complexity, ambiguity or risk.
3. Establish governance before agents scale
Governance becomes more important as AI systems gain access to company information and the ability to act across business applications.
Organisations need to determine:
- What information an AI agent can access
- Which applications and tools it can use
- What actions it can perform
- Which steps require human approval
- How usage is monitored
- How additional capacity is approved
- What information is retained
- How sensitive data is protected
These decisions should be made before an agent becomes widely adopted.
The risks are considerably higher when AI can use connectors, plugins or computer-control capabilities. A poorly governed chatbot may generate an incorrect answer. A poorly governed agent could enter that answer into a business system, send it to a customer or trigger another process.
OpenAI’s ChatGPT Work administration features provide centralised controls for connected tools, approved context, permitted actions, usage and expenditure. Administrators can set workspace defaults, group limits and individual overrides, while allowing employees to request additional capacity with information about the relevant project.
Privacy, security and approval pathways should form part of the original workflow design—not be added after deployment.
4. Fund workflows that can compound in value
OpenAI recommends treating enterprise AI as a portfolio of investments rather than a single technology rollout.
That portfolio could include:
- Broad access for everyday employee productivity
- Function-specific workflows for repeatable departmental work
- Strategic agents built around proprietary company information and processes
The strongest investment candidates are workflows that occur frequently, operate at meaningful scale, have clear ownership and can be measured for quality, risk and value.
Funding should also reflect the workflow’s maturity.
Early exploration should determine whether an AI model is capable of performing the task. Validation should test that capability against representative cases and a defined quality standard. Production investment should cover the integrations, security controls, reliability, training and change management needed for wider adoption.
This staged approach reduces the risk of funding elaborate infrastructure before the underlying business case has been demonstrated.
OpenAI also recommends centrally funding shared capabilities such as identity management, trusted connectors, curated knowledge, evaluations, observability, model routing and reusable agent patterns. Each successful implementation can then make subsequent workflows faster and safer to deploy.
5. Match capacity to demonstrated demand
Once an AI workflow has proven its value, the organisation can select a commercial and technical structure that reflects its actual usage.
OpenAI identifies several options for different workload patterns:
- Guaranteed Capacity for production agents requiring predictable availability
- Scale Tier for consistently high-volume API workloads
- Batch API for large tasks that do not require immediate results
- Flex processing for workloads that can trade processing speed for lower costs
- Prompt Caching for applications repeatedly using the same context
The objective is to avoid purchasing large amounts of capacity before demand is established, while ensuring that business-critical agents have reliable access once deployed.
For larger strategic implementations, OpenAI is also positioning OpenAI Frontier and its Deployment Company as options for enterprises developing AI coworkers that operate across multiple systems.
The shift from licences to AI operating economics
The framework highlights a broader change in enterprise technology management.
Traditional software is often purchased according to the number of users. Agentic AI introduces a more complicated model in which expenditure may depend on the duration and complexity of the work, the model selected, the tools used and the number of attempts required.
Two employees with the same AI licence could consume dramatically different amounts of capacity. One may use ChatGPT for occasional summaries, while another operates an agent that analyses documents, searches connected systems, writes code and completes recurring tasks every day.
Organisations therefore need to connect AI expenditure with business activity and measurable results. Simply reporting the number of active users or tokens consumed will not provide enough information to guide future investment.
What organisations should do now
OpenAI’s framework can be converted into a straightforward operating process:
- Establish a baseline of current AI usage and expenditure.
- Identify workflows generating sustained demand.
- Define an acceptable outcome for each priority workflow.
- Measure the full cost of reaching that outcome.
- Introduce access, approval and expenditure controls.
- Fund reusable capabilities that benefit multiple workflows.
- Increase capacity only after value has been demonstrated.
The organisations that manage this transition successfully will not necessarily be those spending the most on AI. They will be those that understand which workflows create value, apply the appropriate level of intelligence and scale investments using reliable evidence.
As AI agents become more capable, the central financial question is shifting from “How much does this model cost?” to “What valuable work did the organisation receive for the money it spent?”