Stop Measuring AI by Token Price: A Practical ROI Playbook for Business Leaders

01Executive summary

02What happened?

In its new AI investment framework, OpenAI recommends five actions: improve visibility into usage and spend, evaluate models by outcome, govern advanced workflows, fund repeatable work, and match capacity to proven demand.

The timing matters. AI tools are becoming cheaper at the model level while doing longer and more complex jobs. OpenAI says the price per million tokens fell 97% from GPT-4 to GPT-5.4. It also says GPT-5.6 completed a coding benchmark with 54% fewer output tokens and 57% less time per task.

Those are vendor-reported figures, not a promise that every company will save the same amount. The more important point is that token price tells only part of the cost story.

03Why should a business owner care?

A chatbot produces an answer. An AI agent may read a customer request, search a knowledge base, update a CRM, draft a reply, and trigger a follow-up. That creates more value, but it also introduces more places for cost and risk to grow.

If leaders only watch the monthly AI bill, they cannot tell whether higher usage means waste or successful adoption. If they only choose the cheapest model, they may pay more for failed attempts and manual rework. If they scale before adding controls, a small mistake can spread across customers, records, or financial processes.

The business question is no longer “How much does this model cost?” It is “How much does it cost to complete this process at the quality and risk level we require?”

04Measure cost per accepted outcome

Start with one unit of work that already matters to the business. Define what “accepted” means before testing the AI. For example:

• Customer support: a case resolved without reopening.
• Sales: a qualified lead added with complete, accurate data.
• Finance: an invoice processed and approved without correction.
• Operations: an order exception identified and routed correctly.
• Software delivery: a tested change that passes human review.

Then count the full cost: model usage, connected tools, retries, processing time, human review, maintenance, and failures. Compare that with the current process using time saved, faster turnaround, revenue protected, errors avoided, or extra capacity created.

05Use different AI for different work

One model does not need to handle every task. Routine classification, extraction, and drafting may work well with a smaller, faster model. Complex analysis, sensitive decisions, or unclear requests may need a more capable model and human approval.

This is often called model routing: sending each request to the lowest-cost option that can reliably meet the quality bar. It can improve margins and reduce dependence on one provider, but only when supported by real evaluations, monitoring, and fallback rules.

Clear instructions also matter. Give the AI only the tools and context needed for the job. Add limits on steps, time, and spend. Define when it should stop, ask for help, or hand work to a person. Better workflow design can remove waste before you negotiate a lower model price.

06Build governance before scale

Governance is not a document added after launch. It is the practical system that controls what an AI can see and do. Before an agent touches business data, decide:

• Which data and systems can it access?
• Which actions can it take automatically?
• Which actions require human approval?
• How are decisions, tool calls, and costs logged?
• Who owns quality, security, and incident response?
• What happens when the model or a connected service fails?

These controls protect customer trust and make successful workflows easier to expand. Without them, companies either accept unnecessary risk or keep useful automation trapped in permanent pilot mode.

07Where the best opportunities are

Good early candidates are repetitive, high-volume workflows with a clear output and an existing human owner. Customer support triage, document processing, internal knowledge search, lead enrichment, quality checks, and operational reporting often fit this pattern.

Avoid starting with your most sensitive or ambiguous decision. An AI that recommends loan approvals, medical actions, hiring decisions, or large payments needs a much higher standard of evidence and control than one that drafts an internal report.

The strongest long-term opportunities combine AI with proprietary business context: your processes, product knowledge, customer history, and approval rules. That is where custom software can turn a widely available model into a capability competitors cannot copy by purchasing the same subscription.

08A practical way to start today

1. Pick one workflow with enough volume for savings to matter.
2. Record the current cost, turnaround time, error rate, and human effort.
3. Define an accepted outcome and the cases that must always be escalated.
4. Test multiple models and workflow designs on representative work.
5. Run a limited pilot with approvals, logs, budgets, and a named owner.
6. Scale only when quality, cost per outcome, and business value meet the agreed target.

This approach keeps experimentation fast without confusing activity with progress. It also gives leaders evidence they can use to stop, redesign, or expand an investment.