OpenAI published a piece on July 14, 2026 titled "How to manage AI investments in the agentic era." It lands on a real problem. As teams move from chat prompts to agents that run for minutes and call tools on their own, the bill stops behaving like a normal software line item. This article is our read on that playbook: what holds up, what is vendor framing, and what a practitioner running the budget should actually do. Managing AI investments well in 2026 is less about chasing the lowest token price and more about knowing what work each dollar buys.
Some scene-setting first. Gartner has forecast global AI spending north of $2 trillion for 2026, and the FinOps Foundation’s State of FinOps 2026 survey found that 98% of practitioners now manage AI spend, up from 63% a year earlier. The money is real and the governance is immature. OpenAI’s article is one vendor’s attempt to give leaders a frame for it, and the frame is mostly sound. It is worth reading past the product placement to the useful parts.
Why the agentic era breaks old budgeting
A chat request is cheap and bounded. You send a prompt, you get a response, and the cost is roughly predictable. An agent is neither cheap nor bounded. It plans, calls tools, reads results, and loops until it decides it is done. Each loop adds tokens, and context accumulates as the run goes on. Analysts put the multiplier all over the map. EY and McKinsey have described agentic tasks consuming many times more tokens than a single chat, and some vendor estimates run to 50 times or more. The exact multiple depends on the workload, so treat any single figure with suspicion, but the direction is not in dispute. If you want the mechanics of why longer runs cost more, our explainer on input versus output tokens covers how generation dominates the bill. And if the word "agent" still feels fuzzy, start with what AI agents actually are.
OpenAI’s core idea: useful work per dollar
OpenAI argues that token price alone tells you nothing about value, and that leaders should measure "useful work per dollar" instead: tasks completed, time saved, decisions improved. That framing is correct, and it is not unique to OpenAI. EY and McKinsey talk about cost per outcome and a "return on AI" ratio. The FinOps community frames it as connecting spend to business value, the hardest part of the job. When a vendor and its critics agree on the metric, that is a good sign the metric is real.
The practical version: for a priority workflow, pick the unit of work that matters, such as a resolved support case or a code change that passes review, then track cost per accepted outcome rather than cost per token. A cheaper model that fails and retries can cost more per finished job than an expensive model that gets it right the first time. That reframe is worth keeping no matter which vendor you buy from.
The five steps, translated
OpenAI lays out five steps. Stripped of the product tour, they read like this.
Get visibility first. You cannot manage a bill you cannot break down. Know who is using AI, which models, on what work, and whether spend and adoption are moving together. This is FinOps applied to tokens, and it is the right starting point.
Judge models by outcome, not sticker price. Define "good enough" before you test, build evals from real tasks, then measure the full cost of reaching that bar including retries and human review. Reserve frontier models for hard, high-stakes work, and use smaller models wherever they clear the bar.
Govern workflows before they scale. Decide what context an agent can touch, which tools it can call, what actions it can take, and who approves the risky steps. OpenAI is right that governance is the layer that decides what can safely scale. This is also where its own ChatGPT Work admin controls get a plug, so read it with that in mind.
Fund like a portfolio. Broad access for everyday productivity, function-specific workflows for repeatable work, and a small number of strategic bets built on proprietary context. Fund by maturity: exploration first, then validation against a quality bar, then production funding for the integrations and controls that scaling needs.
Match capacity to proven demand. Once a workflow earns its keep, put it on the commercial structure that fits its usage pattern rather than making every team rebuild its own plumbing.
None of this is wrong. The gap is what OpenAI leaves out, and where its recommendations conveniently point.
Where DM parts company with the pitch
This is a vendor document, and the seams show. Every step routes to an OpenAI product: ChatGPT Work for governance, Guaranteed Capacity and Scale Tier for capacity, OpenAI Frontier and Deployment Engineers for the big bets. The governance principles are portable. The plumbing is not, and a portfolio that is all one vendor is not really a portfolio.
Three things a practitioner should add to OpenAI’s list.
First, install real circuit breakers. OpenAI mentions spend controls, but it soft-pedals the blunt instrument: hard spend ceilings, call-volume caps, and automatic shutoffs that stop an agent gone into a loop. Without a kill switch, a runaway run is only visible after the invoice arrives. This is the single most common gap in early agent deployments, and it belongs in step one, not as an afterthought.
Second, treat build versus buy as a per-layer decision, not a vendor loyalty test. The independent consensus on portfolios is to buy commodity capabilities, partner for speed, and build only where you get durable differentiation. That calculus should run separately for models, orchestration, retrieval, and evaluation. OpenAI’s version of "portfolio" is a portfolio of OpenAI products. The real version routes across models and may keep some work in house, and multi-model routing is itself a cost lever OpenAI has little reason to highlight.
Third, watch the pricing model shift, because it changes how you budget. The agent economy is moving toward usage and outcome pricing. Intercom’s support agent, for example, charges per resolved conversation rather than per seat. Outcome pricing aligns cost with value on paper, but it also makes spend less predictable and harder to cap, which loops right back to the need for circuit breakers. Budget for variability, not a flat monthly number.
One more caveat on the numbers. OpenAI opens with impressive efficiency stats, such as a 97% drop in price per million tokens from GPT-4 to GPT-5.4. Those are OpenAI’s own figures, from a benchmark it selected, and we have not reproduced them. Falling unit prices are real and welcome, but they are also the framing a seller wants in front of a buyer. Efficiency per token does not settle whether a workflow is worth funding.
What to actually do this quarter
If you own an AI budget, the honest starting list is short. Get usage broken down by team, model, and workflow. Pick two or three priority workflows and define cost per accepted outcome for each. Put hard spend ceilings and shutoffs on anything agentic before it goes wide. Decide build versus buy per layer rather than per vendor. Treat OpenAI’s article as a useful checklist whose conclusions happen to sell OpenAI. Managing AI investments is finally more a measurement discipline than a procurement one, and our guide to AI implementation sets the foundation this budgeting work sits on top of.
Frequently Asked Questions
What does “managing AI investments in the agentic era” actually mean?
It means governing AI spend when the workload is no longer a single chat request but an agent that plans, calls tools, and loops until it finishes. That behavior makes cost variable and hard to predict, so managing AI investments now centers on visibility, cost per outcome, and controls, not on picking the cheapest per-token model.
What is “useful work per dollar”?
It is OpenAI’s headline metric: judge AI by the tasks it completes, time it saves, and decisions it improves, not by token price. Analysts at EY and McKinsey describe the same thing as cost per outcome, so it is worth adopting regardless of vendor.
Why do AI agents cost so much more than chatbots?
An agent runs multiple steps, calls tools, and carries accumulating context, so one task can consume many times the tokens of a single chat. Estimates vary widely by workload, so treat any single number as a rough signal.
Is OpenAI’s five-step framework worth following?
The principles are sound: get visibility, judge models by outcome, govern before scaling, fund like a portfolio, and match capacity to demand. The caveat is that every step points to an OpenAI product, so keep the governance logic and discount the shopping list. Independent analysts reach the same conclusions without the sales pitch.
What is cost per accepted outcome?
It is the total cost of reaching an acceptable result for a defined unit of work, such as a resolved support case or a code change that passes review. It counts retries, tool calls, and human review, not just the model call. A cheap model that fails and retries can cost more per finished job than a pricier one that succeeds first time.
What are AI cost circuit breakers or kill switches?
They are hard limits that stop spend before the invoice does: per-workflow spend ceilings, call-volume caps, and automatic shutoffs that halt an agent stuck in a loop. OpenAI mentions spend controls but underplays these blunt instruments, the most commonly missing safeguard in early agent deployments.
How does outcome-based pricing change budgeting?
Vendors are shifting from per-seat fees toward charging per action, per workflow, or per delivered result, such as a fee per resolved conversation. This aligns cost with value but makes spend less predictable and harder to cap, so plan for variability and pair it with your own spend ceilings.
Should we build or buy our AI stack?
Decide it per layer, not per vendor. The common pattern is to buy commodity capabilities, partner where speed matters, and build only where you gain durable differentiation. Run that judgment separately for models, orchestration, retrieval, and evaluation.
