How should companies calculate AI ROI? OpenAI has a new answer.

In OpenAI's newly released "Managing AI Investments in the Agentic Era" guide, it proposes a core metric: Useful Work per Dollar. This metric signals that enterprise AI adoption has officially moved from the "pilot experimentation" phase to the "scaled ROI accounting" phase.

Why Old Metrics Failed

For the past two years, companies calculated AI costs using basically two numbers: "how much did API calls cost" and "how many people did AI replace." OpenAI argues both metrics break down in the agent era.

First, call costs. In the single-turn conversation era, costs were indeed linear with usage — you ask a question, the model answers once, you pay once. But agents are different: an agent completing one task might call the model a dozen or even dozens of times, with reasoning, trial and error, tool calls, and self-correction in between. Looking only at per-call costs makes agents "seem expensive"; but if you look at completed tasks, per-unit output costs can be much lower than single-turn conversations.

Second, "headcount replacement." This framing is even more problematic. AI isn't here to "replace people" — it's here to "amplify people." A customer service AI doesn't replace three agents; it lets the remaining agents handle more complex issues, increases CSAT, reduces churn — value that "how many people replaced" can never capture.

The core idea behind OpenAI's "useful work per dollar" is: don't look at how much you put in — look at how much business-value work actually got done. For a sales agent, don't calculate API costs; calculate how many qualified proposals it generated for the sales team per month, how many dormant leads it followed up on, how much it shortened sales cycles — then convert those outputs to dollar value, divide by investment, and that's your real ROI.

Three Phases of Scaling

The most practically useful part of the guide is OpenAI's breakdown of enterprise AI adoption into three phases.

Phase 1: Pilot validation (0-6 months). Don't chase ROI at this stage. The goal is "finding scenarios that work." Pick 1-3 high-value, low-risk workflows (customer support Q&A, document summarization, coding assistance), run them at small scale, prove AI actually delivers value. The worst mistake here is "boil the ocean" — trying every scenario means no scenario gets done deeply.

Phase 2: Productivity multiplication (6-18 months). The focus here is "turning point value into process value." Don't let employees "use it if they want" — embed AI into formal workflows. For example, mandate that AI writes the first draft of proposals in the sales process; have AI handle 80% of common customer issues before escalation. This is where ROI starts showing up, typically 3-5x return on investment.

Phase 3: Workflow reengineering (18+ months). This is where value explodes. You're not bolting AI onto existing processes — you're redesigning entire workflows around AI. Where an analyst used to produce one industry report a week, with a team of agents they can produce three a day covering more dimensions. ROI here can hit 10x or more — but only if you're willing to restructure business processes rather than just slapping AI patches on old ones.

The most valuable insight in this guide is that it punctures a common corporate illusion: "buy an LLM API, give employees Copilot accounts, and that's AI transformation." It's not. Buying tools is step one. The real challenge is whether you're willing to redesign workflows, adjust org structures, change performance metrics — those "soft" problems are the real bottleneck to AI adoption.

Models will get cheaper, agents will get more capable — those are certain trends. What's uncertain is whether your organization will be ready on the organizational and process side before that trend arrives.

See you tomorrow.