Is AI Spending Worth It: OpenAI's Framework for Enterprises to Measure Value

Once ChatGPT and Codex are rolled out across a company, team leads, IT administrators, and business executives will inevitably face the same question from finance and executive management: Is the AI budget worth the spend?

How many accounts have been provisioned or how many credits have been consumed only shows that the tools are being used; it cannot answer whether these investments have led to faster delivery, higher quality, or revenue growth. On September 16, 2026, OpenAI published a guide explaining how enterprise administrators can use the analytics features in the ChatGPT Admin Console to see what teams are doing and where spending goes across ChatGPT Work and Codex, and then partner with business leads to connect this data to measurable business outcomes. All console screenshots in the article use demo data, and the ROI example for the sales team is based on hypothetical figures.

Why Billing and Active User Counts Alone Cannot Clarify ROI

Administrators initially see two types of figures: how many people are using the tools, and how much has been spent. The Usage dashboard in the Admin Console brings together active user counts, credits, and token usage for ChatGPT Work and Codex. Credits are the unit used in the console to track usage and spending, and subsequent task dashboards convert credits into estimated dollar amounts.

This dashboard can be filtered by group or individual. If a team's adoption remains consistently low, the administrator has a reason to check in with the team lead: to see whether the initially designed workflows are running smoothly or if supplemental training is needed. Administrators can also use this data to decide where to allocate support resources, which costs to review, and how to evaluate requests from teams for additional quota.

However, the Usage dashboard only answers "who is using it and where the money goes," not "what was gained in return." To determine whether spending is justified, admins must first understand the specific work employees are doing with AI.

Seeing What Work Credits Are Spent On

The Insights module in the Admin Console includes a task classifier: it samples messages sent by employees and categorizes them by "use case" and "task." For example, the software engineering use case includes feature development and code maintenance, while sales and revenue includes account research and planning.

The Insights overview page displays a chart showing how credits are distributed across tasks. In the demo data, Implementing features consumes 2.8M credits, estimated at approximately $194,535.63; Account research and planning consumes 1.2M credits, estimated at approximately $85,494.52. In this way, billed amounts can be mapped to specific types of work.

To look closer, administrators can switch to the Use cases breakdown table and filter by group. The table lists credits for each task, its share of credits within that use case, message count, cost per message, and active user count.

Taking the sales team in the demo data as an example: the sales and revenue use case consumed 1.46M credits in total, of which Account research and planning alone accounted for 1.22M, or 83.9%, corresponding to 958.45K messages, a cost of $0.09 per message, and 2,448 active users. This represents the sales team's single largest credit expense, giving administrators and sales leaders a concrete starting point for discussion: has AI actually changed how reps prepare account materials?

These classifications are based on message sampling. The guide does not state the sampling rate or classification rules, and the documentation page linked at the end of the article displayed "Page not found" when accessed on September 17, 2026. A safer approach, therefore, is to treat classification results as clues for deciding which workflows to investigate first, rather than as precise audit figures.

Model Settings and Plugins: Do Teams Need Adjustments or Training?

Once primary tasks are clear, the next question is whether the current usage for those tasks is appropriate.

Opening the details for a task reveals three breakdowns: Models (which models were used), Reasoning (reasoning tiers, such as Medium, High, and Extra High), and Speed (Standard or Fast mode), each showing the share of credits for that task across settings. Administrators can use this to determine whether configurations match the work and focus training on "how to choose a model." For instance, when drafting a routine brief, it may be worth testing a faster or cheaper configuration while comparing output quality and the time spent on manual review and revisions.

The same details page also includes a Plugin leaderboard and Skills usage, showing which tools support the task. Plugins connect ChatGPT to external systems such as Google Drive, Salesforce, and Slack; skills are reusable, established routines, which in the demo data include Account Brief, Customer Handoff, and Renewal Account Plan.

Reviewing these two lists yields two common signals. If a plugin that is clearly relevant to a task sees few calls, employees might lack access permissions or not know how to use it, signaling a need for admin follow-up; if a skill is invoked frequently, it requires a designated owner and regular updates.

Bharadwaj Tanikella, an AI Product Manager at Datadog, noted that OpenAI's analytics helped them understand how teams use AI, laying the groundwork for subsequent guidelines and policies. Datadog has already incorporated OpenAI's task classifications into its own AI agent monitoring product, Agent Console, showing customers what types of work their agents are performing; he believes that sourcing this data directly from OpenAI will make delivering such insights more reliable as customer AI adoption grows.

Engineering Teams: Connecting Codex Contributions to Engineering Outcomes

The Outcomes view is designed for engineering teams, tracking the share of merged commits and lines of code merged into repositories that involved Codex contributions, alongside code review activity. Trends can be filtered by group, user, or repository, helping administrators and engineering leaders gauge adoption and decide which teams should receive expanded access or additional support.

In the demo data, merged commits involving Codex accounted for 90% (108,000 of 120,000 commits), and lines of code accounted for 85% (15.3M of 18M lines), with both curves trending slightly upward over a month.

Having Codex involved in merged code does not equate to improved delivery efficiency for the team. If the share of merged code contributed by Codex is rising, engineering leads must look at this trend alongside review time, defect counts, and rework volume. Writing code faster does not necessarily lead to faster overall delivery if reviews take longer, defects increase, or rework mounts.

The guide does not explain how "involving Codex contributions" is specifically determined. When citing this metric, it is best to align on definitions with the team first and evaluate it alongside actual data from your own repositories and release workflows.

Using the Admin Plugin and API for Analysis and Reporting

The Admin plugin in ChatGPT Work allows administrators to conversationally compare usage, spending, and tasks across teams, synthesizing findings into reports for budgeting and rollout decisions. It can also produce ready-to-share deliverables directly, such as executive briefings featuring charts, key takeaways, and recommended next steps to share with cross-functional colleagues.

In the demo conversation, the administrator asks the plugin to prepare a monthly AI rollout review. The plugin concludes that credit usage over the last 30 days reached 6.81M, up 45.3% from the prior 30 days; within that, Work grew 43.1% to 3.76M, and Codex grew 48.1% to 3.05M. In the breakdown by team, the engineering team's credits were predominantly spent on Codex (2.8M), while the sales team used only Work (899.2K), with Codex at 0.

The Admin API allows teams to automate these reports and embed them into internal dashboards alongside data from core business systems. For example, a customer support dashboard can display credit usage side-by-side with ticket resolution times.

Moving from Console Data to Business Value

Usage and task data are only a starting point. To determine value, administrators need to investigate alongside business leads. Business leads understand the context invisible to the console: how workflows actually changed, whether outcomes improved, and what that improvement is worth. Only by combining both perspectives can in-product usage behaviors be tied to metrics like delivery time, quality, and profitability.

Taking a sales team as an example, this chain can be broken down into the following steps. The first two rely primarily on the Admin Console, while subsequent steps depend on sales leadership and data from business systems:

1. Identify clues in the console: The task classifier shows that account research is prevalent within a specific sales group.
2. Admin optimizes usage: Verify whether model settings fit the task; provide supplemental training where CRM plugin usage is low; share a skill so that team members produce account plans in a consistent format.
3. Compare against the baseline: The sales lead compares preparation time and account plan quality against pre-AI benchmarks (the baseline).
4. Track where saved time goes: If time was indeed saved, track how much of it was redirected into customer interactions.
5. Track what interactions yield: Track how many of these interactions converted into qualified opportunities (vetted opportunities worth formal follow-up), and how many eventually closed.
6. Calculate the financial return: Examine the revenue and contribution margin (revenue remaining after variable costs) generated by this business to assess whether the recovered capacity yielded a financial return.

Evaluating the value of a workflow can begin with five questions:

1. What would you like to improve? Choose an outcome that genuinely matters to the team, such as faster preparation, higher work quality, lower costs, or increased sales.
2. How does the process look today? Establish a starting point: how often the task occurs, how long it takes each time, and what constitutes a good outcome. Without this baseline, claims of "improved efficiency" lack a frame of reference.
3. What changes with AI? Compare results over a fixed period. Factor in review and revision time to ensure that while work gets done faster, outputs still meet the team's quality standards.
4. What does this make possible for the team? Saved time might allow for more customer conversations; higher quality might mean less rework. Talk with the team to identify where these improvements matter most.
5. Is the benefit worth the investment? Compare the team's gains against what was spent on AI, configuration, and ongoing support. The outcome helps determine which workflows to scale and which teams need more help.

An Annual ROI Example for Account Research

Imagine a sales team where each rep prepares 2 account briefs per week. Each brief requires synthesizing account history and industry research, and formulating talking points for the next interaction.

Preparing a brief manually takes 4 hours. With AI, materials are organized and summarized by AI, and then reviewed and revised by a human, totaling 1 hour—a net savings of 3 hours per brief. If half of the saved time is redirected to customer calls, at 30 minutes per call, the time saved per brief is enough for 3 additional calls. Beyond time savings, this can bring several advantages: stronger account plans, more personalized outreach, and shorter sales cycles.

Calculating the total value of these changes involves three steps:

Annual hours saved. For a team of 20 reps, assuming 46 working weeks a year:

20 reps × 2 briefs/week × 3 hours saved/brief × 46 weeks = 5,520 hours

Value of recovered capacity. Assume the team redirects 50% of this time toward productive work—such as prospecting new leads, following up on opportunities, or spending more time with clients—valued at a fully loaded hourly cost of $75. Fully loaded cost refers to the total cost an employer incurs for an employee, covering salary as well as payroll taxes, benefits, and allocated overhead.

5,520 hours × 50% × $75 = $207,000

Compare against first-year investment. Assume first-year expenses across AI, configuration, training, and ongoing support total $60,000:

($207,000 − $60,000) ÷ $60,000 = 245%, representing an illustrative ROI.

All figures above are hypothetical. This ROI reflects only estimated capacity value—the monetary value of saved hours based on labor costs—rather than realized revenue; nor does it factor in potential revenue gains from other sales outcomes like higher win rates or larger deal sizes. The value of a workflow also extends beyond efficiency gains: in sales, better account research and a shared understanding of customer accounts support deeper engagements, stronger client relationships, and new opportunities, allowing teams to subsequently measure whether these improvements lead to shorter sales cycles or higher win rates.

Across these three formulas, net hours saved per brief, the share of saved time truly redirected to productive work, and first-year costs are the factors most likely to vary by team. The calculator below applies the same formulas, allowing you to plug in your own team's assumptions to see how results change.

Outcomes Reported by Organizations Already Using the Tools

Approaches and metrics from three customers are detailed below, drawn from OpenAI's respective customer case studies and reflecting self-reported results:

1Password (software development with Codex): Uses Codex to build, review, and test software across planning, technical design, cross-stack development, PR reviews, testing, and production troubleshooting, shipping features and internal tools faster while maintaining strict security policies. It estimates that Codex delivers an ROI of 553% and approximately $800,000 in annual engineering capacity value. The calculation provided on the case study page is: 50 sustained Codex users × $250,000 annual fully loaded cost per developer × 20.9% measured productivity gain × 40% Codex contribution (directional estimate) × 75% capacity conversion rate = $783,750. The case study also notes a 10.9% reduction in median PR cycle time, and troubleshooting time for a defect spanning more than 10 microservices dropped from approximately 2 hours to 5–20 minutes.

ATV Big Air Tour (two-person team using ChatGPT Work for operations): This nationwide touring stunt show operates with a management team of just two people. They use ChatGPT Work to verify event details across publishing channels, plan merchandise inventory, and improve website visibility in AI search. Weekly time spent verifying event listings dropped from around 8 hours to 1 hour, and merchandise auditing and restocking decreased from two to three days to two to three hours.

Playco (game prototyping with GPT-6 Astra via API): Playco uses GPT-6 Astra via OpenAI's API to build and test playable game prototypes. Starting from the same greybox prototype (a base version built with simple geometry before thematic art is added), the team produced three prototypes with distinct themes, requiring 50% fewer manual fixes than with the previous model and allowing developers to playtest and evaluate more concepts.

How to Get Started

Getting started requires only three steps:

1. Open Insights in the Admin Console and select a common task that supports a strategic business priority.
2. Walk through it with the relevant business lead, agree on a baseline and the outcomes to measure, and set a date for review.
3. When the review date arrives, decide based on results whether to expand the workflow, refine how the team uses it, or try a different approach.

Regarding data privacy, OpenAI states that it does not train models on organizational business data by default.

The division of labor in this framework is clear: the Admin Console tells you which tasks money is spent on, whether usage patterns are appropriate, and what share Codex represents in code; answering whether those changes translate into faster delivery, more closed deals, and greater profit requires business leads to establish baselines and inspect data in core business systems.