ChatGPT Enterprise usage report for a companyWho uses it, for what, and which AI tools to buy next
How to find out how employees use ChatGPT Enterprise: five admin console exports turn into adoption depth, uniformity across teams, main tools, a task map, data checks and an interview shortlist.
ChatGPT Enterprise is a version of ChatGPT for organizations with an admin console that shows who uses the tool and how. This report is built from five console exports and shows adoption depth, uniformity across teams, the nature of tasks and the people to interview. It is used to decide which AI tools to buy, whom to train and which assumptions to test.
Business task: find out for which work tasks the current set of AI tools falls short and which solutions are worth testing before buying. The decision is made by the owner of the AI tools budget together with the leads of the teams where the pilot runs. Data access and security are approved by the information security team. The decision criteria are the signs in the table below.
That the current set is suboptimal somewhere is an assumption: the report and the interviews test it. "Suboptimal" breaks down into testable signs. The conclusion "no new tool is needed" is also a result.
Data limits: the report only sees the corporate tool and only for the export period, 3 months. Personal accounts and other tools do not show up in it.
How answers turn into a purchase decision
The report shows where to look, and the purchase decision is made only after these five steps, so the money goes to a task many people have and nothing covers today.
1
Interviews
Which tasks and difficulties exist and how people work around them now
2
Company-wide questionnaire
How many people face a difficulty, how often, and how much it gets in the way
3
Compare
Task scale, severity of consequences, workarounds, cost of the solution
4
Test tools
Promising options on real tasks together with employees
5
Decision
The budget owner decides by the six signs above: quality, time and rework, reproducibility, scale, data access, cost; "no new tool is needed" is a valid outcome
Input
Console exports: messages, spend, projects, and connectors for each person
What the report does
Turns exports into segments, team maturity, and specific people as material for a decision
Decisions at the output
Which tool to pilot, for which team, who to talk to before buying, and where budget is not needed
Workflow
Five steps from export to a finished list of people.
1
Export the data
Five console sections, details in "Where to get the data"
2
Build the big picture
Adoption depth, uniformity across teams, main tool, segments
3
Map the nature of work
Which tasks people solve, with which built-in tools, and with which integrations
4
Check against raw data
Verify each summary against the source export
5
Select people and draw conclusions
An interview list and the questions to test in them
What a ChatGPT Enterprise admin sees: where to get the data
Everything is built from five sections of a single ChatGPT Enterprise admin console. Hover over i to see exactly what each section contains.
How to export data from the ChatGPT Enterprise console: by hand and with the Claude in Chrome browser agent
How to export. The console does have APIs, but they cover only part of the data: Codex analytics, credit costs, and conversation logs for compliance. The users, projects and skills exports and the per-product leaderboard are easier to collect in the interface. Exporting is a good job for Claude in Chrome: it opens the right section, sets the filter, clicks export, and saves the CSV. Every step is visible on screen, so a wrong filter or section is caught immediately, before the numbers reach the report.
Access. The console API is not always available: company security policies often restrict admin API keys. In that case, export every section through the interface.
Messages, activity dates, and built-in tools for each person
iWhat's insideCSV export for a period of up to 12 months. Fields: messages, tool_messages, gpt_messages, project_messages, department, groups, seat_type, last_day_active. The console itself marks "power users" with a badge.
Spend, tokens, and lines of code for each person, separately for the Chat, Work, and Codex products
iWhat's insideExport button next to the heading → "User ranking", CSV or JSON. Fields: Name, Email, Credits, Estimated costs, Tokens, Lines of code. The "Product" filter changes what the export contains. History covers about 120 days.
Who created a project, how many messages it has, and whether it is active
iWhat's insideFields: project_name, project_creator_email, messages_workspace, unique_messagers_workspace, is_active. Some names come through as [content redacted]: the console hides them on its own.
The console's own task classification: code or knowledge work (everything except code)
iWhat's insideShare of code and knowledge work, top tasks by credits. An OpenAI model labels tasks on a sample of about 10% of requests over 30 days.
Big picture
The next four sections are the report itself, each with its goal. Goal: show which groups of people exist in the company and how large they are.
Adoption depth
Adoption depth shows how deeply the tool has entered the company's work: how many people have used it at least once, regularly, and every week. Those who never write in chat and work only through Codex are shown separately. Decision: which group to invest training in and whether to buy more licenses now.
71
88
97
23
61
every weekregularlyoccasionallyCodex onlyfully inactive
340 people
Source: Users · Leaderboard for those who work only through Codex
Takeaway
In this example82% of people have tried the tool, 47% use it in their work, and for 21% it has become a weekly habit. The room to grow depth is the 97 people in the "occasionally" group.
In generalThe gap between "tried it" and "use it every week" shows how many people stopped after the first attempts. The benchmark for mass adoption is 50–70% of employees using it in daily work. Until adoption reaches that benchmark, more licenses will not drive growth: something other than access is holding people back.
Next stepPut training into the largest group between "tried it" and "every week": it gives the fastest growth in regular use. Recount the breakdown after the training to see where people moved.
Important. Pick 5–10 people from the "occasionally" group, find out in interviews how they get their tasks done now, and run a review of real work cases with them. Recount the weekly share after the review. How to ask without suggesting the answer is covered in "How to interview people about their AI use".
Uniformity
Uniformity shows how evenly the tool has entered different teams: what share of each team uses it regularly and how many messages one person writes on average. Decision: which team to take working scenarios from and which team to bring them to first.
In this exampleIn Team A, 74% of people use the tool regularly; in Team E, 18%: a spread of 56 percentage points.
In generalA large spread between teams means working scenarios already exist but live in one or two teams: moving them is cheaper than buying something new. A uniformly low share across all teams points to a shared barrier: access, training, or unsuitable tasks. A uniformly high share means the tool has taken root and you can move to the next step. A team with barriers is not ready for a new tool yet.
Next stepWith a large spread, run a meetup: the strongest team shows three of its scenarios to the two weakest. Before the meetup, run 2–3 interviews about barriers in the weakest team. Recount the spread after the meetup: it should shrink.
Main tool
Which of the three tools takes more than half of a person's spend over three months. The tool type shows how the person works with AI.
In this exampleChat is the main product for 164 of 340 people, Codex for 38, Work for 47.
In generalThe main tool shows the type of work: a conversation, a task handed to an agent, or code. It determines what the person needs next and which pilot to put them in. Group size sets pilot size: a wide pilot for a common scenario, a narrow one for a rare scenario.
Next stepThe decision is to move each group to the next step. For the Chat group: Work with connectors and MCP to work data, which removes prompting and hallucinations. For the Work group: specs, harness setup and shared skills, which make the result reproducible and transferable to other teams.
Top by spend
The top list ranks people by the amount spent on a specific product, for example Codex. This way the list includes someone who spends a lot on Codex even if most of their spend goes to Chat.
In this exampleHana Sato and Luis Moreno spend about five times more on code than the third person on the list. With similar spend, Hana gets one and a half times more tokens per dollar than Luis.
In generalTop users are change agents: they find working scenarios first, and their experience affects whether a specialized tool pays off. Next to them are more cautious colleagues who need more time and a live example on their own tasks. Ranking by share of spend hides the change agents, so look at absolute spend.
Next stepInvolve top users as mentors: let them show their scenarios to colleagues who are just starting, on those colleagues' tasks. Invite cautious colleagues through examples and results, without mandates. Offer the top two or three an experiment in a specialized tool with a before-and-after report.
The leaderboard has both spend and tokens. Compare the tokens-to-spend ratio within one group: one product and similar tasks. Whoever gets noticeably more tokens per dollar usually picks models and modes better. These are the most advanced and efficient users, and they are the first to invite as mentors. Introduce them to colleagues from the same group whose ratio is lower: working through one task together often gives more than a general training. Pairs from different teams bring cross-pollination: good techniques move between teams.
During rollout, top users can spend an order of magnitude more than everyone else. It is better not to introduce limits at this stage: they easily demotivate the people who adopt the tool first. Instead of limits, agree on experiments and a results report. Limits make sense later, once practices are proven and the normal cost of a scenario is clear.
Behavioral segments
segmentinterviewproblemtool
Groups by behavior intensity built from several exports: how much a person writes, spends and what they do.
In this exampleThe largest segments for customer interviews are intensive research (58) and code (44, 19 of them outside technical teams). For the 61 inactive people the question is different: what stops them from starting.
In generalData shows what people do but not whether the current tool is enough for them. So a segment is a hypothesis: it suggests whom to invite to customer interviews, which questions to ask, and what problems people may run into. Tool candidates appear already here, and the choice between them is made after the interviews, for a confirmed problem.
Next stepFor each segment, talk to 2–3 people: where the current tool falls short and what they still do by hand. Gather candidates for a confirmed problem from three sources: consultations with innovators inside the company, outside experts, and market benchmarks, meaning what similar companies use. Next, follow the decision procedure: questionnaire, comparison, and testing candidates on real tasks. If the problem is not confirmed, postpone the segment until the next recount.
Nature of work
Goal: show which tasks people solve with the tool, with which built-in features, and through which integrations.
Connectors
Connectors link the assistant to work data: chats, documents, code, calendars. They determine the volume of hallucinations: without access to data, the model fills in the answer itself. Decision: which teams and which systems to connect first.
In this exampleSlack and Notion account for 70% of all connections. The model does not see systems outside this list: it answers questions about them with a guess.
In generalThe volume of hallucinations drops where the model has the context it needs. Without it, the agent tries to fill the gap on its own. Teams with zero connections work with the model blind, and almost any answer about their processes is a guess. The sign in a customer interview: "I have to double-check everything".
Next stepFind teams with zero connections and ask in interviews where their work data lives. Connect one such source and compare before and after on 10 identical questions: how many answers are wrong and how many rely on work data.
Skills
Skills show process maturity: a scenario once described as a skill repeats the same way for anyone who runs it. The stats show which skills are used and who builds their own. Decision: which processes can already be rolled out to other teams, and which still depend on individual people.
In this exampleThe spreadsheet and PDF skills have 88 and 57 calls: these processes are already reproducible. 29 calls to skill-creator: some people are turning their scenarios into skills.
In generalProcess maturity grows in steps: one-off requests → a personal skill → a shared team skill. At the first step, the result depends on who asked and how. With a skill it is reproducible: the same input gives the same result. A shared skill makes the process transferable to other teams.
Next stepCollect skills into one list with their authors and mark the step of each process. Turn the two most used personal skills into shared ones: the author shows them to a neighboring team, and the team checks that the same inputs give the same result. Scenarios without a skill that repeat every week are the first candidates for new skills.
Project themes
The summary groups projects by meaning and shows the concrete tasks people run in ongoing projects. For customer interviews this is a ready list: which tasks to discuss and with whom. Decision: which tasks and whose projects to cover in interviews first.
Theme
Projects
Tasks
Question for the author
SalesPitch library · Elif Demir
11
pitches, objection handling, commercial proposals
"Which part of a pitch does the model do well, and what do you rewrite?"
OnboardingNew hire guide · Rafael Costa
8
guides for newcomers, answers to typical questions
"Which newcomer questions does the model already handle without you?"
Market researchCompetitor notes · Jonas Keller
5
competitor summaries, price comparisons
"Where do you get the data, and how quickly does it go stale?"
In this exampleSales (11) and onboarding (8) have the most standing projects. The authors of the largest projects are the first people to interview.
In generalA project is a knowledge base: files and instructions the model relies on in every answer. The fuller the base, the more accurate and stable the result. What a project contains suggests what to ask in an interview: which files were added, what is missing, and where answers still need fixing.
Next stepOpen the three most active projects in each theme and list the concrete tasks. In the interview, ask the author the question from the table and which files and instructions changed answer quality the most. Turn a successful project setup into a template for that theme.
Verification
Goal: check the summaries against the raw exports, the project list and the leaderboard. Any number in the report can be traced back to its source row in an export.
Full project list
All projects with author and status. This list is used to check the theme summary and to find the author of a specific project.
In this exampleThe "Personal draft" project has zero messages.
In generalAn empty project is an abandoned attempt: the person started, but the tool never became part of their work. If such projects stay in, the topic summary overstates usage.
Next stepFilter out projects with zero messages before counting and invite 2–3 of their authors: the reason they quit is a valuable answer.
Leaderboard
The billing export for each person and product. Main tools and top lists are calculated from it.
Name
Email
Credits
$
Hana Sato
hana@example.com
10,955
438.20
Mira Castellano
mira@example.com
2,425
97.00
Jonas Keller
jonas@example.com
470
18.80
Source: Leaderboard · three exports with different "Product" filters
Takeaway
In this exampleThe three product exports add up to the total spend of each person on the list.
In generalThe three exports with different Product filters add up to each person's total spend. A gap above 2% means an export is incomplete or the periods differ.
Next stepBefore sharing the report, sum the three exports per person and compare with total spend. If they differ, re-export the data for the same period.
People selection
Goal: name specific people for each way of getting tasks done.
Interview candidates
A short list of people for each direction found. In the example the final list is 16 people out of 340; the table shows the first three. Your company will have its own number. The ChatGPT report is one of the sample sources: users of other tools and people who quit are added by the rules in "Who to interview", and the check across all employees is done with a questionnaire.
Person
Direction
Signal
Hana Sato
code
$438 / 3 mo
Mira Castellano
agentic tasks
$97 / 3 mo
Jonas Keller
research
3,200+ msgs
Source: the users, leaderboard and projects exports, joined by email
Takeaway
In this exampleHana Sato, Mira Castellano, and Jonas Keller represent three different ways of working: code, agentic tasks, and research.
In generalThe shortlist has one person per direction: code, agent tasks, research. Each conversation then shows its own way of working and does not repeat another one.
Next stepSchedule three interviews, one per direction. Afterwards, decide which direction needs 2–3 more people.
Table of all people
Everyone, with filters by team, activity, direction, and main tool.
Team BActivity: every weekTool: Work
Person
Team
Chat / Work
$ / 3 mo
Tomás Reid
Team B
112.00
Mira Castellano
Team B
97.00
Source: the users, leaderboard and projects exports, joined by email
Takeaway
In this exampleThe "Team B" and "Tool: Work" filters left two people: Mira Castellano and Tomás Reid.
In generalA table of all people with filters by team, activity and main tool is the base for any next sample: a list for a new question needs no new export.
Next stepSave each set of filters as a separate table view so the next list takes one click.
Where to start today
You do not need the whole report at once: the first useful result, a list of most active users, appears after the second item.
Export users and the leaderboard for three monthsThe leaderboard three times, once per product. A browser agent can do the export
Calculate main tools and the top by amountThe first list of most active users is ready
Calculate adoption depthAt least once / regularly / every week, with those who work only through Codex counted separately
Build a list of people and schedule interviewsHow to pick these people is covered in "Who to interview"
Common questions
What the ChatGPT Enterprise admin console shows
Five sections: users, leaderboard, projects, overview and tasks. They give adoption depth, uniformity across teams, the main tool, the nature of tasks and a check against raw data.
How to export data from the console
The console has APIs, but they cover only part of the data: Codex analytics, credit costs and conversation logs for compliance. The users, projects and skills exports and the per-product leaderboard are easier to collect in the interface, and the export can be handed to Claude in Chrome.
How to tell whether new licenses are needed
Adoption depth shows how many people use the tool at least once, regularly and every week. It decides which group to train and whether to buy new licenses now.
How to choose people to interview
People are picked for each direction found: in the example the final list is 16 people out of 340. Users of other tools and those who stopped are added by the rules in “Who to interview”.