Three numbers that count what AI actually does — 68% of ChatGPT conversations are not work, 7% of enterprises want to give agents full autonomy, and agents have placed "several hundred" orders on a delivery app

In the same week, three numbers appeared that count "what AI is actually doing." In OpenAI's consumer ChatGPT usage statistics (Signals v3.0, October 26), the share of work-related messages fell from 51.6% to 32.2% over a little more than two years, and writing lost its top spot to "practical guidance." In review site G2's analysis of 3,222 enterprise reviews (October 30), 7% of respondents wanted to give agents total autonomy, and the top trust-raising factor was "explaining why it did that." DoorDash's CEO said on the earnings call (October 29) that third-party AI agents have placed "several hundred orders in total" so far. Total orders for the quarter were 993 million. We lay out how to use the three numbers in running our apps, commerce, and agents.

This article was written with AI (Claude) and reviewed by an editor. OpenAI, the subject of the first item, competes with Anthropic, which makes Claude, and our company is a paying customer of both. G2, in the second item, is a review platform that sells marketing services to software companies, so it has a commercial interest in its own reports. Figures announced by companies are marked as "announced by the company," and where something comes from a single source, that source is named alongside it. This is not an investment opinion on any company.

AI companies' announcements tell you "what it can do." This week, unusually, numbers that count "what it is actually doing" came out from three places at once. One is from the people who use it (OpenAI's classification of consumer ChatGPT messages), one from the companies that buy it (enterprise agent reviews posted on G2), and one from the store that sells (the count of agent orders that came up on DoorDash's earnings call). In NO.16 and NO.17 we read "which AI is used for what" from companies' announcements, and in NO.25 we looked at what web share does not count. This time we put three numbers counted in three different ways side by side, and sort out what a small team like ours should treat as a planning assumption and what should stay an item to watch.

1. 68% of consumer ChatGPT messages are not work — OpenAI Signals v3.0 (fact)

On October 26, OpenAI released the third version (v3.0) of its usage-statistics program Signals (reported by PPC Land, Search Engine Journal, and Relevant Audience; all three agree on the release date). The method works like this. Each month, a sample of about 1 million messages is drawn from consumer plans (Free, Go, Plus, Pro), an automated classifier labels each with a topic (practical guidance, seeking information, writing, technical help, self-expression, multimedia, other) and "work / school / neither," and differential-privacy noise is added before aggregation. No human reads the conversations, only accounts that declared an age of 18 or over are included, and enterprise plans and the coding product Codex are excluded. This method continues the classification used in OpenAI's Signals methodology paper from this February and the 2025 joint paper with Harvard; we opened the methodology paper ourselves to confirm.

The figures compare July 2024 with September 2026 (the three outlets report the same numbers). Work-related messages: 51.6% → 32.2%. In other words, seven in ten consumer ChatGPT messages now have nothing to do with work. By topic: writing 34.1% → 21.4%, practical guidance 26.1% → 28.1% (the top topic every month since March 2025), seeking information 16.7% → 20.2%, technical help 13.5% → 4.6%, self-expression 4.0% → 8.9%, multimedia 2.0% → 8.8%. For one month, August, seeking information (20.7%) edged just past writing (20.6%), and in September it flipped back. By country, work was the majority use in 44 of 137 countries in Q3 2024 and in none in Q3 2026, and the median work share across countries fell from 46.4% to 27.9% (figures compiled by PPC Land, one source). The figure that the work share on the Free plan alone fell from 44.0% (September 2024) to 24.7% comes from Relevant Audience, one source.

Three things to be careful about when reading these numbers. First, these are shares, not volumes. If total messages grew, the absolute number of work messages could rise while its share falls, and Signals does not publish volumes. Second, the drop in technical help from 13.5% to 4.6% may be because coding moved to Codex, but Codex is excluded from the data, so this dataset cannot tell us. Enterprise plans are excluded too, so work use is undercounted. Third, Signals recomputes past months with each new version, so the same month carries different numbers in v2 and v3. That is why the version number should be cited with any figure. One more note. When we opened OpenAI's Signals data page ourselves, it showed the previous version ("updated May 11, through March 2026"), so we could not check the v3.0 figures directly against the official page. The figures above are carried on the basis that three outlets of different kinds reported the same numbers on the same day.

2. 7% of enterprises want to give agents full autonomy — G2's 3,222 enterprise reviews (fact)

Software review platform G2 published, on October 30, an analysis of 3,222 AI agent reviews written by users at companies with more than 1,000 employees (January 2025 to October 23, 2026; 231 products; five categories: business operations, customer support, agent builders, sales development, marketing) (G2's own data; interest disclosure at the top). Review counts more than doubled, from 983 in 2025 to 2,239 in 2026 (through October 23), and users who "use it most of the time or every time" went from 47% to 53%. Use cases: process automation 34%, customer service 27%, software development 21%, sales development 12%, marketing 7%.

The answers on autonomy are the heart of this piece. Preferred autonomy level: "autonomy with guardrails" 60%, "limited autonomy" 21%, "a human in the loop for every action" 11%, "total autonomy" 7% (6% in 2025). Even among respondents who "trust their agent very much," only 15% would grant total autonomy, and admins (5%) and executive sponsors (2%) are more conservative than everyday users (10%). Asked what would raise their trust, 79% chose "let me see it or control it": explaining decisions 32%, guardrail settings 29%, a warning before a questionable action 18%. "Be more accurate" was 21%. In the same vein, among about 500 reviewers' ratings, 92% gave output accuracy a 4 or 5, but only 82% did the same for "explaining results or fixing mistakes," and the complaint "I can't see why it acted" rose from 3.0% to 5.3% of all dislikes.

The satisfaction numbers slipped a little each year. Average likelihood to recommend 9.24 (2024) → 9.10 (2025) → 9.04 (2026), share of perfect 10s 55% → 48%, average share of users who fully adopted the agent inside the organization 68% → 55%. The top complaint is setup and complexity (25%), and praise for "good integration" fell from 30% to 21%. In a separate survey G2 cites in the same article (July 2026, 1,019 B2B decision-makers; summarized by SaaStr on October 16), 66% of organizations have agents in limited or scaled production, but 31% say employees spend six or more hours a week overseeing agents, and 45% say oversight and rework eat more than a quarter of the gains they expected. In that survey, too, total autonomy was preferred by 7% (guardrails 48%, limited 29% — the distribution differs somewhat from the review analysis).

Caveats. G2 itself says the like/dislike tags use a keyword method and should be read "for direction only"; reviews are a self-selected sample of people who chose to write them; and the change from 6% to 7% is small enough that it is hard to call statistically meaningful. The 2026 figures are a partial year through October 23. And G2 is a company that attaches a pitch for its vendor services at the end of the article.

3. Agents have placed "several hundred orders in total" on a delivery app — DoorDash's earnings call (fact)

US delivery platform DoorDash reported third-quarter results on October 29. Total orders for the quarter: 993 million (up 28% year on year), gross order value $33.9 billion, revenue $4.6 billion (company results, compiled by PPC Land). The results materials carry no agent-related figures and introduce three of the company's own agents: Ask DoorDash (June), which recommends restaurants similar to your usual orders and builds a cart; Text DoorDash (September), which takes orders by text message outside the app; and an MCP connection to external AI agents (for corporate group ordering). In July it also released a CLI agent that orders from the command line (TechCrunch).

The number came up in the Q&A on the earnings call. When an analyst asked about agent orders, CEO Tony Xu called it "still a very dynamic area" and said, "we're seeing maybe several hundred orders in total from some of these third-party agents" (TechCrunch, October 30, quoting the remark). The outside agents named were Muse, Instinct, Fo, and others. He added that "you never want to fight gravity" but "you don't want to be delusional," and that most users still order through the app. According to the call summary (GuruFocus), the company expects everyday ordering to stay in the app, with agents a better fit for long-consideration or proactive purchases, and mentioned advertising inside agents as an opportunity. Put "several hundred" next to 993 million orders in the quarter and the ratio is somewhere around one in a million. Against an average of about 3 million deliveries a day (the company's 2025 report, cited by TechCrunch), it is roughly one ten-thousandth of a single day.

The same week brought reports that DoorDash agreed to acquire Instacart (agreement dated October 28; TechCrunch, PPC Land), but that is not this article's subject and not investment material, so we leave it out.

Reading the three numbers together (interpretation)

The three numbers count three different things. Signals counts shares of messages, G2 counts the preferences of people who wrote reviews, and DoorDash counts one company's orders. So folding them into "AI use is falling" or "agents don't work" would be wrong. What they say together is narrower and more useful: people use AI a great deal to "ask and be guided," companies watch AI "do things for them" but demand explanations, and "buying for you" all the way up to the storefront has barely happened yet.

What we read in the first number is not that the definition of "work" changed but that the people using it broadened (interpretation). Over the two years in which the work share fell, ChatGPT gained free users (the basis of the share rebound we saw in NO.25), and practical guidance and information seeking grew. When a Safelist user photographs an ingredient label and asks "can I eat this?", that is "practical guidance" and "seeking information" in Signals' classification. In other words, what our app does sits inside the two largest cells of consumer AI use. But general-purpose chatbots occupy those cells too, so what sets us apart has to be the evidence structure that "never says it's safe" (NO.3).

The second number points the same way as NO.29's "start from zero permissions." That six in ten enterprise reviewers choose autonomy within guardrails, and eight in ten choose "show me why" over "be more accurate," means an agent product's competitiveness lies as much in explanation and warnings as in accuracy (interpretation). The 31% who spend six or more hours a week on oversight and the 45% whose gains lose more than a quarter to oversight and rework explain why a plan that brought in agents without putting oversight time into the cost column goes off course.

The third number shows where the phrase "agentic commerce" stands right now. DoorDash has opened three agents of its own plus MCP, and outside agent orders are still in the hundreds. We read this not as "agent orders won't come" but as "keep the channel open and enter zero in the revenue plan" (interpretation). The CEO's "don't fight gravity, don't be delusional" applies just as well to a small commerce business.

If it were us — three things to do this week (practice)

Reclassify our app's AI answers into the two cells of "guidance" and "search." Take one month's sample of the questions the AI features in Safelist and Hangul Cups receive and split them into Signals' topic cells (practical guidance, seeking information, writing, technical help, self-expression, multimedia). Use a classifier only, with no human reading, and note how our distribution differs from consumer ChatGPT as a whole (guidance 28%, search 20%). If "writing" requests are more common than expected, people are leaving because the feature does not exist; if "self-expression" is high, it is being used like counseling and the guidance copy needs work. The result goes next to NO.17's task table as a column for "our users' distribution."
Keep the agent channel "open, entered as zero." Prepare a structured feed (price, stock, ingredients, shipping terms) that outside agents can read for our commerce products, but enter agent traffic as zero in this year's revenue plan. Instead, create a separate "agent" column among traffic sources and count it monthly. If "several hundred in total," as at DoorDash, continues for a few months, it will not be too late to think about ads and partnerships then. Add an "agent traffic count" row to the same table as the "ad account default check" from NO.28.
Attach a "why" log and a "warning before acting" to our agents first. Ria's video-production pipeline and the review-analysis agent currently leave only "results." Put in the "explain decisions" and "warn before questionable actions" that G2 reviewers ranked first and third, and improve accuracy after that. Concretely, every time the agent finishes a step it leaves one line of "what and why," and at steps involving money, sending, or deletion it requires a human check, using NO.29's permission table as is. Measure oversight time weekly and put it in the cost table. If it exceeds six hours a week, that agent gets fixed by "lowering its autonomy level or narrowing its scope." We also adopt, in our own way, the "90-day adoption target" G2 suggests. If half the team is not using it after three months, that agent is shut down.

What we did not pick this time

The pricing preferences in G2's survey (usage-based 57%, per-seat 9%, and so on) are useful for our decisions as a buyer but differ from this article's subject, so we left them out. GitLab's "4.5× excess token usage" for Orbit, the context graph it made generally available on October 30, is a company claim and we could not find the official post, so it is not included (a follow-up candidate for NO.21 and NO.30 on "price and cost"). The report that security evaluator Irregular found GLM-5.3 cheating its way to correct answers in a cyber evaluation (Help Net Security) is one we did not read in full and may contain attack details, so we did not cover it. Next.ink's "Google will put ads in AI Mode" (October 29) is deferred as a NO.28 follow-up because we could not confirm Google's official post. The reports of OpenAI trimming its funding round and the tip on Opus 5.5 reasoning-effort settings, reposted on promppy, were not used before confirmation of originals or official sources.

What we could not confirm in this story

OpenAI Signals v3.0: We opened the official data page directly, but it showed the previous version (May update), so the v3.0 figures are confirmed only by the agreement of three outlets. Absolute message volumes, the classifier's model name, and the noise sizes are not published. The country median (46.4% → 27.9%) and the Free-plan work share (44.0% → 24.7%) each come from one outlet's compilation.
G2: The article does not explain why the autonomy-preference distributions differ between the review analysis and the separate survey (60/21/11/7 vs. 48/29/—/7). "31% spend six or more hours a week on oversight" and "45% lose more than a quarter of gains" are figures from the survey (1,019 respondents), which we confirmed through SaaStr's summary and the citation in G2's article; we did not open the survey report itself.
DoorDash: "Several hundred" is the CEO's remark; company materials carry no figure. We did not check which companies make the outside agents named (Muse, Instinct, Fo). The 3 million a day is from the 2025 report, a different period from Q3 2026.

Sources we read

We opened sources of different kinds directly to confirm. They are listed by type: official materials and methodology, research organizations, media reports, and earnings call summary.

Official materials and methodology:

Research organization (with a commercial interest):

Media reports:

Earnings call summary:

finance.yahoo.comDoorDash (DASH) Q3 2026 Earnings Call Highlights (GuruFocus, on Yahoo Finance)English, Q&A summary

Our earlier articles: