A billion people use AI chatbots. You do not need one for everything.

Google calls Gemini its fastest-growing product ever. Its August 11 announcement says 63% of users talk to Gemini, one in five Gemini Live interactions includes a camera feed or screen sharing, and more than 100 million active users are on iOS. Google also says people generate more than 150 million images in Gemini each day. These are company-reported usage figures, not an independent study of whether the resulting advice, image or action was good.

OpenAI's August 6 material gives a different picture. The company says ChatGPT has 1 billion weekly users and that, at work, people are more than twice as likely to ask it to produce something or perform a task than they are outside work. Its Signals analysis covers consumer Free, Go, Plus and Pro accounts; it excludes enterprise accounts and Codex, and the public dataset is based on sampled messages through March 2026.

The Verge notes the denominator problem. Gemini's public milestone is monthly activity. ChatGPT's is weekly. A person opening an app once in a month and a person using another app every Tuesday do not represent the same habit. The useful conclusion is not that one vendor won. It is that hundreds of millions of people are already trying voice, images, research, writing and practical help without waiting for a final verdict on AI.

The vendors would prefer the answer to be everything. That is how a chatbot becomes the front door to search, shopping, files, calendars, creative tools and eventually paid recommendations. The user has a different interest: finishing one annoying thing without creating three new things to check.

Good first jobs have a visible after-state. Turn these rough notes into an agenda I can edit. Compare these two repair estimates and list the differences. Read this school email and extract the dates, costs and forms. Explain this error message using the manual I attached. Draft a reply, but do not send it. You can look at the result and decide whether the chore got smaller.

Weak first jobs hide the standard of success inside the answer. Tell me what career to choose. Decide whether this symptom is serious. Run my entire business. Tell me which news to believe. The chatbot may produce a fluent response, but fluency does not give you a clean way to notice what it missed. High-consequence decisions also make a bad place to learn the tool's limits.

There is a simpler split than chatbot versus AI agent. Use an assistant when the source material is available, the result is inspectable and a mistake is cheap to catch. Keep a person, primary source or qualified professional in charge when the job changes money, health, rights, access or a relationship you care about.

Pick a task you already do, not a task invented to justify the subscription. Keep the original material beside the answer. Then run the same kind of job three times during a normal week.

After each attempt, write down four things: minutes spent, corrections needed, whether you reopened the source, and whether the result was actually used. Do not count a polished draft as finished if you spent the next half hour checking names, dates and links. Do not count research as useful because it was long. Count the decision, document or solved problem that survived contact with the rest of your day.

Set a dull stop rule. If two comparable attempts take longer than doing the job yourself, move that task back to the old method. You can revisit it after the product changes. What you should not do is turn every bad fit into a weekend course in prompting. A billion other users do not owe the tool your persistence.

If it works, save the small pattern: the source you provide, the boundary, the output shape and the final check. That is enough. You do not need a folder of 200 prompts or an AI version of every app you already have.

Priya Rao would treat the billion-user milestone as evidence of reach, not value. Her test stays attached to one recurring job: did it finish, how many corrections came back, how often did the source need reopening and who kept the saved time? If the measurement ends at messages sent or drafts created, the tool can look busy while the person inherits the cleanup.

Mina Torres would remove the feeling that everyone else has already mastered some secret language. A useful chatbot should help with the document, photo, question or error already in front of you. If the first result requires a prompt tutorial, three settings pages and a new vocabulary before it becomes safe to use, the product handed its onboarding job to the customer.

Priya's view is stricter about proof. Mina's is stricter about burden. Together they leave a reasonable bar: the assistant should complete something you recognize without making you study the assistant.

A billion people using chatbots does not mean a billion people want an all-day synthetic coworker. Google's own figures show people talking, sharing screens, making images and solving immediate problems. OpenAI's data separates asking from doing and shows that the mix changes between work and life. People are fitting the tool to different moments. That is healthier than treating maximum usage as the goal.

Use it for the explanation that saves a trip through a manual. Use it for the first pass that lets you leave work on time. Use it to compare, extract, draft or rehearse when the source and final judgment stay visible. Then close it.

The companies have finished proving that AI chatbots can attract a crowd. They still have to prove, task by task, that the crowd gets something back besides another place to spend attention.