AI Strategy

AI Token Limits: Where to Use AI and Where to Stop

By Alyssa Caridi · Last updated July 5, 2026

• AI Costs
• AI Governance
• Token Limits
• AI Strategy

Usage limits have changed the calculation for where to use AI. Through 2024 and 2025, many leaders told their teams to use AI for everything. That approach no longer works, because the major providers have added weekly caps, metered credits, and higher costs for heavy use. A more selective approach makes sense now. Use AI for high-value work that benefits from scale and speed, such as call recaps, complex data analysis, deck production, and recurring research. Keep people on the work that carries your judgment and voice, such as executive summaries, corporate communications, and first-pass thinking. Spending your token budget on tasks a person should own is how teams end up paying more for worse output.

Why does everything made with AI look the same now?

Start here, because this problem predates the pricing changes and would have forced a pullback on its own. Over the last two years, a lot of work made with AI has converged on the same shapes. Decks look alike. New websites look alike. Client question lists and feedback read like they came from one template. This happens because general-purpose models pull toward the average of what they have seen. A 2024 study in Science Advances found that generative AI raised the quality of individual writing while making the resulting stories more similar to one another, which reduced the collective diversity of ideas across the group (Doshi and Hauser, 2024). When an entire team drafts from the same tools, the output homogenizes. For anyone whose job is to make a brand distinct, sameness is a direct threat to the work. It is the same convergence problem that makes standing out in AI answers harder every quarter.

What did over-using AI do to how teams read and think?

The flattening came with a second cost. Colleagues and clients have spent two years complaining about generic AI feedback, generic AI question lists, and generic AI responses. Some of that traces to a habit that AI made easy: skipping the reading. People stopped reading project briefs and RFPs in full and let a model summarize instead, which left them underprepared and unclear on what they were actually responding to. A 2025 Microsoft and Carnegie Mellon study of 319 knowledge workers found that the more people trusted AI, the less critical thinking they applied to its output, while those with more confidence in their own skills engaged more carefully (Lee et al., 2025). The practical fix is old-fashioned. Read the brief yourself and form a first opinion before you open a model. Bringing reading back into the workflow improves the quality of everything downstream.

What changed with AI usage limits and costs?

The economics shifted just as the quality concerns were mounting. The providers that once encouraged unlimited use have added caps and meters. Anthropic introduced weekly rate limits on Claude in late August 2025 after unannounced summer restrictions frustrated heavy users (TechCrunch, July 2025). A class-action lawsuit later alleged that the Claude Max plans deliver far below the advertised amount of usage (Engadget, 2025). OpenAI moved its paid business plans onto a credit system in which advanced features draw from a purchased pool and pause when it empties (OpenAI Help Center, 2026). For a leader, the message is plain. Telling a team to run every small task through AI is now a line item that grows every month.

Does ChatGPT Enterprise actually save you money on usage?

Not by default, and this catches teams off guard. Under OpenAI's flexible pricing, ChatGPT Business seats come with included per-seat limits for advanced features such as deep research, thinking models, and image generation, and a workspace can add a shared credit pool only if it wants to unblock users who exceed those limits. Enterprise and Edu workspaces instead buy a shared credit pool at the contract level, and once that pool is exhausted, advanced features pause unless an owner enables overages or purchases more credits (OpenAI Help Center, 2026). ChatGPT Business runs around 25 dollars per user each month, while Enterprise is negotiated directly and typically requires a 150-seat minimum (IntuitionLabs, 2026). A team that has not yet learned to keep its consumption at a level it is comfortable with can spend more on Enterprise credits than it would have on Business seats. Get your usage under control first, then decide whether the move to a metered Enterprise contract is worth it.

What should you use AI for?

Point AI at work where scale, speed, and analysis produce clear value and where a strong model earns its cost. For the heaviest tasks, use a frontier model such as Claude Fable 5, Claude Opus 4.8, or GPT-5.5, since the quality gap is largest exactly where the work is hardest.

• Call recaps and transcription. Recording and transcribing calls is standard practice for a reason. It keeps everyone looped in and gives you a detailed record you can search and re-reference at any time.
• Complex data analysis. A strong model can find patterns across large or messy datasets faster than a manual pass, which is a good use of both the tool and your budget.
• Deck creation. Turning approved content into a first-draft deck saves real production hours, as long as a person edits the narrative and design.
• Recurring research automations. Daily or monthly research runs, press-hit aggregation, and standardized reporting are repetitive, high-volume tasks where automation pays off.
• Secondary brainstorming. Use AI as a second pass after your own thinking, to pressure-test or extend ideas you have already started.
• SEO, competitive research, and positioning. AI is well suited to scanning competitors, optimizing content for search, and helping you find a sharper position in the market.

What should you stop using AI for?

Keep people on the work that carries your voice, your judgment, and your understanding of the account. Some of these will be controversial, and that is fine. The cost of getting them wrong is generic output and a team that cannot speak to its own work.

• Basic emails. A short email is faster to write than to prompt, edit, and send, and it keeps your own tone intact.
• Calendar availability. Checking your own schedule is not a task worth a token, and the tools add friction rather than removing it.
• Reading your inbox and building your to-do list. This is the controversial one. Handing your inbox to a model means you stop reading what people actually sent you, which is how details and context get lost.
• Executive summaries. The summary is where you prove you understood the material. Writing it yourself is part of the thinking, not a chore to remove.
• Corporate communications. Internal and external comms carry the organization's voice and stakes, and readers can tell when they were generated.
• Social copy. Social is a place to sound like a person, and model-drafted posts read like everyone else's feed.
• Initial brainstorming, questions, and feedback. Your first take should be yours. Starting with AI anchors the whole team to the same average answer and skips the reading that makes feedback worth giving.

How do you decide who gets more tokens and when?

Treat expanded AI access as a budget you allocate on purpose. OpenAI's plans let owners and admins use role-based access control to set spend limits by group and cap overages, so heavy access can be given to specific roles rather than the whole workspace (OpenAI Help Center, 2026). Decide which functions genuinely need frontier-model access, such as data and research roles, and give them a clear path to request more. Require a short justification tied to an approved use case, so an increase maps to work you have already agreed is worth the spend. Review the usage reports monthly and adjust, rather than letting access creep quietly upward.

How should AI usage show up in hiring and job descriptions?

Make AI expectations explicit before someone joins. Write into the job description which tasks the role is expected to do with AI and which it is expected to own directly, so the standard is set from day one. In interviews, ask candidates how they decide when to use AI and when to keep a task human, because that judgment now predicts the quality of their work. A candidate who can explain where AI helps and where it flattens is showing you the exact skill this moment requires.

How do you keep track of all of this?

None of this works without honest, recurring conversation. Run workflow assessments that map which tasks actually use AI today and what each one costs in time and tokens. Hold regular feedback sessions where people can say where AI is helping and where it is producing slop. Publish a short directive that shows the team how you want AI used, using the same use-for and do-not-use-for logic above, and update it as the tools and prices change. These discovery conversations across teams are the mechanism that keeps usage aligned with value, and they pair naturally with the five steps of gen AI change management.

What does this mean for your brand?

A brand that sounds like every other brand loses the reason a client chose it. The pullback from all-purpose AI use is a chance to protect the parts of your work that are genuinely yours: your point of view, your voice, and the judgment your clients pay for. Use AI where it gives you leverage, keep people on the work that makes you distinct, and put the governance in place to keep both true as costs rise.