AI at work is evolving

The difference I am drawing is treating AI like a partner in the process versus treating it like an employee you give work to, it goes away, and it comes back with the work. That employee path is Grok Bot. Using AI as a tool, sitting with it, is Cowork or ChatGPT Work. As people move from Excel to AI, they may not be ready to jump all the way to something they just give work to and walk away from.

ChatGPT Work is the surface you pay for on ChatGPT Business and Enterprise seats.

Work and Cowork keep you on the file

Your PM drops last quarter's deck and the metrics export into ChatGPT Work. The slides stay in a preview. She changes a chart, rewrites the speaker notes, exports a tab, drops it back, and watches the deck redraw while she corrects the file in place.

Claude Cowork is the knowledge-work twin of Claude Code. Both ride paid Claude plans. She points Cowork at a folder on the desktop or in the cloud, the deck and the metrics export together. Plugins and connectors are the path Anthropic publishes, with computer use as a preview. Chat, Code, and Cowork share one usage pool, so a long Cowork session spends the Code budget for anyone on that seat.

In Work she sits with sheets, docs, and decks in a ChatGPT window, Apps and Company Knowledge attached. In Cowork she sits with files she has, plus plugins. Codex is Work's coding twin, with a separate history, sharing credits on Business. Admins get a second loop on Claude through Cowork spend controls, and on OpenAI through Business admin and Enterprise RBAC.

She can leave a Work project running, schedule follow-ups, and come back to the sheet, doc, or deck in ChatGPT. She can leave Cowork running in the cloud on scheduled tasks against the same folder, then come back to her files and interrupt the draft there.

Grok Bot works on its own computer

Grok Bot sells a different job. You message it like a colleague and it signs into real apps, each bot on a cloud computer that keeps going after you close the laptop. SpaceXAI and Cursor launched the public beta on 11 Aug 2026 and widened access on 21 Aug 2026. Those posts are the buyer sources; treat the demos as vendor demos. Access rides Cursor Pro+, Ultra, and Teams, plus SuperGrok Plus. There is no standalone Grok Bot SKU. Enterprise is a waitlist on those pages.

Your controller has the consolidating workbook open. She watches Work or Cowork fill a column, sees a SUM that double-counts intercompany, edits the formula, and keeps the pack under her signature. Dispatch that job to a Grok Bot and she waits on a file from a computer off her desk. She signs packs she has watched, so she buys the window that keeps the formula on screen.

Overnight jobs she trusts are a different buy. The public pitch is a bot that signs into the vendor portal, pulls the invoice, files it, and pings Slack, with multiple bots in parallel, a routine taught by showing the job once, and work that continues after the laptop lid shuts. Launch posts, Cursor pricing, and x.ai/bot are the public sources for that framing.

Included usage, as published

Read the included usage in the vendor's unit. Figures below are estimates from public pages, checked 22 Aug 2026. No invented token floors.

1/ ChatGPT Work on Business. Per-seat included plan limits first. OpenAI does not publish an absolute Standard token or message pool. Premium, from the 10 Aug 2026 seats post on openai.com, is 5x Standard, with no five-hour usage cap and weekly resets. After included limits, Work and Codex burn shared workspace credits, priced as credits per 1M tokens on the rate card (help.openai.com). Instant chat is described as unlimited subject to abuse guardrails. Enterprise is contact sales. Flexible deals use a shared credit pool with no default per-seat cap.

2/ Claude. Team Standard ($20 annual / $25 monthly) gives more usage than Pro ($20). Premium ($100 / $125) is 5x Standard. Limits reset on a rolling five-hour window plus weekly caps. There is no fixed message count. Chat, Code, and Cowork share the pool (claude.com/pricing). Enterprise is $20 per seat per month billed annually, plus usage at API rates, with $0 included tokens on the seat. Usage bills from the first token.

3/ Grok Bot on Cursor. Other Models pools list at $20 / $70 / $400 on Pro / Pro+ / Ultra (Cursor docs). Teams Premium is 5x the Standard team allowance. Standard's Other Models dollar figure is unpublished. Cursor Models are a separate pool vendors call generous; exact size unpublished. Grok Bot adds a weekly included allowance: Ultra highest, Pro+ below, Teams by seat. Absolute Bot units in steps, tokens, or hours are unpublished (cursor help + x.ai/bot). Overage continues on-demand if you enable it.

Buy the workbench. Then the coding remainder.

You can buy tokens on an API and assemble the rest, which means you own identity, connectors, approvals, the preview window, spend caps, and the audit trail. The workbench sells that stack: SSO, plugins, Company Knowledge, Cowork spend controls, a bot with a computer, a draft you can sit with. I called that stack the harness in an earlier piece, because models get swapped under you and a successor inherits it.

Engineering buys a coding plan on top of the workbench.

1/ Assist near $19 per eng seat. GitHub Copilot Business lists at $19 with 1,900 pooled AI credits per user per month (Enterprise $39 / 3,900). One credit equals $0.01. Completions stay outside the credit pool. Gemini Code Assist Standard lists at $19 on a 12-month commit ($22.80 monthly) with 1,500 agent and CLI model requests per user per day (Enterprise $45 annual / 2,000 per day). Amazon Q Developer Pro lists at $19. AWS blocked new signups on 15 May 2026 and ends IDE plugin support on 30 April 2027. The public migration path is Kiro, from $20 (1,000 credits) to $200 (10,000 credits), add-ons at $0.04 per credit.

2/ Cursor is the AI-native IDE that also carries Grok Bot on Pro+, Ultra, and Teams. Windsurf is folding into Cognition's Devin Desktop. Public individual Pro sits near $20 and Teams near $40 on the Mar 2026 plan post. Those vendor pages returned 429 on the 22 Aug 2026 scan, so confirm live before a contract.

3/ Claude Code shares Anthropic plan limits, or burns API tokens on Enterprise. Codex shares credit logic with ChatGPT Work. Keep the workbench and the IDE as separate buys until a team runs both jobs from one seat.

Prove it on one team: pick the loop they sit in, the draft on screen or a bot on its own computer, and instrument usage in the vendor's unit. Expand on pull, when a second team shows up with a use case and a budget owner. Wire identity, connectors, and approvals so the workbench runs after you leave.

Prices and meters checked 22 Aug 2026 against public vendor pages. Confirm at purchase time. Estimates are assumptions, not Finance quotes.