September 5, 2026. SpaceXAI opened Grok Bot to enterprises on September 3, with two weeks of free usage for Grok and Cursor Enterprise customers and the right to invite an entire organization, including people who do not hold a seat. The day after, it published something more useful to a small business than the launch itself: the complete system prompt of the procurement bot it runs internally, along with the numbers it produced. The bot, which the company calls Haggle Bot, found more than $100,000 in direct savings by auditing SaaS seats, renewals and recurring purchases. For an agency or a ten-person company, the prompt is a copyable specification for the cheapest automation project most of them have never run, a seat audit of their own software stack, and the permission lines inside it are the operating rule any business should apply to an agent that can touch money.
What SpaceXAI shipped this week
- Enterprise availability with governance controls. The September 3 release adds access, network and audit controls for managing bots at scale. Each user's bots run in an isolated cloud environment, and a bot has no access by default beyond the accounts it is signed into. Enterprise customers activate it from the admin dashboard; the free window runs two weeks from launch.
- Adoption figures and where the use is. SpaceXAI says thousands of organizations have adopted Grok Bot since the August 11 launch, naming Legora, Supermicro and ServiceTitan, and that the heaviest use is outside engineering: sales bots that watch webinars by people at target accounts and leave LinkedIn and email drafts for morning review, recruiting bots that build overnight shortlists, marketing bots that pull Zoom Q&A and message the right account executives. Investing.com's report repeats the same customer names and the two-week terms.
- Haggle Bot's results, itemized. Given Slack, Notion, Drive, Gmail, Hex and Ramp, the bot mapped roughly 125 active vendors. On one product it found 43 paid seats with no activity in 90 days, worth $14,220. On another it found $85,662 a year in unused SKUs, and because that product was month to month the cut took effect immediately. On a weekly office order it shopped a $14,629 cart across Amazon, Costco, Uline and Walmart and brought it to $6,143, a 58% reduction.
- The prompt is public, in full. It has a setup block naming where spend data, contracts, seat data and dossiers live, a people block naming the human who makes final decisions, three permission lines, negotiation dials, and a worked example of a weak finding versus a strong one.
- Pricing for everyone below enterprise. Per the Grok Bot page, it is included in Cursor Pro at $20 a month, SuperGrok at $30 a month, and Cursor Teams at $40 per seat per month, with enterprise pricing through sales. The macOS download and the enterprise contact form are both served from Cursor's domains.
What the prompt actually tells the bot to do
Most agent prompts you see published are aspirational. This one is operational, and four parts of it are worth reading twice.
First, the output format. Every finding must open with four lines: TODAY, what we pay now, annualized, from live data, with the source; SAVE, the realistic saving and the mechanism; REC, one committed recommendation, with the note that a menu of options is not a recommendation; and NEXT, what the bot has already set in motion internally. The prompt bans the phrase "I can ping X if you want." Internal actions happen before the report, not as offers.
Second, the evidence bar. A real opportunity needs a dollar figure traced to live spend data, a specific mechanism, and a reason it is actionable now, such as a renewal window or a competing quote. Anything less is labeled a lead, with the missing data named. The example the prompt gives of a weak finding is "Renegotiate our CRM (~$50k)." The strong version names the tool, the renewal date, 210 seats with 74 idle for 90 days per admin logs, the saving of about $18,000 a year from dropping to 150, and that the owner confirmed.
Third, the negotiation dials. Prioritize renewals inside a 120-day radar. Open 5 to 10% below the internal target and never more than 25% off the vendor's latest quote. Give the vendor one acceptable reason per ask. Never reveal usage data, seat counts, internal projects, urgency, or that you have already decided to renew. A rejected draft is rebuilt from the plan, never edited.
Fourth, and most important, the permission lines. Reading spend data, messaging colleagues, requesting admin access and updating dossiers are always allowed. Any vendor-facing send needs the operator's explicit go, every time. Signing, buying, subscribing, approving charges or making any binding commitment is never allowed, under any circumstances.
What it means for operators
Start with the arithmetic, because the enterprise numbers scale down cleanly. A ten-person agency typically runs 25 to 40 paid tools. If a fifth of the seats on the five most expensive ones are idle, which is the pattern SpaceXAI found on both of its audited products, the saving is usually one or two months of a junior salary, found in an afternoon. The reason nobody does it is not that it is hard. It is that it is nobody's job, and the prompt's own framing agrees: the work is worth doing but hard to justify doing by hand.
The second lesson is the permission model, and it is the one we would lift verbatim into any AI automation that touches vendors, invoices or customers. Three lines: what the agent may do without asking, what it must ask about every time, and what it may never do. SpaceXAI's bot found six figures of savings without once being allowed to sign, buy or approve a charge. That is not a limitation on the result. It is why the result is trustworthy.
Third, read the caveats SpaceXAI put in its own post. The team still revises the bot's vendor emails to calibrate tone and decide how much to disclose. The results came from a short window. And the environment mattered: Haggle Bot had Ramp for spend and IT-provided last-used data for seats. If your company cannot produce a list of who logged into what in the last 90 days, the seat audit cannot run until it can, and getting that data out of a dozen admin panels is usually the first real task, not the bot.
Fourth, know whose infrastructure you are standing on. The Grok Bot download resolves to Cursor's update servers and the enterprise sales form is on cursor.com, which is now part of SpaceX. OpenAI is winding down Cursor's access to its models on November 12, a story we covered in our vendor-risk analysis. None of that changes what the procurement prompt teaches, but a bot you have trained on your workflows is a dependency, and the prompt's own advice about vendors applies to the vendor of the prompt: keep the routine written down somewhere you control.
How to run the audit this month
- Export spend. Twelve months of card and invoice data from your accounting tool or corporate card, tagged by vendor.
- Get last-login data. From your identity provider if you have one, otherwise from each admin panel for the top ten tools by spend. This is the step that takes the time.
- Fill in the setup block. Use SpaceXAI's prompt as written, substituting your systems, your operator and your dials. Keep the three permission lines exactly as they are.
- Demand the four-line format. If a finding arrives without TODAY, SAVE, REC and NEXT, send it back. The format is what makes the output checkable.
- Approve every vendor-facing message yourself. Read the anchor and the walk-away before it goes. The prompt requires the plan to be shown first; hold it to that.
- Measure in dollars a year, not in tasks completed. That is the only number that survives a conversation with whoever pays for the tool.
We covered Grok Bot's plan expansion last week and argued then that its listed jobs were the ones agencies sell. The procurement post is the first time the company has shown its full working for one of those jobs, with the prompt, the systems and the figures side by side. If you would rather have the audit built and run for you, on your own accounting and identity data, that is the kind of bounded, measurable first project we recommend to any agency starting with agents.
Frequently Asked Questions
Grok Bot became available to enterprises with access, network and audit controls for governing bots at scale. Grok and Cursor Enterprise customers get free usage for two weeks from launch and can invite their whole organization, including people without an existing seat. Each user's bots run in an isolated environment and have no access by default beyond the accounts they are signed into.
SpaceXAI reports more than $100,000 in direct savings. Itemized examples in its September 4 post: $14,220 from 43 paid seats with no activity in 90 days on one product, $85,662 a year in unused SKUs on a month-to-month product, and a weekly office order reduced from $14,629 to $6,143, a 58% cut, by shopping it across Amazon, Costco, Uline and Walmart.
Slack, Notion, Drive, Gmail, Hex and Ramp. From those it built a working map of roughly 125 active vendors. Seat usage data came from the IT team's assigned-seat and last-used records, which is the input most small companies will need to assemble before a similar audit can run.
No. The published prompt has three permission lines. Reading spend data, messaging colleagues, requesting admin access and updating vendor dossiers are always allowed. Any vendor-facing send needs the operator's explicit approval every time. Signing, buying, subscribing, approving charges or any binding commitment is never allowed under any circumstances.
According to the Grok Bot page, it is included with Cursor Pro at $20 a month, SuperGrok at $30 a month, and Cursor Teams at $40 per seat per month. Enterprise access, with the governance controls added on September 3, is priced through sales. The enterprise free period runs two weeks from the September 3 launch.
Yes. SpaceXAI published it in full with a setup block for your own systems, people, permission lines and negotiation dials. The practical prerequisites are twelve months of vendor spend data and last-login data for your most expensive tools. Without usage data the seat audit cannot produce evidence-backed findings, which the prompt itself requires before anything counts as an opportunity.