Yes, but not for the reason the headlines suggest. This week SpaceXAI, the company behind Grok, opened a marketplace where you can copy pre-built AI agents into your own account, and its showcase agent, Haggle Bot, reads a company's software bills and hunts for subscriptions nobody uses. SpaceXAI's own case study claims the bot found more than a hundred thousand dollars of waste in its first week. That figure is company-reported and deserves some suspicion, but the launch still matters to small businesses, because it points at a problem I see in almost every client I take on: you are paying for software you have forgotten you own.
I build websites for small businesses in Chennai, and before any redesign I ask the owner what tools they run. The answer is usually a pause, then a mumbled list. Booking app, WhatsApp Business, a CRM someone installed two years ago, Canva, Google Workspace, a domain renewal, some project tool an old employee signed up for. Nobody can say what each one costs, and nobody wants to check, because the checking itself feels like work.
What actually launched this week
Grok Bot itself has been in beta since August 11, but the marketplace went live on September 4. It lets people package an agent's instructions, skills and routines as a template that others can copy into their own accounts. The catalog had 69 public bots from 43 creators across 10 categories as of launch day, and the one SpaceXAI pushed hardest is Haggle Bot, its in-house procurement agent.
The idea is simple. Haggle Bot connects to the places a company's money and usage already live, things like Ramp for spend, Drive or Notion for contracts, Slack for colleagues. It builds a renewal calendar from billing data, checks seat counts against actual usage, looks for duplicate tools, shops real alternatives against current prices, and drafts the emails asking for a better rate.
Two details about how SpaceXAI runs it are more interesting than the savings claim. First, the agent never signs, buys or subscribes to anything. A human approves every vendor-facing message, and staff edit the emails before they go out. Second, the instructions tell the agent to never reveal usage data, seat counts or internal plans to a vendor, because that information is exactly what kills a negotiation. That is a genuinely good design for a trainee with access to your accounts, and I will come back to it.
On the money: the case study reports $85,662 a year in unused software SKUs from one product, 43 paid seats with no activity in 90 days, and a weekly office supply order that got shopped around and cut from $14,629 to $6,143. Coverage this week noted the figures are the company's own, with no invoices or vendor names published, so treat them as a demonstration of what such an audit looks like rather than a market statistic.
What the agent is actually good at finding
Zombie subscriptions are the obvious one. A seat is bought for a new hire, the hire leaves, the seat renews forever. The agent spots that because it can read the usage log, not just the bill.
The quieter wins are renewal timing and duplication. Most owners I meet cannot tell me when their big tools renew, which means they negotiate nothing and get whatever price the vendor decides. An agent that keeps a renewal calendar and starts shopping three months out changes that conversation completely. And duplication is everywhere once you look: a salon running a booking app, a separate billing sheet and a third tool for reminders, all doing parts of the same job.
What I notice about the agent's real output is that it is boring in the best way. No strategy memos. Just: this tool renews in November, 74 of 210 seats have been idle for 90 days, drop to 150 seats and save roughly this much. That is the unglamorous coordination work that small businesses never assign to anyone because everyone is busy with customers.
You can do most of this audit yourself this weekend
Before you pay anyone, an agent or a consultant, do the obvious pass yourself. It takes one afternoon.
Pull twelve months of bank and card statements and list every recurring charge. You will be surprised how many there are. Match each one to a login and ask who used it in the last 90 days, not who has an account, who actually opened it. Check seat counts against headcount. Export whatever data you need before cancelling anything, because support cannot restore a deleted workspace three months later. And write down every renewal date with the price you currently pay.
This first pass usually finds the worst waste, because the worst waste is the stuff you forgot, and you remember it the moment you see the charge on a statement. The businesses that need an agent are the ones where this first pass is already done and the remaining money is hiding across twenty dashboards and a hundred seats. Everyone else just needs discipline and an afternoon.
Where the agent genuinely earns its keep
The case for an agent gets strong at a specific size: when you have more tools than one person can hold in their head. A fifteen person agency with a design tool, a project tool, a CRM, email, storage, a code host and a half dozen subscriptions per department has real money scattered around, and the person who could find it is busy doing client work.
What an agent adds at that scale is not intelligence, it is diligence. It reads the vendor bills, matches them against usage, notices the seat that went quiet when someone left in March, and chases it before the renewal. It never gets bored and never decides it is someone else's problem. The companies that build these things are careful to say the output still needs a human owner, and that is true, but the difference between a tool that nags you with a list and one that assembles the list itself is the difference between a chore and no chore.
What would I let an agent do, and what not?
If you do hand an agent access to your accounts, copy the permission structure SpaceXAI published with Haggle Bot, because it is the right one. Read access to spend and usage data, no questions needed. Anything that goes to a vendor, explicit approval every single time. Signing, buying, subscribing or approving charges, never, under any circumstances.
In practice that means the agent can investigate and draft freely, but it cannot commit you to anything. That division is the whole trick. The moment an agent can cancel or sign on its own, the risk stops being about missed savings and starts being about a wrong click on the wrong account with no undo. I would also think hard about what it can read. An agent that needs billing data does not automatically need your customer records or your payroll.
The part most owners skip: your tools do not talk to each other
Here is the thing nobody in the agent marketing says out loud. Haggle Bot only works well at SpaceXAI because its spend, contracts and usage data live in connected, readable systems. Most small businesses do not have that. Your bookings sit in one app, your bills in another, your customer list in a spreadsheet someone maintains by hand, and your website, if it is old, does not feed anything at all.
That disconnectedness is the real reason software waste accumulates. When nothing talks to anything, no single view of your tools exists, so nobody owns the question of whether they are worth it. Fixing that is less about buying an agent and more about getting the basics connected: a website with real enquiry forms, a CRM that actually receives those enquiries, and billing that you can export cleanly. I wrote last week about which daily tasks AI agents can genuinely take over, and the answer there starts the same way. The agent is only as good as the systems it can read.
If you run a salon, a clinic or a small agency, the practical version of all this is small. Do the statement pass this weekend. Cancel the obvious dead weight. Put one person in charge of the tools you keep. Set reminders on the three biggest renewal dates. That alone will likely cover the first year of any agent subscription you were thinking of buying, and it costs you one afternoon.
Then, if you still have twenty tools and no idea which ones matter, an agent starts to make sense, and I would be happy to help you get the website and data flowing in a way that makes that possible. See what I do, look at past work, or start a project.
The launch this week was useful for one reason above all: it showed what a competent software audit looks like when someone actually does it. The agent is a neat demo of that idea. The idea itself, that your software bill deserves an owner, works whether the owner is you, a spreadsheet, or a bot.