What xAI’s Haggle Bot can’t see
xAI’s procurement agent optimizes what it can measure. Most of what procurement calls value falls outside that view.
Naming a procurement agent ‘ Haggle Bot’ tells you exactly what the builders think procurement is for.
In early September, xAI published a brief account of having their Grok bot go live on the company’s own vendor spend. The bot was connected to Slack, Notion, Drive, Gmail, Hex and Ramp. It mapped about 125 active vendors and, according to the post, identified over $100k in direct savings. It could research and operate internally on its own, though anything relating to money, terms or a note to a vendor required a human to approve that.
The post is well written and in places quite refreshingly honest (more on that later). It also demonstrates a very narrow view of procurement, one in which value is simply a lower number on the next invoice. That view is easy to automate and that’s one of the reasons it’s so attractive. The problem is that it ignores almost everything procurement is meant to do.
Get rid of the name and most of what Haggle Bot did was spend hygiene. It found 43 paid seats on one SaaS product with no activity in 90 days, worth $14,220. It flagged $85,662 per year in unused SKUs on a month-to-month product. It also pushed back on a renewal quote that was adding seats before any demonstrated usage, priced the alternatives and drafted a counteroffer. And for a weekly order of new-hire kits, it compared prices across Amazon, Costco, Uline and Walmart, swapped in another brand if it couldn’t find the product in question for less and then handed the final order back to a second agent to place.
The renewal counteroffer is the only real haggling in the post, and people set the minimum quantities, edited the draft and approved the send.
On the autonomy scale I used in my sourcing automation post, this falls somewhere between assisted sourcing and decision support. It is not bounded autonomy, and it is certainly not autonomous negotiation. The design itself is sound; the permission model (free to research inside, gated for anything that leaves the building) is the risk-tiered approach I would recommend to anyone. What doesn’t hold up is the framing, since a bot named for haggling spent most of its time on housekeeping.
Goodhart’s law is usually put in the anthropological form of Marilyn Strathern’s: when a measure becomes a target, it ceases to be a good measure. An agent incentivized to save will optimize what it can see and Haggle Bot can see unit prices, seat counts and usage logs. Its own examples show what is outside that view.
Start with the office/IT supplies. Swapping brands every Friday lowers the basket price. It also means new hires get different docks and chargers from week to week, so IT supports a growing mix of hardware and warranty terms vary by brand. None of that shows up in the 58%.
The renewal is a subtler case because a counteroffer backed by competitor quotes works once. Run every renewal that way, every year, and the vendor learns something about its customer; whatever that costs in service, priority, or goodwill won’t appear in a spreadsheet either. (It also raises a question for another article: what happens when the vendor’s side runs an agent too?)
The 43 ‘idle’ seats tell a larger story because getting rid of them is the right thing to do, but an idle seat is generally one that hasn’t been touched for 90 days (it’s usually indicative of a failed rollout, a team that stopped recruiting or a tool that simply didn’t work in the process). Canceling the licenses is just treating the symptom. Instead, procurement should ask why the company bought capacity nobody ever used. To its credit, the agent did go around and chase the owners on Slack to find out the company’s plan for one of the vendors, but none of the outcomes reported go beyond cancellation.
This is the classic price-versus-value argument made tangible. Procurement value includes all of the following: total cost, supply continuity, quality, risk and the relationships that will get a supplier to pick up the phone when something goes wrong. It also includes the cheapest option of all: not buying. A price comparison agent is very good at one element of that list.
I learned that back in the days when I was a buyer. We’d select a supplier on the basis of an excellent purchase price and the decision would look absolutely sound on paper. The problems only emerged when deliveries started to be late and the quality declined. We then had to pay for air freight to expedite the deliveries and we still ended up with stock shortages and were forced to bring in a second source. By that stage, the supplier we’d selected on the basis of the best price had become extremely expensive, and those additional costs hadn’t been part of the initial comparison that made the original choice look so good. A price comparison agent looking at the same quotes would have picked the same supplier.
The headline figure is worth a bit of scrutiny. The post states the agent “identified” over $100k in direct savings, “identified” is not the same as “realized”. The post only references the 14,220 recurs and the office supplies reduction is from one run of one order. The maths here is important because the two SaaS figures total $100k and, if you simply add them together as a single figure, you’re combining annualized savings, one-off savings and a price gap on a basket which may or may not have been like-for-like (the post does mention brand substitutions but doesn’t quantify how much of the 58% that accounts for).
Get Bertrand Maltaverne’s stories in your inbox
Join Medium for free to get updates from this writer.
Anyone who’s ever handed a savings figure to a CFO will know the follow-up questions: is it a reduction or an avoidance? Is it an identified or realized saving? Is it a one-off or a recurring saving? And crucially, is it on the same scope as before?
The post doesn’t state what it cost to run. Two agents (the Haggle Bot and a separate Amazon Bot that handles weekly orders) pull data from six different sources. For xAI, given that they’re running their own model, the marginal cost is essentially zero. For anyone else replicating the setup, it’s a real cost, and the comparison that matters is the savings against running costs plus the staff hours that would otherwise be spent reviewing the agent’s output. The post is quite clear that “we still go through and edit the emails to fine-tune the tone and to make sure we’re giving the vendors the right level of information.” That’s a very honest statement (the second part is about judgment, not tone), and those editing hours are part of the business case.
The post is also completely silent on the security implications. An agent that reads Gmail treats every vendor email as input and an agent with access to Ramp has something that could be attacked. Prompt injection via a supplier’s email is the obvious scenario and a case study making that argument should at least mention it.
None of that renders the experiment pointless. Hygiene spend on the tail, and indirect spend, is a very sensible first job for an agent. The risk is low, the outcome is measurable, and the data is already in-house. You could call it the easy end of procurement and it absolutely makes sense to automate it.
Agents are best at structured tasks (usage data, price tables, a calculator sheet, etc.) and absolutely terrible at tasks where the relationship is expressed via language. In the xAI case, the agent produced the comparison sheet without any assistance, while the humans still had to go back and rework the emails.
One final point in the human-in-the-loop discussion. People approved every outbound email and the 43 names were then sent back for review prior to the downgrade. The post makes no mention of who actually tested the analysis underpinning the renewal recommendation or the price comparisons. Were the alternative brands really equivalent? Did anyone actually check? If the reviewers are only looking at the ‘feel’ of the output and not actually assessing the logic then the approval becomes a formality.
If a vendor or an internal team is pitching something of the sort, then a few questions will take you from demo to deployment:
- What is the agent optimizing and what is outside its view?
- Who defines a saving and does finance validate that figure?
- How much of the savings identified is actually realized six months later?
- What does it cost to run (token cost plus human review time)?
- What untrusted input does it read and what can it do with that input?
- What’s the supplier’s experience on the other side?
One more thing about the xAI setup has nothing to do with cost savings: there was no dedicated procurement suite in the stack at all. Just a generic agent, Ramp (the spend management platform) and the company’s collaboration tools. That’s the subject of my next post! Stay tuned.
In the meantime: if your team deployed a Haggle Bot tomorrow, which part of your value story would it miss first?
AI disclosure: I research and edit with AI; the arguments and judgments are mine
Originally published at https://procurementtidbits.substack.com on October 26, 2026.
