An AI agent found $100K in SaaS waste in its first week — but the checks it runs take an afternoon
xAI shipped a marketplace for Grok Bot templates, and the one they built for themselves is the interesting one. Haggle Bot is a procurement agent: it connects to Slack and Ramp, finds unused SaaS seats, price-checks recurring purchases, and negotiates vendor contracts. First week, they say it identified over $100K in savings.
That's their number about their spend at their scale, so don't anchor on it. But the checks are the same four that show up in almost every software budget:
- Seats billed for people who aren't using them. Pull the user list from each subscription, compare last-login against what you're charged for. Anything untouched for 60+ days is where you start.
- Two teams buying the same capability. Marketing has an email platform, sales has another, nobody knows. Compare categories, not vendor names.
- Intro pricing that quietly expired. Year one discounted, year two at list. Nobody decided to accept the increase.
- Per-seat plans where most seats only read. Downgrade to a viewer tier instead of cancelling — easier to get approved, often most of the saving.
The real argument for the agent isn't that it's smarter than an audit. It's continuity. A manual audit is accurate in January and wrong by April, because nothing in it updates itself. A bot checking weekly catches a departed employee's licence in week two instead of eleven months later.
My suggestion: do it by hand once. Two hours, and you'll know whether the bot is worth wiring up.
Full breakdown: https://choosesaas.com/guides/ai-agent-saas-waste-audit
The continuity argument makes sense, but I’d measure accepted savings rather than identified savings. A seat flagged as unused can still be contract-bound, seasonal, or expensive to restore; mistaken removals create review and reprovisioning cost. The weekly report should separate dollars identified, approved and actually realized after renewal, plus analyst minutes, reversals and false-positive removals. The first manual audit can also become the baseline and eval set for the agent. Did they publish how much of the $100K moved from identified to realized?
The continuity point is interesting — but it also creates a second problem once the agent moves beyond identifying waste and starts negotiating or initiating consequential procurement actions.
What happens if the authority behind an action changes between the agent identifying the opportunity and the downstream action actually executing?
Technical access to Ramp or another system can still exist. The original instruction can still exist. Neither necessarily proves the agent remained authorised at the moment of consequence.
That authority-at-execution gap is what we’ve been working on with OpsWatch. Procurement seems like a particularly good real-world test case.