NOTES: Salesforce (Agentforce): The Great Agentic Fleecing of 2026
A critical teardown of the Agentic Work Unit, exposing the hidden compute costs, LLM token volatility, and systemic drift within Salesforce's autonomous agent ecosystem.
Listen on
Listen (episode)SpotifyApple PodcastsRSS
The Cassandra Files — forensic audio drama. Katie audits the books, Marcus kills the spin, Killian opens the file. About · Latest · Themes
I’ve seen my share of enterprise software mirages. Back in Shinjuku in 2018, I watched slick founders peddle algorithmic vaporware that evaporated faster than a spilled Asahi on August asphalt. But what Salesforce is pulling with Agentforce? It’s a beautifully orchestrated fleecing. They’ve built a casino where the house always wins, but nobody can actually account for the chips.
The official yarn from Marc Benioff is a utopian fantasy where "AI elevates humans." They are flogging this dream hard to pump up a $1.8 billion ARR narrative before those 2027 IPO rumors crystallize. In Q4 FY2026 alone, they closed 22,000+ Agentforce deals, promising frictionless scale, 85% autonomous resolution, and lightning-fast 2-5 week deployments. To budget for this miracle, they invented the Agentic Work Unit (AWU). It looks brilliant on a boardroom slide deck.
But peel back the spin, and you’re staring at pure cactus tech.
The AWU is nothing but a jargon-laced friction point masking variable LLM inference costs. At $0.10 per action under their Flex Credits model, enterprises are bleeding cash into a black box. Salesforce processed 771 million AWUs and served up an eye-watering 11.14 trillion tokens in Q4. Why the massive volume? Because of Agent Drift.
Without real-time drift detection—a basic safeguard mandated by the EU AI Act that Agentforce's Trust Layer conspicuously lacks—these autonomous agents become liability time bombs. They just spin in recursive loops. When you ask them to handle multi-step workflows, escalation rates instantly spike to 15%. That "seamless" voice infrastructure? Live architects report it routinely fails SIP integration during high-volume call surges. And that two-week deployment promise? It took Siemens twelve grueling weeks just to tune a basic lead qualification agent.
The cultural rot here is profound. We are replacing human workers with a system that hits a physics wall. RAG latency limits these bots to a maximum of nine skills, and retail agents default to a pathetic 1-2 actions outside of peak seasons. The human cost isn't just the specter of lost jobs—it's the exhausted engineers left manually patching a hollow audit trail while competitors like Freshworks easily hit 92% compliance scores. Meanwhile, Oracle is already undercutting them at $0.07 an action, and Microsoft is shipping deterministic guardrails.
The Bottom Line: Salesforce isn't selling an autonomous workforce. They are masking compute cost volatility behind a proprietary metric to fund their trillion-token OpenAI bill. When the vendor charges by the task, and the vendor's machine autonomously decides how many tasks to run, you aren't buying efficiency. You're just paying the house.
