Salesforce flagged a data overage. The account had jumped from roughly 5GB to 40GB in about 30 days, around 21 million records. No human on the team had typed any of it. Every record came from agents running production workflows across the CRM.
That detail, from Jason Lemkin's latest episode of The Agents (#013), is the clearest signal yet that the economics of systems of record are breaking under agentic load. If you're running demand gen on top of a CRM you pay per-seat for, the math is about to change whether you plan for it or not.
The Pricing Model Assumes Humans
CRM pricing was built around a simple premise: people log in, people type, people run reports. Seats scale linearly with headcount. Agents don't work that way. They write data continuously, query at machine speed, and generate records at volumes no human team would produce. Lemkin's team estimated storing those same 21 million records on Postgres would cost roughly a thousand times less than on Salesforce.
Salesforce's own telemetry shows the average organization now runs 13 AI agents across its SaaS portfolio, up from 5 in February 2025. Average time to create and activate an agent dropped 53% to 1.9 days. Workload per account grew at a 31% compound monthly rate. Vendors are responding by raising API prices. A 20% bump on API calls feels manageable until your data volume is growing 10x to 100x. Lemkin's team takeaway: customers increasingly want costs tied to outcomes, not usage.
A Renewal Agent Built in Half a Day
Lemkin's team built a renewal agent in roughly half a day. It pulled data from multiple sources, drafted custom proposals, and generated multiple versions targeting different sponsor tiers. The initial output was missing context, so they added a step: the agent now drafts a narrative before generating the deck. A human reviews the narrative, approves or adjusts, then the deck gets built.
Salesforce reported each deployed agent automates an average of six business actions. But accuracy is a real problem. Lemkin's team found the AI occasionally generates inaccurate numbers despite explicit instructions. Their operating rule: let the agent initiate contact, but humans handle follow-ups. A Gartner-based survey backs this up: 87% of customers want the option to contact a human when AI is used in support. For B2B renewals, where contract terms and account history matter, fully autonomous agents carry real retention risk. The hybrid model (agent drafts, human reviews, human handles escalation) isn't a compromise. It's the architecture that holds.
When Your System of Record Becomes a Competitive Weapon
ServiceTitan ended a nine-year partnership with Podium this year, affecting about 1,000 shared customers. Podium had shifted to an agentic model competing directly with ServiceTitan's core services. ServiceTitan's position was blunt: partnerships can't be used to displace core functionality.
If your agent strategy depends on third-party integrations into someone else's system of record, you're exposed to exactly this kind of policy change. Certification requirements, marketplace rules, API restrictions can all shift when the platform owner decides agents are their product now. And 79% of senior executives say AI agents are already being adopted in their companies, so every major platform vendor is thinking about this. Lemkin's team described their own architecture as a "headless Salesforce," with an AI VP of Revenue (called 10K) operating across multiple integrated systems. The irony they noted: the more integrated the architecture became, the easier it got to leave any single vendor.
What to Measure Before You Ship
If you're piloting agent workflows in renewals, support routing, or proposal generation, instrument these from day one:
- Primary metric: qualified pipeline influenced by agent-initiated touchpoints (not just ticket deflection)
- Secondary metrics: time-to-first-response on renewals, proposal acceptance rate by tier, data accuracy rate (human corrections per 100 agent outputs)
- Guardrail: escalation-to-human rate. If it drops below your baseline, customers may not be finding the handoff. That 87% stat is the stop-loss threshold for customer trust.
The hypothesis: if we deploy a renewal agent with human-reviewed narratives, then proposal turnaround drops below 4 hours and acceptance rate on mid-tier sponsors improves, because speed and personalization at that tier were previously bottlenecked by headcount.
The trade-off you're accepting: data costs will rise before you renegotiate storage terms. Your CRM bill may look alarming before the pipeline impact justifies it. Accuracy will require a human review layer for anything touching revenue, which means you're not cutting headcount. You're reallocating it.
Lemkin's 40GB surprise wasn't a failure. It was the first invoice from a future where agents are the primary users of your systems of record. The teams that instrument, measure, and design handoffs now will own the operating model. The ones that don't will get a storage overage notification and wonder what happened.
