"The atomic unit of productivity in AI is a process, not a person."
That line from Sierra CEO Bret Taylor, during a recent conversation with Stripe co-founder John Collison, cuts against how most companies are actually deploying AI right now. The dominant playbook is still "give everyone Copilot and hope productivity goes up." Taylor's argument is that this misses the point entirely.
Companies Ship Their Org Charts
Taylor's core observation: most organizations optimize AI deployment department by department because that's how budgets and accountability work. But the biggest efficiency gains come from processes that span multiple teams.
His example was supplier onboarding - a workflow touching legal (contracts), procurement, IT integration, and business sponsorship. The typical timeline: 17 days. An AI agent treating this as one unified process instead of four departmental handoffs could compress it to a single day. But no single department owns that workflow, so no one is tasked with automating it.
This maps directly to what smaller teams experience too. If you're a freelancer or small agency, your bottlenecks aren't usually inside one task - they're in the handoffs between tasks. Client sends brief, you research, draft, get approval, revise, deliver, invoice. The AI opportunity isn't making the drafting step 30% faster. It's collapsing the entire sequence.
The Numbers Behind Sierra's Bet
Sierra builds AI agents for customer experience - phone, chat, email. The company hit $165 million in annual recurring revenue in roughly two years, which is fast by any SaaS standard.
The economics Taylor shared are stark. A typical customer support call costs a company $10 to $20. Sierra's AI agents handle the same interaction for 10 to 20 cents, sometimes as low as 1 to 2 cents. Clients like Ramp are automating 90% of support cases. SoFi raised its Net Promoter Score (a measure of customer satisfaction) by 33 points after deploying Sierra's agents.
But here's the wrinkle Taylor flagged that most AI vendors won't mention: one retailer saw total support volume increase by almost as much as they saved through automation. When you make support dramatically better and cheaper, people actually use it more. Economists call this the Jevons Paradox - when something becomes more efficient, total consumption often rises rather than falls.
Outcome Pricing Over Token Counting
Sierra charges per resolved case, not per API call or seat. If the AI agent handles the issue without human intervention, there's a pre-negotiated rate. If it escalates to a person, that's free to the client.
This pricing model is worth paying attention to because it signals where business software is heading. The traditional SaaS model - pay per seat, per month - doesn't make much sense when an AI agent can do the work of multiple seats. Taylor argues token consumption (the raw compute cost of running AI models) doesn't correlate with business value, so pricing should be tied to outcomes instead.
For anyone evaluating AI tools right now, this framing is useful even at small scale. The question isn't "how much does this tool cost per month" but "what's the cost per completed task, and how does that compare to doing it manually?"
Taylor also made a candid admission about AI valuations: software company valuations have dropped 20-30% recently as investors question whether AI agents will replace existing systems of record (think CRMs, project management tools). His take is that tools serving as actual records - ledgers, databases - keep their value. Tools that are primarily interfaces for human work are more vulnerable to being bypassed by agents that can work across multiple systems directly.
None of this is theoretical for Taylor. He chairs the OpenAI board and previously served as Salesforce co-CEO, so he's watching AI reshape the exact category of software he helped build. When someone with that vantage point says the playbook is changing from "AI as productivity boost for individuals" to "AI as process automation across teams," it's worth taking seriously.