Why Postscript Shut Down Its 100-Person Sales Center to Deploy AI

Three and a half years ago, Postscript opened a massive office in Phoenix, Arizona. They called it the E-commerce Sales Center. It was staffed with “tons of humans” — people trained to respond to customer questions, handle orders, and manage customer interactions via SMS.

It worked. Brands got real human service. Customers got responsive, personable interactions. Postscript built Shopper, a conversational commerce agent, to replace it.

That transition reveals something important about how AI agents are actually deployed in production: they’re not replacing jobs because they’re smarter. They’re replacing jobs because they can scale when humans can’t.

The Economics of Human Chat

The Phoenix Sales Center was a good business. Brands paid Postscript to handle customer conversations. Postscript paid humans to have those conversations. Margin was solid, customers were happy.

But economics constrained growth. Every new customer meant hiring more staff. Every conversation required a human. You could train people to handle edge cases, to build rapport, to sell. But you couldn’t have one person respond to 10,000 conversations in a single day.

“We have this thing called Shopper right now,” Alex explains. “Like three years ago, we started a business, maybe more, three and a half, called SMS Sales. And SMS Sales was a giant office in Phoenix, Arizona called the E-commerce Sales Center where we staffed up tons of humans.”

The Sales Center represented a ceiling. Great customer experience, linear cost structure, impossible to scale. When Postscript eventually wanted to offer conversational commerce to all 10,000 customers simultaneously, humans were no longer an option.

The Data Advantage

Here’s what the Phoenix Sales Center gave Postscript that most AI startups never get: data.

Every conversation had a human writing messages, making real-time decisions about how to handle questions, when to escalate, when to close. That created a dataset of thousands of human customer service interactions, all tagged with outcomes (customer satisfied, sale made, escalated, etc.).

When Postscript built Shopper, they used that data to train the AI. The AI learned patterns from how humans actually handle customer conversations. It didn’t learn from an LLM’s generic knowledge of customer service. It learned from Postscript’s specific customers, categories, and conversation patterns.

This is the unglamorous truth about AI deployment in enterprise: the companies that win aren’t the ones with the best models. They’re the ones with proprietary data from actually running the business.

The Handoff Problem

Shutting down the Phoenix Sales Center and moving to Shopper wasn’t a simple swap. It was a migration.

Brands needed to trust the AI. Early Shopper wasn’t perfect. It would misunderstand questions, give wrong recommendations, miss sales opportunities. So Postscript didn’t go cold turkey. They ran hybrid operations: Shopper handled simple conversations, humans handled escalations and complex cases.

Over time, as Shopper improved, that ratio shifted. More conversations went to the agent, fewer to humans. Eventually, the human center was no longer needed.

But here’s what’s important: Postscript kept getting richer data throughout this transition. Every conversation Shopper got wrong became a data point showing what to improve. The AI learned from failures in production, not from a test dataset.

What Made the Transition Possible

Three things had to be true:

1. High volume and low complexity (on average). Customer service conversations for e-commerce aren’t bespoke. Customers ask similar questions repeatedly. “When does this ship?” “Do you have this in another size?” “Can I get a discount?” An AI can be trained to handle these reliably, even if it struggles with truly novel questions.

2. Asymmetric outcomes. It’s okay if the AI gets 10% of conversations wrong, because you’re handling millions of conversations. Even with a 10% error rate, you’re serving more customers better than humans could. The errors are absorbed as acceptable cost.

3. Continuous learning. The AI doesn’t ship static. It gets better as it encounters more conversations. Bad outputs are captured, analyzed, and fed back into retraining. This means the early AI is deliberately imperfect — it improves with scale.

Most AI deployments fail because they expect the AI to be perfect from day one. Companies that win treat the AI as a system that improves in production, not as a finished product.

The Specialist Problem

This also reveals why specialization wins in the agent wars. Postscript’s Shopper is incredibly good at handling e-commerce SMS conversations because that’s all it does. It’s not a general chatbot. It’s specifically trained on Shopify stores, SMS-based conversations, customer service in that context.

A general AI agent might talk about anything, but it would be worse at e-commerce than an agent trained specifically for that category. Scope is a feature, not a limitation.

FAQ

Why did Postscript keep the Phoenix Sales Center running while building Shopper instead of shutting it down immediately?

Because they needed data and they needed trust. The human conversations generated training data for Shopper. And customers needed proof that the agent could reliably handle their conversations before going all-in. A hybrid operation de-risked the transition.

How does Postscript gather training data for Shopper now that the human center is closed?

Every Shopper conversation that gets flagged as problematic becomes a data point. Customer escalations, negative feedback, manually corrected responses — all of these show where Shopper is failing and what it needs to improve. This creates a continuous learning loop.

What kinds of customer service questions can’t Shopper handle?

Complex issues that require account lookup, fraud investigation, or deep product knowledge. Billing disputes, lost packages, custom product requests. These still escalate to humans. But in Postscript’s volume, these are the exception, not the rule.

How much did the Phoenix Sales Center cost to run?

Alex doesn’t provide the exact number in the interview, but implies it was expensive enough that scaling it was economically impossible. A 100-person operation in Phoenix, handling conversations all day, costs millions per year in salary alone.

Could Postscript have expanded the human center instead of building Shopper?

Theoretically yes, but it would have meant hiring hundreds more people, building new offices, managing quality across distributed teams. The unit economics make this unsustainable for high-volume customer service. At a certain scale, AI becomes cheaper and faster than hiring.

Does Shopper still make mistakes that would have been caught by humans?

Yes. The error rate isn’t zero. But the error rate per conversation is acceptable because the volume of successful conversations justifies the overall accuracy. When one person handles 100 conversations perfectly, and one mistake happens out of 1,000 total conversations, that’s still better than one person handling 5 conversations per day.

How did Postscript decide which conversations to send to Shopper vs. humans during the transition?

Likely based on confidence scoring. Simple questions with clear answers went to Shopper first. Complex or ambiguous questions went to humans. As Shopper’s confidence improved, it handled more edge cases. This is a gradual, data-driven migration.

Is the Phoenix Sales Center completely gone, or does Postscript maintain any human customer service operation?

Alex mentions closing the center, but doesn’t specify if some human escalation team exists. Large-scale AI operations typically maintain a small human team for the 1-2% of conversations that need human judgment.