A Call Center Contact Costs $13. A Company Agent Can Do It for Under a Dollar

A typical inbound call to a contact center, counting labor and systems, costs about $13 to fulfill. With AI, Bret Taylor says, you can bring that cost well below a dollar. An order-of-magnitude drop in the price of a conversation is not a rounding error: it changes which brands bother to talk to customers at all, and how often. The companies that capture the surplus are the ones with a lot of customers, not a lot of theory.

Taylor's résumé is unusually stacked even by Silicon Valley standards. He helped build Google Maps, was CTO of Facebook, co-CEO of Salesforce, chaired Twitter's board, and now chairs OpenAI. In 2023 he and Clay Bavor founded Sierra to do one thing: build customer-facing agents that wear a company's own brand. Sonos and SiriusXM are already live. On No Priors, Sarah Guo and Elad Gil walk him from the definition of an agent through business models and the devices people will talk to next.

Taylor starts with a cold splash of water: "agent" does not mean the same thing in academia and in industry. The academic version is broad — software that can reason and act autonomously, from the word agency — and that breadth turns the term into a Rorschach blot. He would rather talk about three categories that industry is actually trying to ship. The first is the personal assistant that triages your inbox, books a vacation, preps a meeting, and runs your calendar. It gets the most airtime online. He is least bullish on it near-term, because a great personal agent demands more technology than exists today: the surface area of reasoning and systems integration is almost infinite. He expects that lane to belong to Apple, Google, OpenAI, and other firms that already own a consumer front door. Startups can carve niches and expand as the models improve.

The second category is what he calls persona-based agents: systems that do a specific job. Harvey does legal work; a cluster of coding agents does the job of a programmer. He calls these cases narrow but deep. Task scope is tight, the tools you touch are tight, and evaluation is unusually honest — coding agents already have benchmarks, compilers emit errors, and integration tests catch the rest. His core observation is that after foundation models arrived, a lot of effective AI is just taking a research problem and turning it into an engineering problem. Those markets can produce meaningful companies, because you need AI skill and domain skill in the same building.

That logic is why he is watching analysts next, especially back-office people who pull data, write the deck, and hand a conclusion to a stakeholder. Summarizing, synthesizing, and reasoning are what large language models are already good at; tables and numbers are the gap, which is why fine-tuning and domain knowledge matter. He is not sure "analyst" is one job or many, but he is sure the work gets a huge lift. The metaphor he wants is an Iron Man suit for analysts, not a pink slip.

The third category is Sierra's: company agents. In 1995, existing digitally meant a website and a listing in the Yahoo directory, with the URL on the business card. In 2025, he argues, it will mean a branded AI agent that can do everything the website can — ask about products, buy something, get service. The point is not automation for its own sake. It is the form your company takes in a conversational world. This one is shovel-ready, because processes and systems of record already exist; the remaining design question is what conversational experience you want around the brand. Look up early websites on the Wayback Machine and you will see a phone number, or a page that looks like a DVD intro. Today's agents will look just as quaint. Consumer companies come first because of that $13 math: conversation used to be so expensive that brands hid the phone number, and entire websites exist to help you find it. Now conversation is cheap enough to put at the center of the experience. A B2B firm with 100 customers can still benefit. The scale, and the decisions you make, are simply different.

The live examples make the job concrete. If a Sonos speaker throws the dreaded flashing orange light, you chat with the Sonos AI — powered by Sierra — to onboard or to figure out whether the problem is hardware or Wi-Fi. SiriusXM's agent is named Harmony, a name Taylor enjoys, and it can upgrade or downgrade a subscription or talk to someone who just got a trial with a new car. These are branded, customer-facing systems. Brand is not decoration. It is the experience.

The industry's default technique is retrieval-augmented generation: hang a knowledge base on a large language model and tell it to answer from your documents, so an off-the-shelf model can sit on proprietary data. Taylor's verdict is blunt. That class of investment makes a beautiful demo and is woefully insufficient for almost any meaningful customer experience. Think about the last times you actually engaged a brand you care about. How many of those conversations were questions? Almost none. You were upgrading or canceling, returning an order, doing a warranty exchange, filing an insurance claim. Each of those is an action across ten-plus systems of record, wrapped in a process that also has a business goal — talk you out of canceling — and a compliance goal, such as HIPAA limits on what a healthcare-adjacent firm may say.

Pretty answers are demo difficulty. Completing the job is product difficulty. So Sierra's real technical bet is a platform that can orchestrate a process of arbitrary complexity, with agency and guardrails at the same time. Fast-forward three or four years, Taylor says, and an insurance company's agent will file the claim, compare plans, and add a child to the policy when they are old enough to drive. That is the product, not the chatbot window around it.

Enterprise software for the past two decades has mostly been rules engines, whether the rules lived in source code or a low-code tool, executing fast. Taylor thinks the expression language is changing. A business process is no longer only a decision tree. It is three statements: what you are trying to achieve, where the AI is allowed to be creative, and where it is not. Customers often start by demanding precise control of every step. The platform can do that, and the result is robotic — you have switched off the creativity that made ChatGPT the fastest product to 100 million users. Crank agency to Spinal Tap 11 and you get hallucinations, policy violations, or a merely mediocre brand ambassador. The real problem, he says, is how you program against non-deterministic, creative software so you keep the delight and can still look a board, a CEO, and a customer in the eye about the guardrails.

In practice Sierra sits with the customer and formalizes the process. An agent, in their telling, is factual knowledge plus procedural knowledge, including the systems it has to touch. Then they experiment, put it in front of real people, and see whether it met the expectations anyone had on paper. AI should not be a technology-team exclusive. The customer-experience team — maybe in the office of a chief digital officer — should have their hands on the steering wheel. Sierra built audit and improvement tools for that reason. The work is never done.

A website is a menu. An agent is a free-text box. Taylor's analogy is Yahoo directory to Google search. Menus you click; a text box you can ask anything, so after launch you get a long tail of needs nobody designed for. Guo has seen wild fraud-return cases where nobody knew what was going on. Taylor reaches for *The Long Tail*: when the web moved from directories to search, the long tail of blogs became the living part of the network. Customer experience is on the same path. Treat unplanned conversations as a paint-by-number list of scenes to fill in. The agent is literally the voice of the customer, always on, less like an A/B test than like raising an organism.

It is also a brand ambassador. Instruction tuning makes models mirror the other person's sentiment and tone; a brand can still set the register — irreverent for a streetwear label, austere for luxury. Implementation is a mix of prompts, post-training, and supervisor models. Sensitive lines, such as no medical advice and no financial advice, go to those supervisors. The office joke is that the solution to every problem in AI is more AI. Taylor's brand example is Apple: Cupertino campus, retail store, and unboxing all smell like the same company, and an agent should too. Language is the other thing that used to be impossible. The agent answers in the customer's language, including slang, jargon, and idiom. Staffing a phone bank in 30 languages never made economic sense; retraining 5,000 agents on a new product took weeks; the model updates with a button. Reflecting the language someone actually speaks is, he says, a remarkably empathetic luxury.

Gil asks the question every application founder lives with: won't OpenAI and Anthropic just fold this platform layer into the model? Taylor zooms out and quotes Mark Twain: history does not repeat itself, but it rhymes. He is betting the rhyme is the last fifteen years of cloud. Cloud consolidated into a handful of IaaS vendors because data-center CapEx has economies of scale. A startup could not sanely build a data center or an IaaS business. He thinks frontier models will consolidate the same way, into a relatively small number of companies doing pre-training, again because of CapEx, not because they have a monopoly on researchers. Heavy assets have to be rented to as many people as possible, and a lot of pre-training startups are discovering the return math is ugly.

Guo wonders whether model companies simply merge into the clouds, since applications already run there and calling out adds latency, a second purchase, and a second security review. Taylor buys only half of that. His picture is cloud providers in partnership with the big research labs — roughly today's landscape. On latency and security he is less convinced: large enterprises already run multi-cloud and SaaS, and after twenty years of software-as-a-service they have stopped asking where the power comes from. They ask what the SLA is. In the middle sit tools companies, the pickaxes: Snowflake in cloud, Scale AI in this wave. On top sit solutions. Most firms never wanted to build their own CRM or ERP; the total cost of ownership was nonsense. The same will be true of a customer-service agent or a legal agent. Work with Sierra or Harvey rather than rolling your own.

His test is sharp. If every time a new model drops, your product's value shrinks, you are probably not a solution. You are a slight add-on, sitting in what he calls the uncanny valley of value, and he thinks a number of startups smell like that. Sierra's customers, on a platform they call Agent OS, define goals and guardrails. When a better model arrives, they plug it in and case resolution and customer satisfaction go up. Stronger models make the platform stronger, the way a SaaS product should quietly get better when the stack does. Guo notes that people said the same thing about SQL wrappers, and Salesforce, Shopify, and ServiceNow are still good companies. One more operational detail: a hand-tuned prompt often breaks when you swap models, because tokens and weights fit tightly. Most companies should not live at that layer, any more than they want to know about a database migration. A real solutions vendor turns the new thing on for you.

Sierra's commercial bet is outcome-based pricing: charge for the job done. Taylor thinks that may be the most important business-model idea since subscription software. The best AI companies are lining their invoice up with the customer's business result, which is a different species from paying for tokens. The analogy he likes is advertising's move from impressions to clicks. Guo and Gil both lived that shift at Google: the closer you get to direct attribution, the more customers will pay. AI is the first time software can complete a countable job — it analyzed this contract, it resolved this contact — and enterprises should use that to hold vendors to a higher standard.

Ask a CIO whether they are getting the value they hoped from the software they bought, and Taylor says the blood drains from their face. Twenty years of productivity claims were indirect. Everyone can make an ROI slide that assumes each employee gets 5% more of something; almost nobody checks whether the department actually got smaller. Gil cites Klarna: after automating support workflows, net promoter score and customer satisfaction went up, time per customer went down, and the team shrank by about 700 people. Taylor then names the first-order value hiding under the case study. The number-one reason a customer experience is bad is waiting. You call, you are tenth in line, and by the time a gifted human picks up you are already unhappy. That is not an indictment of the people who used to do the job. It is a scale mismatch. AI's opening is instant. He tells customers not to overthink it. Instant gratification is the main value of these systems. The rest is gravy.

Guo asks whether, now that voice works and video avatars exist, the next generation of human-computer interaction still runs through the phone. Taylor's last-fifteen-years narrative is that the smartphone ate almost every adjacent device. He would not swear to the number, but he would not be shocked if touchscreens account for 90% or more of human-computer interaction, with keyboard and mouse retreating to professional tasks. Conversational interfaces, in his view, crossed an inflection around GPT-4 — a date he admits is debatable — the way multi-touch let people put down BlackBerrys. Will multimodal voice reshape the thing in your pocket?

His guess is that the supercomputer in the pocket remains, coupled to AirPods, CarPlay, and other modalities that steal more of the interaction. He is open-minded about a smart-speaker comeback; 2015 and 2016 put an Alexa on every counter, and better models might make those boxes into computers again. Video avatars excite him more: FaceTiming with a brand is a cool idea. What he actually cares about is whether AI can filter the bag of push notifications so we stop picking up the phone on reflex, and whether technology can recede into the background instead of stacking on top of the screen we already have.

The App Store analogy is the one he wants product people to steal. Early apps were flashlights, skeuomorphic copies of hardware. Later people asked what GPS plus a screen plus the network could become, and you got WhatsApp, Uber, Instacart, and DoorDash. Conversational-first software, he suspects, is at the flashlight stage. On the human relationship with agents he gives the idealistic answer: with the right safety rails, software acts on our behalf, and we get to be present in the world instead of poking buttons. Games and the metaverse can stay a serious form of entertainment. You can still put the phone down, take off the headset, and let the agent do the errand. The point of the technology, if it works, is that the technology melts away.