A Broken Belt Buckle Turned Gucci's Call Center Into a Sales Floor

Clara Shih, CEO of Salesforce AI and one of Time's 100 Most Influential People in AI, sat down with Redpoint partners Jacob Effron and Pat Chase for episode 29 of Unsupervised Learning. Ten years from now, she says, people will look at today's office the way we look at early-1990s filing cabinets and dictation. The near-term proof is already in a luxury call center.

Last summer Salesforce shipped its first generative apps under the Einstein GPT brand: Service GPT for support, Sales GPT for sales, and Commerce GPT for commerce. The two features that blew Shih away were service reply recommendations and case summaries—the tasks no agent enjoys and no customer wants to wait for. Grounded with basic RAG on data already in Salesforce, they worked out of the box. Customers called it magical.

Gucci was one of the first Service GPT customers, with agents split between Florence, Italy, and New Jersey. Average handle time fell. Then something else happened. After a customer came in with a belt-buckle that would not work, the AI noticed a handbag browsing session or a cart sitting for several weeks. The representative, coached in real time, stayed on and sold. Conversion rose by double digits. A cost center started to look like a revenue center.

Salesforce had assumed AI would stay in its lane: service for service, sales for sales. Instead the lines blurred. Agents became salespeople and brand storytellers, and the job description of a service representative changed. People still ask whether AI will take jobs. The Gucci agents Shih meets say they are doing the best work of their careers.

The week of the taping, Salesforce generally available'd Einstein Copilot, a natural-language assistant grounded in a customer's Salesforce data, metadata, and flows, plus Copilot Studio. Studio is a Prompt Builder for golden prompts that reference specific Salesforce fields, and a Model Builder for customers who want to fine-tune or bring their own predictive models into the stack.

Team structure kept evolving. AI people started decentralized in each cloud; Shih was running Service Cloud when that cloud built its own AI. Generative work exposed shared needs—the Einstein Trust Layer, a model gateway, multiple providers—that every app would require. Salesforce splintered Service AI and Sales AI into a shared platform team for the trust layer, the gateway, Prompt Builder, and Einstein Copilot. Product clouds kept predictive AI for their own use cases and now build actions: sales actions, service actions, Slack actions.

Trust is engineered on three levels. Technically, the Einstein Trust Layer does data masking, grounding with Data Cloud to cut hallucinations, citations, an audit trail, prompt defense, and zero-retention prompts. Policy requires customer-facing bots to self-identify as AI, even though that is not a technical requirement. The third layer is stakeholder work: an open-sourced set of trusted-AI principles around accuracy, honesty, and empowerment, shared with the industry and with regulators. Data masking itself predates generative AI, because names, gender, and zip codes can introduce bias; Salesforce already flags those sensitive fields and recommends masking them.

Copilots also force a shift from deterministic software to stochastic systems, which means a different way of building and a lot of user education. Shih's team sits in call centers, listens to the fear that AI will replace the job, and writes that concern into onboarding copy and the product. Even the name is a tell: it is Copilot, not Autopilot—a coworker you talk to in natural language, not a system that takes the wheel.

Salesforce AI Research has existed for a decade, staffed largely with former Stanford CS professors. They have worked on large language models since 2018, shortly after the Transformer paper, and published on prompt engineering when talking to a model in natural language still sounded radical. The summer before ChatGPT, the team was already playing with GPT-2 and 2.5. Quality jumped exponentially. It caught Shih's attention.

Gucci visited chief scientist Silvio Savarese and Shih in November 2021—her first in-person meeting after vaccination, everyone in double masks. Savarese demoed their codegen model. Jaws dropped, Shih's included. Gucci wanted to be the guinea pig and co-create. A year and a half later, when every enterprise wanted an LLM, Salesforce shipped Service GPT in a matter of 60 days. How did you do it in 60 days, people asked. The truth, Shih says, is that they did not.

That success produced a new org. Shih and Savarese now jointly run a Frontier AI group that prototypes off the Salesforce stack, as if it were a startup, on 30-, 60-, and 90-day experiments. What works graduates into core product. What fails still teaches the platform team which capabilities the core stack must support.

Everyone recites that whoever has the best data will have the best AI. Shih's first correction is ownership: it is not Salesforce's data. It is the customer's, and unlike many consumer companies, customer data is not the product. The second correction is that Salesforce actually holds four kinds of data that matter for models.

First is the structured CRM record, the company's heritage. Second is unstructured data pushed over the last two or three years: knowledge articles in Service Cloud, and an exponentially growing pile of Slack channels, contact-center voice, chats, and email. Data Cloud is adding vector search and hybrid reranking across both, plus zero-ETL partnerships with BigQuery, Redshift, Snowflake, and Databricks so enterprise lakes can come in.

Third is the metadata layer created 25 years ago to support multi-tenancy. Nobody guessed it would tell an AI which objects, tables, and functions to use. Fourth is feedback data. Salesforce is not only the world's largest database of customer data, Shih says; it is the largest database of customer outcomes—whether an opportunity advanced or died, whether a campaign earned an open, a click, a conversion. That stream is a reward function for predictive and generative models alike, captured as a pipeline in Data Cloud.

On models, Shih will not call a winner. Leaderboards change in weeks. The architecture is open: customers pick models on the service, bring third-party models, or bring first-party models through Model Builder. Over time, different tasks will want different models. Salesforce's own research models are being fine-tuned for code generation, Salesforce Flow generation, financial-services sales, and high-tech sales. Existing large Google Cloud customers who already live in Workspace and BigQuery tend to prefer Gemini because the purchase order is open. For the hardest work—agentic planning, plan generation—GPT-4 is still the best in class they are seeing.

Evals are vertical, not generic. The company is building sales-specific benchmarks for pharma and for wealth management. Industry experts define what a good response or a good outbound looks like; that becomes the eval set a new model is run through. It is not finished, Shih says, and it is complicated: the base model, the fine-tune, the data, and the entire RAG pipeline all move the score, so they want to benchmark every step.

Inference cost is passed through in product SKUs. Pilots are not cost-sensitive; scaling to a full contact center and multiple use cases is. That is why Salesforce is routing tasks to the right-size model, and why customers still have to write a business case. Gucci's math is the template: lower average time to resolve, plus incremental revenue from the extra inference. On incumbents versus startups, Shih sees room for both, as in prior software waves. Brett Taylor, a former classmate and former Salesforce co-CEO, is someone she is a fan of. Salesforce Ventures is writing checks and hoping startups build on Data Cloud and the UI.

Asked to rank what blocks a company from going from a trial to production, Shih does not put cost first. The biggest barrier is trust. Data-security and privacy risks are real, and the problem is not only leakage outside the firm. Different employees and departments already see different workflows and records. The AI has to honor those sharing rules. Two coworkers asking Einstein Copilot for the quarter's forecast should get different answers, because they own different opportunities. On the surface Copilot looks like ChatGPT or a former Bard; the difference is the data and metadata you do not see.

Hosts floated Intercom's advice to turn AI on for nights and weekends. Salesforce's baby step is turnkey use cases from last summer. Asking a customer to write prompts from scratch is a non-starter. Service GPT, Sales GPT, and Commerce GPT take about five minutes to stand up, and people immediately see the business value. Then they ask how to make it theirs. Gucci's brand voice is not AAA Insurance's, which is not Ford's. Gucci wants a custom style field and shopper frequency; Ford wants make, model, and how many Fords the family owns. You can fine-tune, Shih says, but it is much easier to put that in the prompt. Prompt Builder and Copilot exist for that.

Support is not only question answering. Copilots can orchestrate workflows: initiate an RMA, email a shipping label, run whatever the customer has already built in Salesforce. Air Canada's chatbot, which issued refunds it should not have, is the cautionary tale. Copilot does not automatically inherit every flow. An admin picks, in Einstein Copilot Builder, which existing flows and Apex the assistant may call, and can create new ones. Start with the least harmful actions—lookups—not issuing refunds.

If she were not at Salesforce, Shih has about 20 ideas. She founded Hearsay Systems in 2009 and spent a decade as CEO in financial-advisor AI: the "hear" half mined social posts and texts for lead events; the "say" half predicted what to say to whom, and when. Today's hearing side is far easier. The old NLP was regular expressions. Advisors now mine millions of texts and social messages for insights at the client, advisor, and firm-region level.

The next enterprise use cases after chat-your-docs, support, and code generation will not be a single product, she says. Every department should rewrite its job description, starting with functional leads and first-line managers, the way 1990s offices had to learn email and Google. Internally, Salesforce already brainstorms V2MOM—vision, values, methods, obstacles, measures, their OKR—on a secure generative model in Slack. The Copilot Studio launch plan was hashed out the same way: many people on one channel, treating the model as another teammate.

Most people still use AI in single-player mode. Chase and Effron argue the real shape of work is team plus AI: tagging a model in a channel so everyone sees "this might raise win rate by X percent." Shih agrees Slack is the place. Slack AI already does conversation and channel summaries. Next is Einstein Copilot in Slack, so a seller can pull open tickets, web engagement, and event history before a meeting and share it with the whole account team. Asked for a percentage of support questions AI can answer, she refuses a blanket number; it depends on the industry, and orchestration matters as much as answers.

Quick-fire: everything in AI is hyped, and we have not seen anything yet. It is 1998, except faster. What sits across the horizon is not just a technology shift but a civilizational one, and the job is to make the benefits accrue to everybody. Marc Benioff, often typed as a sales-and-marketing icon, also understands engineers, code, and AI at a fundamental level. Shipping this fast, Shih says, has taken a village. After she left, the hosts lingered on Slack as a multiplayer UI, on always-be-launching as a Salesforce-ism, and on two Redpoint bets: Abridge, which turns doctor-patient audio into EHR notes and kills "pajama time," and Ideogram, a text-to-image lab that is unusually good at rendering words in pictures.