Seven People Did the Work of 700. Salesforce's AI Chief Says Enterprise AI Is Still in the Second Inning

Sarah Guo introduces Clara Shih as CEO of Salesforce AI, a job that covers the copilot and agent platform, model development, go-to-market, adoption, partnerships, ecosystems, and secure, responsible AI. Guo jokes that she got tired just reading the list. Before that, Shih ran Salesforce Service Cloud, co-founded and led Hearsay Social, and sat on the Starbucks board. The dual resume — scaled operator and founder — is why Guo and Elad Gil wanted her view of enterprise generative AI in late 2023, when every company was asking how to actually use the stuff.

Shih's AI work predates the current wave. At Hearsay, the company already used NLP to mine financial advisors' social messages for lead-generation opportunities and compliance infractions, all before large language models. When she joined Service Cloud almost three years earlier, customer service was already thick with chatbots running early, pre-GPT transformer models. As those models and the OpenAI ecosystem improved, it became obvious this would be core to the product. About a year and a half before the recording, she and engineering leader Jayesh doubled down on experiments and prototypes, including with Gucci, that became the early shape of Service GPT. They were learning and iterating as they went.

Then ChatGPT launched, and every customer was suddenly interested. Across Salesforce there was a wakeup: how do you apply large language models to every cloud? Sales, marketing, commerce, and Slack could each have built their own stack. Shih's group argued for a common platform and shared services instead — model fine-tuning, Prompt Builder, a trust layer, a gateway — so the company and its ecosystem could all go fast. That argument was formalized into the Salesforce AI CEO role she had held for about six or seven months when she sat down with No Priors.

Salesforce has been in the AI story since Einstein, now evolved into Einstein Copilot and Einstein GPT. The company has built large language models for four or five years and open-sourced many of them on Hugging Face, another Salesforce Ventures investment. Shih is not trying to freeze a single-model strategy. Large enterprises arrive with their own models or want to fine-tune. SMBs want Salesforce to figure everything out. A third group wants Salesforce to pick the right model for the right task from data, feedback, cost, performance, and latency. A software company serving that mix cannot write strategy as one model assumption.

The supply side is therefore three layers. In-house models include CodeGen from the research team, which powers Apex code generation in Developer GPT, plus fine-tuned domain versions for customer service, sales, healthcare, and financial services. Customers can spin up and fine-tune their own models on data in Salesforce Data Cloud. Third-party options include Anthropic and Cohere, both Salesforce Ventures investments, OpenAI as a close partner, and Google Vertex — buy through Salesforce or bring your own API keys.

In March 2023 the company announced out-of-the-box AI features in every existing Salesforce cloud. Service GPT, Sales GPT, and Marketing GPT are, in Shih's telling, prompt templates product managers wrote against operational bottlenecks. The most popular in market is service reply recommendations: a customer emails or chats, and the system drafts a response grounded in knowledge articles and similar past cases. It is generally available, with customers using it and sending feedback.

In parallel, the platform team is building Copilot as a natural-language interface across the clouds and Slack, plus Copilot Studio in three parts. Prompt Builder, piloted the week before the interview, lets customers take official templates and change brand tone, model, and grounding — including custom fields that exist only in their org. Action Builder gives the copilot agent powers: decades of Salesforce workflows, sharing rules, permissions, and MuleSoft integrations, with their SLAs and security guardrails, can be designated as actions in one click. Einstein Studio is bring-your-own-model, training or fine-tuning predictive and generative models on data Salesforce holds or indexes.

Gil wants the unglamorous number: how much generative AI is actually in production versus experiments and pilots? Salesforce has so many customers that the answer is all of the above. A handful have rolled Service GPT or Sales GPT across the contact center, already changing reps' days. Shih has talked to agents who feel they are doing the best work of their careers because the manual lookup and rote tasks that used to bog them down — and make customers angry — are largely automated or accelerated. Most customers sit in the middle, still experimenting, suddenly aware that they need their data ducks in a row, connecting Salesforce Data Cloud to the multiple data lakes every Fortune 500 company seems to own.

Asked whether large-scale enterprise adoption is one, two, or three years away — a question that matters to eval and observability startups whose futures ride on that ramp — Shih refuses a single date. There is a distribution. Aggregated, it is early: probably the second or third inning. She is not a baseball expert, she adds, but that is roughly the location. Enough companies, though few and far between, have proved you can transform business processes in a big way. Most enterprises still have data all over the place. That is step one. Data Cloud has become the fastest organically developed product in Salesforce history, driven by the need for data to train, fine-tune, and retrieve. Zero-ETL sharing integrations with BigQuery, Databricks, and Snowflake are meant to pull structured and unstructured data into one place.

Gil's sequence — data, then internal or external prototypes, then push to end products — is largely right, Shih says, with a smaller bite available. Because Salesforce is a CRM company, the most common entry is customer service. You do not have to clean all enterprise data first. Bring knowledge articles across silos together with Data Cloud connectors, vector search, and embeddings, and you can drive decent RAG for questions that hit the self-service agent, formerly Einstein Bots, or a human. That is a production-shaped wedge while the rest of the lake is still a mess.

Guo asks how Shih educated product managers and engineers across such a wide suite. It has not been one-way. Curiosity is high; much of the roadmap is coming from people across the organization. As agents mature, she thinks software development itself will change. Branching logic, execution paths, and painstakingly specced screens may get handed to an agent to do dynamically. Guo then asks the better consumer question: efficiency in the contact center is easy to picture, but what should an ordinary end user expect to get better?

Shih's line is the episode's thesis: AI is the new UI, or maybe Slack is the new UI for AI. Slack, Tableau, and MuleSoft were not acquired in the name of AI. In hindsight they look perfectly placed — a conversational interface, data visualization, and every plugin an enterprise might want. On product UX, Salesforce is testing a raw prototype called Generative Canvas. As you talk to Einstein Copilot about a sales opportunity, components pop up: Tableau and Salesforce report visualizations, record updates, the pieces a Lightning web component page used to hard-wire. The components already exist; the model has to call the right ones. It is messy and not ready for prime time. They are showing it to customers, using it themselves, and expecting to roll out something radically different. Slack work is also exploring agents that serve teams, not just one person.

The day-to-day change is already visible in service. Gucci hired a wave of new representatives during Covid, with high early turnover. Luxury service is hard: customers spend a lot and expect the advisor to know the product. Retrieval augmentation now arms every advisor, even a new hire, with expert brand storytelling and troubleshooting. Average handle time on support issues has gone down, but the call does not get colder. Because Salesforce holds a 360-degree customer view, the advisor can see that the broken purse buckle is fixed and that a belt and shoes sit in the cart, or that the customer has been browsing the site. The same person then tells you about the heritage of the shoes, and that Jackie Kennedy used to wear them. Service, sales, and marketing collapse into one question: what does the customer want, and how do you give a new hire all the knowledge required to meet it.

Guo asks what enterprises have learned about sensitive data, ownership, and quality. Shih says they do not have all the answers, but one distinction is already load-bearing. Unstructured data is not one thing. A PRD or a knowledge article was written to communicate; you can probably assume every sentence matters. Call transcripts, chat logs, and Slack channels are the opposite: most of it was not intended for other people — back-and-forth, clarification, someone talking about their dog. That second kind cannot go straight into retrieval. It has to be preprocessed and mined, a capability Salesforce is building into the platform. Vectorizing is not enough, either. A retail call may reveal that a customer's favorite color is blue and that she has a teenage daughter she also shops for. Those facts should land in structured Salesforce fields.

Asked for unexpected emergence, Shih says everything — the fact that it works so well, a surprise every week. Two to four years out, she analogizes to the early cloud. Some applications clearly belonged on demand and on a phone; some still want to stay on-prem. AI will split the same way. Workflows that must be fully deterministic and reproduce the same way every time — authentication, a financial transaction, a healthcare procedure — will stay as they are. Everywhere else, the job of the engineer, product manager, and designer shifts from prescribing the how to describing the why, the what, and the goal, leaving the AI to figure out the how stochastically.

Pricing is the hard business problem because AI has real compute COGS. Salesforce has put prices in market and talks about the tension constantly: cover costs, but do not make customers run complicated token calculators. The requirement that actually matters is ROI. For Gucci and other retailers, is average handle time down, and is sales conversion up versus baseline? If the value is there, customers will pay. It cannot be an added cost without a clear benefit. Gil is optimistic for two reasons: many applications are net-new capabilities or order-of-magnitude productivity gains, and the cost of doing the same task with models should fall as the stack improves. Shih's extreme sample is a meeting the day before with Runway ML founder Cristóbal Valenzuela. Runway powered much of the effects in *Everything Everywhere All at Once* with a video-editing team of seven; a movie that size would traditionally have needed 700. Jobs and the Hollywood strikes sit behind that number. From an ROI standpoint, she is sure the studio and Runway can find a model that works.

Guo's last question is the founder one: with Salesforce's product, channel, and data advantages in the room, where should startups push? Shih lists four layers. Foundation-model companies, if you can still call them startups given their size. Domain-specific work in legal or medicine. Tooling, because different organizations have different needs. And applications, which will be built very differently than they are today. Her concrete advice for the application layer is to align with existing data graphs, because usefulness needs an anchor in data. We do not know what we do not know, she says. It is a time to invent and test. Salesforce is doing a lot. It cannot do everything.