Bret Taylor (Sierra, OpenAI): Why the future of enterprise software belongs to vertical agents

Sierra founder Bret Taylor on AI agents, trust, and why the next era of enterprise software will belong to vertical specialists

Every new tech era involves a defining shift that determines which companies come out on top. Just as the internet created space for Amazon, Google, and Salesforce, the rise of AI agents is surfacing a new generation of products, teams, and business models.

In a recent conversation with Pigment Co-CEO Eléonore Crespo, Bret Taylor shared what it means to build in the middle of that wave. 

Bret’s perspective is broad. He founded Google Maps, served as CTO of Facebook during its hyper-scale years, and later became Co-CEO of Salesforce, where he helped steer the company through a major cloud transformation. Today, he serves as chairman of the board at OpenAI in addition to leading Sierra – an agentic AI solution built to automate and improve customer service.

Why the wave matters more than any one company

After two decades of building products at the center of major technological shifts, Brett has learned a humbling truth: the wave is always bigger than you are.

He points to the internet as the clearest example. Companies like Amazon, Google, and Salesforce may appear very different on the surface, but their origin stories share a similar factor: timing. Each was born into a major technology moment that changed consumer behavior and opened up net-new markets.

Those companies did not force the internet into existence. They recognized that the web was rewriting how people discovered information, bought products, and ran their businesses. And they built into that momentum.

Bret sees AI agents and large language models (LLMs) the same way. Sierra is not creating that wave. It’s riding it.

Instead of leading the company as the sole author of its destiny, Bret continuously asks what the externalities look like. How are customer habits shifting because of AI? What new workflows are suddenly possible? Where are new markets appearing that didn’t exist three years ago?

For leaders, the takeaway is to zoom out. Roadmaps and feature lists matter, but they sit on top of a much larger movement. The more clearly you understand that movement, the easier it is to build products and teams that stay relevant as the wave grows.

Turning customer love into a habit, not a heroic act

Bret has a reputation for being close to his customers. Many of Sierra’s early adopters have his phone number. He and his team bring that mindset to events as well, inviting people like Will Guidara, the restaurateur behind the idea of “unreasonable hospitality,” to explore what exceptional service looks like in practice.

What stands out is how operational he is about it. For Bret, great customer experiences cannot rely on the heroic action of a few teammates. They have to be designed.

Agents as the new foundation of customer experience

Before the internet, almost every interaction with a company happened in person or over the phone. Today, most of that work runs through digital channels on a mobile device.

AI agents push that transition further. Phone support is the last analog channel. Once agents can answer calls, the same intelligent system can meet customers on the website, over messaging, and through voice. One team can design a unified experience that spans the entire surface area of the brand.

That experience is not just more consistent. It can be multilingual, always-on, and more personalized than the scripts many call centers run on today. Agents can combine policy knowledge with an understanding of emotion. They can learn to notice when someone is upset and decide to go above and beyond in the moment, the way a great human representative would.

The economics change even more dramatically. For many consumer brands, a human phone call can cost $20 to $25 dollars just in labor. In many cases, that’s more than the revenue the company earns from the average customer, which makes live service nearly impossible to offer at scale.

When AI agents cut that cost by orders of magnitude, entirely new experiences become achievable.

Bret compares it to the web in the late nineties, when many sites were simple translations of physical experiences. The next four or five years will be about discovering the equivalent of WhatsApp and Uber for agents, the products that feel native to this medium rather than borrowed from the last one.

Designing for trust in systems that are powerful but imperfect

Trust is the hardest question in advancing AI, and Bret is direct about that.

Traditional software felt like a very fast rules engine (if this, then that). It ran cheaply and deterministically. Teams could tune it for specific reliability targets and know that the same input would always produce the same output. 

LLMs behave differently. They are slower and more expensive per call than a typical page view or database query. They are also non-deterministic in that the same input can produce wildly different outputs. On top of that, LLMs can be both super intelligent and shockingly wrong in trivial ways.

Layer those traits on top of high-stakes workflows, and trust becomes a design problem, not just an engineering problem. Bret’s approach rests on a few core principles:

• Don’t overpromise. Agents should clearly identify themselves as AI and acknowledge that they sometimes make mistakes. Many consumers already understand this from their own use of tools like ChatGPT and can calibrate their expectations accordingly.
• Wrap models in guardrails. Supervisor models can watch underlying models. Deterministic checks can sit around AI calls. And clear policies can prevent a model from, for example, quietly rewriting a return policy in pursuit of better satisfaction scores.
• Treat oversight as a continuous loop. Humans will always be imperfect, and AI will be imperfect in different ways. The answer is not to wait for perfection. The answer is to design controls and audits so that, when mistakes happen, you notice them quickly, learn from them, and adjust systems without losing customer trust.

Bret’s advice to leaders is to stop asking when AI will be flawless and start asking how to make its imperfections visible and manageable. Trust comes from honesty about what systems can and cannot do, plus a clear plan for what happens when they’re wrong.

Why vertical agents will define the next era of enterprise software

When Bret looks at the future of enterprise software, he frames it in terms of atomic units.

Most of those agents will be highly specialized. One will answer the phone. Another will qualify leads. Another will audit financials or review supplier contracts. 

The pattern looks a lot like hiring specialists inside a company, with the main reason for this being domain expertise. A consumer packaged goods company, for example, might dream of an agent that optimizes its holiday supply chain. To work well, that agent needs to understand how the company structures contracts, how shipping timelines behave, how different markets behave seasonally, and more. A public company that wants an agent to support its financial close needs something that understands accounting rules and reporting practices, not just database fields.

Bret highlights Harvey as an early example. Harvey is a legal agent that helps with antitrust reviews – a software category that didn’t really exist before AI. It exists now because the workflows that lawyers already run are both highly valuable and highly structured, making them a good fit for careful automation.

This is where Bret sees a huge opportunity. Pair people who deeply understand obscure back office processes with AI engineers, and entirely new categories of software appear. For Pigment’s own customers, this maps closely to planning workflows where context, governance, and cross-team collaboration matter as much as the model itself.

AI as a mainstream building block, not a niche capability

One of the most encouraging shifts Bret sees is that AI has moved from a niche machine learning discipline into everyday engineering.

Until recently, building AI capabilities often required an in-house machine learning team, data labeling, and expertise in training and reinforcement learning. That limited where AI could realistically be applied.

Today, full stack developers can reach for models the way they reach for databases. They can use off-the-shelf APIs and tools to add intelligence to applications without deep research backgrounds. That lowers the barrier to experimentation and creates space for long-tail applications that would never have justified a dedicated machine learning team.

There are still open questions. Many powerful techniques in reinforcement learning remain opaque, which can be problematic for business processes that need strong guarantees. If an agent is told to improve customer satisfaction and quietly does it by weakening the returns policy, the result might be better scores but an unacceptable business tradeoff.

Bret is interested in product patterns that pull those methods into environments where humans can set boundaries and understand tradeoffs. The goal is to get the most out of advanced training techniques while keeping humans firmly in the loop.

AI, mental health, and the cost of getting it wrong

When asked about the philosophical questions that keep him up at night, Bret points to mental health.

Still, Bret believes AI could be a powerful force for good.

As a parent, Bret considers tools like ChatGPT to be more like a search engine than social media. In his own home, social platforms are off limits for his children, but AI tools are not. He wants them to use AI as a way to amplify their agency, not to avoid doing the work. At home, using AI to write an essay is out of bounds, but using it to critique their own writing and suggest improvements is encouraged. The goal is to teach his kids how to use these tools without letting the tools quietly do the learning on their behalf.

How leaders can keep up when the ground keeps shifting

Every tidal change affects the sand beneath it. To keep steady enough to take advantage of the wave, Bret’s first recommendation is to seek out smart people and listen deeply.

His next piece of advice is to keep your hands on the tools. It is easy to extrapolate wildly from one demo or one article. The only way to build realistic intuition is to tinker.

For designers, Bret believes personal agents will eventually flip the idea of interfaces. Instead of people learning software, software may increasingly be learning people. That will put more emphasis on the core job a product does for the customer and less on surface-level choices like button size or hero images. Design will still matter, but more at the level of experience, trust, and value than at the level of individual screens.

Conclusion

Bret considers the AI agent wave larger than any one company riding it. Leaders who thrive will be the ones who stay humble about that fact while still building ambitious products and systems.

That means:

• Treating customer love as a machine, not a miracle
• Designing trust into every layer of an agent rather than assuming perfection
• Betting on vertical domain expertise where workflows are rich and under-served

Perhaps above all, it involves learning through people and practice. The wave may be bigger than any company, but the leaders who stay curious, grounded, and hands-on will be the ones who ride it best.