The Trust Threshold: Why Google DeepMind’s Tulsee Doshi Says AI’s Next Chapter Hinges on User Confidence

When Google DeepMind’s product VP, Tulsee Doshi, sits down to talk about the future of artificial intelligence, she doesn’t lead with benchmarks or model parameter counts—she leads with trust. And that’s a signal every B2B leader should take seriously.

At Google’s I/O developer conference this week, the company rolled out a firehose of new AI capabilities: personal agents that act on your behalf, code generators that accelerate development, search tools that rethink how we find information, and a “world model” designed to produce physically accurate video. All of it runs on the latest Gemini 3.5 models, developed inside DeepMind. But the real story isn’t the tech stack—it’s the delicate balancing act between capability and confidence.

For revenue teams, product leaders, and CROs building AI into their go-to-market motion, Doshi’s perspective offers a playbook for navigating the tension between shipping fast and earning trust.

The Safety-Quality Spectrum Every AI Product Must Navigate

Doshi frames the core challenge as a spectrum. On one end: the “blank response rate”—where an AI simply refuses to answer because it’s safer to say nothing. On the other: going too far, saying something nuanced but crossing a line into harmful or sycophantic territory.

“That’s always the spectrum that we’re trying to find the right balance on,” Doshi told Fast Company. She added a personal note: “I feel assured by an agent that chooses not to answer a question.”

For B2B buyers, this is a crucial distinction. In enterprise sales and customer success, the AI tools you embed into your workflows must know when to say “I don’t know” or “I can’t answer that.” A model that confidently hallucinates a customer’s contract terms or misrepresents your product’s capabilities isn’t just embarrassing—it’s a liability.

The lesson for product and GTM teams: ship models that are calibrated for confidence, not just speed. If your AI assistant is generating outbound email sequences or summarizing CRM data, make sure it has guardrails that prevent overpromising. The most valuable AI tool is the one your reps can trust to tell the truth—even when the truth is “I need human help.”

Beyond Traditional Harms: Sycophancy and Agent Safety

Doshi highlighted two evaluation criteria that deserve more attention from B2B practitioners:

• Sycophancy: The tendency of AI models to agree with the user or flatter them, even when the user is wrong. In a sales context, this could mean an AI rep agreeing with a prospect’s incorrect assumption about pricing or capabilities.
• Agent safety: As AI moves from answering questions to taking actions (scheduling meetings, drafting contracts, updating pipelines), the stakes go up. An agent that accidentally deletes a deal from your CRM because of a misinterpreted command is a nightmare scenario.

Both of these are now front and center at DeepMind. They’re building verifications into the product experience to catch these failure modes before they reach users. For your team, this means you should be testing not just “does the AI answer fast?” but “does the AI contradict itself when I push back?” and “can the AI recover from a mistake gracefully?”

The Agentic Persona Shift: What Happens When AI Acts With You, Not Just For You

One of the most provocative points Doshi raised was about persona. When Google’s Gemini moves from being a conversational assistant to an agent that acts on your behalf, the entire relationship changes.

“As we enter this more agentic era of Gemini acting with and for you, there’s a switch in persona that you also need to think through,” she said.

In plain language: an AI that writes an email for you is one thing. An AI that sends that email, monitors the reply, and schedules a follow-up meeting without your direct input is something else entirely. The trust required from the user jumps dramatically.

For B2B revenue teams, this shift is already happening. Platforms like Gong, SalesLoft, and Outreach are embedding agentic features into their workflows. Imagine an AI that not only suggests a next step but actually executes it: books a meeting in your calendar, drafts a proposal, and sends it to the prospect—all while you’re in another conversation.

That’s powerful. It’s also terrifying if trust isn’t baked into the design.

How to Build Persona-Worthy Trust in Your AI Products

Doshi emphasized that the persona is “an area we are actively investing in” and that it will evolve based on user feedback. This is a critical insight for any B2B company building AI features: the persona isn’t set at launch. It’s refined over time through listening to what users actually need.

Here’s a practical framework for applying this to your own GTM:

1. Start with transparency. Tell users explicitly when an agent is acting on their behalf, and give them a way to review or undo actions. Think of it like two-factor authentication for AI decisions.

2. Design for recovery. Every AI will make mistakes. The question is whether your system helps users fix them easily. Doshi noted that DeepMind builds “the right verifications” into the product experience. Do the same for your sales or customer success tool.

3. Let users choose the persona. Some reps want an AI that’s assertive and direct. Others want one that’s cautious and deferential. Giving users control over the tone and assertiveness of your AI can dramatically increase adoption.

4. Measure trust, not just usage. Standard metrics like DAU or message volume don’t tell you if users trust the AI. Instead, track override rates: how often do users reject or modify the AI’s suggestions? A low override rate is a sign of trust. A high one means you have work to do.

The Bigger Picture: Why Small Business AI Adoption Numbers Matter

The newsletter also touched on new data about small businesses adopting AI. While the source didn’t provide exact figures, the trend is clear: AI is moving from a “nice to have” for early adopters to a “must have” for mainstream businesses.

For B2B SaaS and tech companies, this means your buyers are becoming more sophisticated. They’ve probably already tried ChatGPT, Claude, or Gemini in their personal lives. They come to your product expecting AI to be embedded, intuitive, and—most importantly—trustworthy.

If you’re still shipping AI features without robust safety evaluations, without clear guardrails, and without persona thinking, you’re going to lose to competitors who take the Doshi approach.

What Revenue Teams Should Do Right Now

Based on the interview with Tulsee Doshi and the broader shift in AI product thinking, here are three actionable steps for B2B leaders:

1. Audit your AI’s sycophancy rate

Run a simple test: give your AI tool wrong information about your product, pricing, or competitor landscape. Does it push back, or does it agree with you? If it’s sycophantic, you need to add disclaimers or retrain the model.

2. Implement agent safety reviews before every launch

Before you ship any agentic feature (auto-scribe, auto-summarize, auto-send), run a failure mode analysis. What’s the worst that could happen if the AI misinterprets a command? Build a manual verification step into the flow until you’re confident.

3. Start a “trust score” dashboard

Track how often users manually correct AI outputs. Track how many users turn off agentic features after trying them. Use this data to prioritize improvements—not just on performance, but on confidence.

The Bottom Line

Tulsee Doshi’s remarks aren’t just a window into Google DeepMind’s internal decision-making. They’re a blueprint for every B2B company that wants to win in the AI era. The winners won’t be the ones with the fastest models or the most features. They’ll be the ones that users trust.

As Doshi put it: “There’s always a trade-off between blank response rate and answering in a nuanced way.” The key is finding the right balance for your users—and being transparent about where that balance lies.

If you’re building AI into your sales, marketing, or customer success stack, stop thinking about the tech first. Start thinking about the trust. Because in mid-2026 and beyond, that’s what separates the tools your team relies on from the ones they ignore.

Ready to build trust into your AI-powered GTM motion? Start with one small change: give your team a clear way to verify every AI action. Then measure the trust, not just the output.