--- title: What CROs should take from Unbound 2026 description: "What Unbound revealed about AI agents, revenue outcomes, context, human oversight, and the evidence CROs need before trusting their decisions." canonical: "https://vasco.app/blog/cro-takeaways-unbound-2026" date: "2026-09-24T16:42:00.000Z" authors: - Justin Hudon jobTitle: Head of Sales and Customer Success readingTimeMinutes: 2 contentType: article intent: strategy-insights pillar: "AI in GTM & RevOps" audiences: - cros --- # What CROs should take from Unbound 2026 _AI agents: 3 questions every CRO should ask_ I've had a few days to sit with Unbound. The most useful part for me was the shift in the conversation from adoption to outcomes. There was plenty of discussion about what these systems can do. The more interesting question is whether they are helping revenue teams make better decisions (thereby growing revenue or reducing inefficiencies/costs). Underneath that sit two questions I think every CRO should ask about the agents they're buying: can we trust them, and how would we know? Three things stood out. ## 1. Context means something different to every vendor "Context" came up constantly at Unbound, but it didn't always mean the same thing. For a CRO, there are four layers worth separating: - Business context: What is the business reality? Who are we selling to? What are my ICP tiers? - Data context: Are we actually looking at the same customer across CRM, billing, product and every other system? Do we have one unified view of all my accounts? - Initiative context: What are we trying to achieve? What are we doing weekly to drive results? - Definition context: What do "qualified," "pipeline" and "closed" actually mean in this business? None of these lives in a single system. The CRM holds what happened. The plan sits in a spreadsheet, the definitions live in one operations person's head, and the same company shows up as a separate record in every tool it touches. An agent reading those systems raw has to guess at each layer, every time it's asked, and nothing in its answer tells you when the guess was wrong. ![Vasco's Justin Hudon visits HubSpot Unbound 2026](https://cdn.sanity.io/images/ys8gstp8/production/357e03013678ed649b78413660e814f981f35b19-768x1024.jpg?w=1600&fit=max&auto=format) We tested how much this matters by holding the model still and changing only what it stands on. Same Claude, same monthly business review, two foundations: 21.9% of answers correct on raw CRM and Stripe data, 99.5% on the Vasco context graph. The model wasn't the variable. An independent study in JAMIA Open (2024) found the same pattern across 60 business questions: 8.3% correct at the low end, 78.3% at the high end. That gap is what we built Vasco to close. Vasco is the GTM brain for AI agents: the revenue context layer that lets go-to-market agents run in production. The brain holds the plan, the definitions and the resolved facts once, so every agent reads the same answer instead of rebuilding it on each request. Having access to more data doesn't solve this on its own. Before buying another tool, ask which of these four layers it actually has, and whether it resolves them once or guesses at them every time. ## 2. Recommending and acting are different jobs That range is useful, but putting everything under the same label hides an important distinction. ![Vasco's Justin Hudon sees Tom Brady at HubSpot Unbound 2026](https://cdn.sanity.io/images/ys8gstp8/production/196035aeceb07db1abdd77161689cfea4bb25f91-768x1024.jpg?w=1600&fit=max&auto=format) If a system answers a question incorrectly, someone can correct it. If it changes a forecast, flags a customer for intervention or triggers a workflow, the stakes are higher. It helps to treat autonomy as a scale instead of a switch. We use four steps: advisor, approver, supervised, autonomous. An agent starts by recommending. It earns the right to act while its record stays clean, and it moves back down when people start editing, correcting or cancelling its work more often. This is what we mean by the harness: the layer that governs how agents are allowed to act. Every agent inside it owns an outcome, and every number it acts on opens to the accounts underneath it, so trust is something you can inspect rather than something you're promised. ### CROs should be clear about where that line sits in their own organization. What can an agent recommend? What can it change? What still needs a person to check it? Those decisions shouldn't be left to RevOps by default. ## 3. The real test is what changed This was the question I kept coming back to afte It's easy to report on seats, usage, prompts, credits or hours saved. Those numbers tell you whether people are using a system. They don't tell you whether it improved revenue performance. Adoption looks healthy almost everywhere. Outcomes don't. MIT NANDA looked at more than 300 publicly disclosed deployments and found that 89% of AI agents never reach production, and 95% of enterprise generative AI pilots returned nothing measurable. The harder question is what changed as a result. ![Vasco's Justin Hudon at the HubSpot Unbound 2026 exhibition room.](https://cdn.sanity.io/images/ys8gstp8/production/33ec3eeb63e2fc9ebca14cb30f9ba60a522520e2-768x1024.jpg?w=1600&fit=max&auto=format) Did the team catch a problem earlier? Did forecast accuracy improve? Did reps spend more time on the accounts actually worth pursuing? Did an intervention reduce churn? Those numbers are much harder to establish, but they're the ones a CRO can take seriously. Most tools stop short of them. They surface a signal, make a recommendation and never see what happened next, so the best they can offer is better advice. The agents worth paying for close the loop: they propose a play, the team runs it, and the agent watches the result and reports whether it moved the number. That record of what worked is what makes the next recommendation better, and it compounds for the team that keeps it. You rent the model. You own the memory. ## What I'm taking back For me, the useful questions coming out of Unbound are pretty simple: - What context is missing? - Where does the work still need human oversight? - What revenue outcome should change if this works? The answers will tell you more than another AI adoption metric. They're also a practical way into the two questions that matter for any agent: can we trust it, and how would we know?