A founder sits through a vendor demo where an AI agent handles a support conversation flawlessly end to end, no handoffs, no hesitation, and signs a contract for full autonomous coverage across the support queue. Three months later the agent is quietly routing anything remotely ambiguous to a human, the 'full autonomy' pitch has become a narrow set of well-defined intents, and the founder is wondering whether they were sold a demo or a product. They weren't alone. At Dreamforce 2026 in September, two of the event's own featured enterprise customers stood up and said the same thing in public, to the company that just spent the conference selling agentic AI as the future of every job function.
What did AT&T and Crocs actually say about agentic AI at Dreamforce 2026?
Both told delegates that automation is delivering real value only in narrow, well-scoped use cases, and that the 'autonomous agent doing your job' narrative has run into edge cases, escalations and brand risk. John Miller, AT&T's VP of Consumer and Business Solutions, described using AI to cut phone-upgrade time from roughly 30 minutes to about 10, a genuine, measurable win. But AT&T deliberately keeps AI agents out of customer cancellations, reserving that specific interaction for humans. Miller's reasoning: the company wants to 'understand from our customers' perspective why they're cancelling, and then be able to help,' a judgment call an agent optimized for call deflection has no way to make on its own.
Crocs told a similar story from the retail side. Feliz Papich, Crocs' SVP of Digital Technology and Experience, runs a consumer-facing chat and voice bot called Rivet alongside internal agents for promotion creation and product categorization. Her framing was direct: 'there still needs to be a human in the loop as we're still learning about capabilities and complexities of automation.' She also flagged something vendors rarely lead with in a sales pitch: the data quality underlying an AI interface gets far less attention than the interface's polish, even though it's the part that actually determines whether the agent is trustworthy.
Why does this matter more coming from AT&T and Crocs than from a random critique of AI hype?
Because these are Salesforce's own featured production users, at Salesforce's own flagship event, publicly drawing a line the marketing hadn't drawn. Dreamforce 2026 went all-in on agentic AI as the headline theme, yet the customers actually running agents in production at scale used their stage time to say the opposite of 'agents can do the job end to end.' That's not a competitor's talking point or an analyst's caution, it's two well-resourced enterprises with real deployment experience describing what actually held up.
It's also consistent with the wider 2026 data on enterprise AI agent deployments. 88% of enterprise AI agent pilots never make it to production, per the 2026 State of AI Agents report, and the research consensus is that the gap is overwhelmingly a governance and scoping problem, not a model-capability problem. AT&T and Crocs are describing, from the inside, exactly what that scoping discipline looks like once you've actually shipped: keep the agent inside a boundary you can defend, and put a human at the exact point where judgment, not pattern-matching, is required.
How do you tell a legitimately scoped AI automation project from an overpromised one before you sign?
A handful of direct questions do most of the filtering:
- **Ask exactly where the human sits, not whether one exists.** Every credible vendor will say 'human in the loop' somewhere in the pitch. The real question is which specific decisions route to a person and why, the way AT&T names cancellations specifically rather than gesturing at 'complex cases.'
- **Ask what happens on the 5% of inputs the demo never showed you.** A demo is tested on curated examples by definition; ask what the agent does with a request it's never seen shaped like that before, and whether that's a designed fallback or an untested gap.
- **Ask what data the agent's judgment actually depends on**, since Papich's point about data quality mattering more than interface polish is the one most sales conversations skip entirely. An agent making decisions on stale, incomplete or ungoverned data is a liability wearing a good UI.
- **Ask for the narrow version of the pitch, not the expansive one.** If a vendor can't describe the smallest, most defensible version of the automation, one clear workflow, one clear boundary, they likely haven't scoped it themselves yet, they're scoping it live in your contract.
- **Treat 'full autonomy' claims for anything customer-facing and judgment-heavy as a red flag by default**, not because autonomy is impossible, but because the two companies with the most production experience in the room just said, in public, that they deliberately don't build it that way yet.
Does this mean AI automation isn't worth investing in?
No, it means the return comes from scoping correctly, not from chasing full autonomy. AT&T's own example, phone-upgrade time cut from 30 minutes to roughly 10, is a real, compounding win precisely because it targeted a specific, high-volume, low-judgment task rather than the entire customer relationship. That's the same pattern behind every AI automation project that actually shows measurable ROI within weeks rather than stalling for quarters: pick the narrow, well-defined bottleneck first, prove it, then expand the boundary deliberately instead of promising the whole boundary on day one.
How AIBOOTSTRAPPER helps
This is the exact discipline we scope every automation and agent build around. Building Leon & Vera for local service studios meant deliberately keeping two tightly scoped agents rather than one sprawling platform: Leon produces a week of ad creative from the owner's own photos, Vera handles every inbound enquiry and books it straight into the existing calendar, and the owner controls ad spend from as little as €10 a day with a hard, owner-set ceiling rather than the agent's own judgment deciding what to spend. Neither agent pretends to do the other's job, and that's exactly why both work reliably in production instead of joining the pile of pilots that never scale past a demo.
If you're evaluating an AI automation or agent proposal and want a second, technical opinion on whether it's scoped like AT&T's phone-upgrade flow or like the overpromised version Dreamforce's own customers pushed back on, book a call or see our AI automation and product development services.
Want this done for you?
Book a free strategy call and we'll show you how to build and market your business with AI.