Talking to the user who hates AI agents

A specific type of person comes up in every customer interview. The AI skeptic.

Not the casually doubtful — those convert quickly when they see specific value. The active skeptic. The person who thinks AI is overhyped, that current agents are flaky, that the industry is in a bubble.

We talk to many of them. Sometimes they convert. Sometimes they don't. Always they're interesting.

This post is about how we talk to AI skeptics. Not to convert them. To engage honestly with their concerns.

What AI skeptics actually believe

The skeptic position isn't uniform. Common variants:

Variant 1: "AI is overhyped, mostly demoware."

These people have watched cloud-LLM demos that don't translate to production reliability. They've tried tools that worked for the demo case but failed in their actual use. Their skepticism is based on direct experience.

Variant 2: "AI is harmful to workers."

These people worry about job displacement, particularly in their own profession. They've seen layoffs framed as AI-related. Their skepticism is values-driven.

Variant 3: "AI is environmental disaster."

These people focus on the energy consumption of large model training and inference. Their skepticism is about externalized costs.

Variant 4: "AI hallucinates too much to trust."

These people have specific examples of AI producing confident-sounding wrong outputs. Their skepticism is about reliability.

Variant 5: "AI is a bubble; this will all collapse."

These people see venture funding patterns, hype cycles, and historical analogies (dot-com, crypto). Their skepticism is about market dynamics.

Each variant deserves a different conversation. Lumping them together as "AI skeptics" misses the substance.

What we don't do with skeptics

→ We don't try to convince them they're wrong. Their skepticism is based on real observations. Dismissing it loses the conversation immediately.

→ We don't claim AI is universally good. It's not. Different deployment contexts have different values.

→ We don't deflect to "everyone else is doing it." Bad argument. Doesn't change the truth value of the skepticism.

→ We don't ignore the specific concern. Each variant of skepticism has a specific worry. Speaking to it specifically respects the person.

How we engage each variant

Variant 1: "Overhyped demoware."

Our response: "You're right that a lot of AI is demoware. We've been frustrated by it too. What we built is specifically for the operational use cases where reliability matters. The boring stuff. Meeting follow-ups, ticket triage, that kind of thing. Local model, repeatable workloads. Demo videos of this look unimpressive. Production deployments of this work."

Sometimes converts. Sometimes the skeptic says "I'll watch and see if you survive the bubble." That's fair.

Variant 2: "Harmful to workers."

Our response: "We share the concern about AI harming workers. The way we think about it: there's work humans love doing, and work humans don't love doing. Most operational drudgery is in the second category. Agents absorbing the drudgery, freeing humans for the relationship + creative work, isn't displacement — it's redistribution of time. We're explicit about not building agents that replace human judgment. We build agents that absorb drudgery."

Sometimes converts. Sometimes the skeptic doesn't trust that the redistribution will happen fairly. That's a legitimate concern.

Variant 3: "Environmental disaster."

Our response: "Training large models has significant environmental cost. Inference cost depends heavily on architecture. Local 7-13B models running on consumer hardware use dramatically less energy than frontier cloud LLM inference. Our local-first architecture has substantially lower environmental footprint per query than cloud-LLM alternatives. We're not zero-cost — running compute uses energy. But it's not the disaster cloud-AI hype suggests."

Sometimes converts. Sometimes the skeptic doesn't trust the energy math. We share our energy estimates if asked.

Variant 4: "Hallucinates too much."

Our response: "Hallucination is real. Cloud-LLM agents have produced confident wrong outputs that we've seen too. Our approach: agents have confidence scoring built in, and high-stakes actions default to human review. We don't claim agents are reliable enough to act autonomously on important decisions. We use the trust ladder approach — start with drafts only, expand authority gradually. For most operational workloads (classification, extraction, structured drafting), output quality is good enough at the configured confidence levels. We don't pretend it's perfect."

Sometimes converts. Sometimes the skeptic still doesn't trust. That's fine — they shouldn't if they don't.

Variant 5: "Bubble; will collapse."

Our response: "AI valuations may be inflated. Hype cycles are real. But the underlying technology produces real value in real workloads. When the hype cycle adjusts, the companies that built durable value will survive. Like the dot-com era — pets.com was a bubble; Amazon survived because they were building durable value. We're trying to be in the durable category. Our approach (local-first, flat pricing, focus on operational use cases) is structurally less hype-dependent than alternatives. We might still be wrong. But the bet is on durability."

Sometimes converts. Sometimes the skeptic says "I'll evaluate again after the correction." That's reasonable.

What we learn from skeptics

Skeptic conversations are more valuable than enthusiast conversations.

Enthusiasts validate. Skeptics challenge.

When a skeptic articulates a specific concern we don't have a clean answer for, that's a signal we need to think harder. Some of our positioning evolved because of skeptic conversations.

Specifically:

Skeptic concerns that improved our product:

→ Audit ledger as foundational (came from compliance-skeptic conversations) → Confidence scoring on outputs (came from hallucination-skeptic conversations) → Local-first architecture emphasis (came from privacy-skeptic conversations) → Trust ladder framework (came from autonomy-skeptic conversations)

Skeptic concerns that improved our marketing:

→ Less universal-value language (came from "overhype" skeptic conversations) → More specific use case framing (came from "I don't see where this fits me" skeptic conversations) → Explicit "what we're not" content (came from "you claim too much" skeptic conversations) → Cost transparency (came from "vendors hide pricing" skeptic conversations)

The product is better because skeptics challenged us.

What skeptics teach about timing

Some AI skeptics are right about timing, even when they're wrong about the long-term trajectory.

Specifically: "AI will be useful eventually but isn't ready yet" is a common 2024-2025 skeptic position. For specific use cases in 2024-2025, this position was correct. Cloud-LLM agents were too expensive and too unreliable for many operational uses.

By 2026, the calculus changed. Local AI capability improved. Cost economics shifted. Tooling matured.

The skeptic who said "not yet" in 2024 might still be right to be skeptical of certain implementations in 2026. The honest engagement is to acknowledge that the technology has evolved and ask the skeptic to evaluate the current state, not their cached 2024 view.

Sometimes they're willing to re-evaluate. Sometimes they're not. Both responses are fair.

What skeptics teach about humility

Engaging with skeptics regularly is a useful corrective to vendor groupthink.

It's easy in the AI agent vendor space to think every potential customer is just one demo away from converting. Skeptics remind you that's not true. Most people aren't going to buy. Most are appropriately uncertain about AI.

We've maintained a more grounded marketing voice partly because we talk to skeptics often. The conversations check our enthusiasm.

If you're a vendor in this space and you don't talk to skeptics regularly, your view of the market is probably skewed. Talk to them. They'll improve your product.

What skeptics teach about quality

Skeptics who convert tend to be the highest-quality customers afterwards.

Why: they evaluated carefully. They had specific concerns. When they bought, they bought because the product addressed their concerns.

Enthusiasts who convert tend to churn more — they bought on excitement, not analysis.

We have customers who started as active skeptics and are now multi-year users. They're often our most thoughtful customers. Their feedback is more useful than enthusiast feedback.

How to engage skeptics if you're not us

If you're trying to convince an AI skeptic in your team or organization:

→ Listen to their specific concern. Don't assume their skepticism is the same as someone else's.

→ Acknowledge what they're right about. Many skeptic concerns have real basis.

→ Show, don't tell. Get them to TRY something small. Their conclusions from their own experience matter more than your arguments.

→ Accept they might not convert. That's fine. Don't push past good-faith engagement.

→ Use their concerns to improve your deployment. Skeptic concerns often point at real risks you should address.

The skeptic who taught us most

There's one specific customer interaction I think about often. A skeptic from a regulated industry. Several layers of concern. Took a one-hour call.

Most of the call was him explaining why our product (and AI agent platforms generally) couldn't work for his use case. I listened. He had specific compliance concerns, specific data flow concerns, specific reliability concerns.

At the end of the call, I said something like: "I think you're right that cloud-LLM platforms can't work for your use case. I think our local-first architecture might. Want to pilot for two weeks and see?"

He piloted. He bought. He's still a customer, two years later. He's also one of our best advocates because his skepticism was real and his conversion was earned.

The principle

Skeptics aren't obstacles. They're useful filters and improvement sources.

Engage honestly. Don't try to convert. Listen to concerns. Accept that conversion may not happen.

The skeptics who convert become your best customers. The ones who don't sharpen your product anyway. Both outcomes have value.

→ avery.software — Free Desktop tier. Tested by people who weren't sure they wanted to use AI at all.