The enterprise AI industry has a math problem. Cisco data presented at VB Transform 2026 shows that 85% of enterprises are piloting AI agents. Only 5% have shipped them to production. That's not a slow adoption curve — that's a wall.
At VB Transform 2026 last week, Bryan Silverthorn, Director of AGI Autonomy at Amazon, explained exactly where that wall comes from. And his answer upends how most enterprises are thinking about it.
The problem isn't capability. The models are capable enough. The problem is reliability — and most organizations don't even have the right framework to measure it.
The Lie That Internal Evals Tell
Here's the uncomfortable truth: your AI agent probably passed your internal evaluation. It demonstrated the use case. It impressed the demo audience. It made it through the pilot. And then you tried to ship it and discovered it doesn't actually work reliably at scale.
VentureBeat's own proprietary research, presented at VB Transform, put hard numbers on this: half of surveyed companies shipped agents that passed internal evaluations but failed real customers. Half.
There's a structural reason for this failure. Most enterprises track uptime while ignoring accuracy. They check whether the agent is running, not whether it's doing the right thing. They've borrowed the metrics from server monitoring and applied them to something that requires an entirely different measurement approach.
Silverthorn described a customer that deployed an AI agent for software QA work — specifically, extracting serial numbers from screens. It worked flawlessly for two months. Then it started reading wrong numbers intermittently. The culprit wasn't a model failure in the obvious sense. The underlying vision encoder behaved differently depending on where the serial number appeared on screen, and a software update that was invisible to human testers triggered the failure.
This is what production failure looks like for AI agents. It's not dramatic. It's intermittent, hard to reproduce, and easy to miss until it's caused real damage.
The 4 Dimensions of Reliability You're Probably Collapsing Into One
Silverthorn credited Princeton research with a framework that his team now uses to untangle what "reliable" actually means for AI agents. The framework breaks reliability into four distinct dimensions:
1. Consistency — Does the agent produce the same output given the same input, across time and context? This is the dimension most enterprises are accidentally testing. But consistency alone doesn't mean correct.
2. Robustness — Does the agent maintain performance when inputs change, when edge cases appear, when the environment shifts slightly? The serial number failure above was a robustness failure, not a consistency one. The agent was consistent (always read whatever was in a given screen position) but not robust (didn't handle position changes correctly).
3. Predictability — Can operators anticipate how the agent will behave in novel situations? Can they reason about its failure modes before they happen? Unpredictable agents are dangerous even when they're usually right, because you can't build appropriate safeguards around behavior you can't forecast.
4. Safety — Does the agent operate within defined boundaries? Does it avoid unintended side effects? Can it be corrected, paused, or rolled back?
"It unpacks different factors that I see tangled together in almost every eval I've ever seen," Silverthorn said at the conference.
Most enterprise evaluations treat all four as a single question: does it work? That conflation is why agents pass evals and then fail in production. Your evaluation showed the agent worked under controlled conditions, which tested one dimension. It didn't show what happens when those conditions change.
What Most Enterprises Are Trusting Instead
There's a deeper problem beneath the measurement gap: most enterprises aren't building their own evaluation frameworks at all. According to VentureBeat's research from the conference, most enterprises default to the model provider's own evaluations — and little else.
That's not an evaluation strategy. That's outsourcing your risk assessment to the vendor with the strongest incentive to show their model in its best light.
To be fair, evaluating AI agents is genuinely hard. There's no standardized benchmark for reliability in the dimensions that matter in enterprise production environments. The benchmarks that exist were built by AI labs to measure capability, not by operations teams to measure production-readiness.
But hard isn't the same as impossible. The gap is measurement rigor, not measurement impossibility. And the enterprises that figure this out first will be the ones who aren't stuck in the 85% for long.
Amazon's "Intern" Framework: A Mental Model Worth Stealing
The most memorable insight from Silverthorn's talk wasn't technical. Inside Amazon's AGI lab, researchers call their agents "interns." As in: "I'll have my intern talk to your intern."
The joke is also an operating philosophy. Interns are capable but occasionally clueless. They can do impressive work and spectacular derailment in the same week. Managing them requires management skills, not just technical skills.
For AI agents, that means:
- Asking what could go wrong before you deploy — and getting the agent to articulate its own failure modes
- Adding backups and undo capabilities rather than assuming the agent will always get it right
- Consciously deciding what risk you can tolerate — and matching your safeguards to that decision
Silverthorn put it directly: "You can ask the intern, 'Hey, what might you do wrong here? How might you mitigate your negative outcomes?'"
This isn't anthropomorphizing AI. It's applying a time-tested management framework to a novel type of autonomous system. The organizations that treat AI agents like junior employees — with supervision, escalation paths, and structured feedback loops — will run them more reliably than organizations treating them like software that either works or doesn't.
Amazon's AGI lab, for example, accepts that agents occasionally run the wrong experiment in exchange for research velocity. One agent runs experiments around the clock on its own high-level research plan. That's a deliberate trade-off, not an accident. They know what they're accepting, and they've built accordingly.
The Irony: Cisco Wrote the Data and Cisco Is Deploying Anyway
Here's the wrinkle worth sitting with. The 85%/5% statistic comes from Cisco. And Cisco, simultaneously, is planning to deploy AI agents to all of its roughly 90,000 employees starting in August 2026 — one of the largest enterprise-wide agentic AI rollouts announced to date.
That's not hypocrisy. That's sophisticated enterprise thinking.
Cisco knows the production gap exists. They're not deploying because they think the gap doesn't apply to them — they're deploying because they've made a calculated bet that the value of moving fast outweighs the risk of production failures, and they're building the management infrastructure to handle those failures when they happen.
Their approach is worth noting: the system dynamically routes each task to the most cost-efficient model rather than defaulting to frontier models for everything. They're treating agent deployment as an operations problem (what's the right tool for each task?) rather than a technology problem (what's the most powerful model?). That's a maturity signal.
What This Means for Business Leaders
If you're a CFO, COO, or business unit leader evaluating AI agent investments, the Cisco/VentureBeat data changes your risk calculus in a specific way.
The question isn't "should we pilot AI agents?" Most of your peers already are. The question is "what will it take to get from pilot to production?" and "what does production failure actually cost us?"
BCG research from earlier this year found that 60% of companies report minimal or no value from AI despite significant investment, while only 5% are creating substantial value at scale. The production gap isn't just a technical problem — it's where AI investment goes to die.
Three questions worth asking your technical teams before the next budget conversation:
-
How are we measuring reliability, specifically? If the answer is "we ran an evaluation," ask which of the four dimensions — consistency, robustness, predictability, safety — that evaluation actually covered.
-
What does failure look like in production, and what's the detection time? The serial number agent failed for an unknown period before anyone noticed. For your use case, what would intermittent failure cost?
-
What's our escalation path when the agent gets it wrong? Every production AI agent should have a clearly defined human override mechanism, an audit trail, and a rollback procedure.
Organizations that can answer these questions clearly are the ones in the 5%. Organizations that can't are in the 85%.
What This Means for Technical Leaders
For CTOs, VPs of Engineering, and AI/ML leads, the reliability framework is immediately actionable.
Start by mapping your existing evaluations against the four dimensions. Most teams will quickly discover they've been testing consistency (same input, same output) while ignoring robustness (does it still work when inputs shift?). That's the gap where production failures live.
Second, build separate measurement for each dimension. Don't collapse them into a single quality score. A 92% accuracy rate on a standard benchmark tells you nothing about how the agent performs when input distributions shift after a software update. You need dimension-specific tests that simulate the actual variability in your production environment.
Third, implement the intern mental model operationally. This means:
- Pre-deployment failure mode analysis — document what could go wrong, including the failure modes the agent itself can identify when prompted
- Graduated autonomy — start with low-stakes, reversible tasks; expand scope as trust is earned through demonstrated reliability
- Active accuracy monitoring — track whether the agent is getting it right, not just whether it's running
Silverthorn's framing is useful here: "Stop asking whether your agent can do something impressive once, and start asking whether it can do it correctly a thousand times in a row."
That thousand-repetitions standard is what separates pilot from production.
The Path Forward: Measurement First
The enterprises that escape the 85% ceiling won't be the ones with the smartest agents. They'll be the ones with the best measurement frameworks and the most honest evaluation practices.
This requires a cultural shift that's harder than the technical one. Most organizations have built internal processes that reward pilots that look good in demos. Production is where you find out what actually happened.
The good news is that the framework exists. The four dimensions of reliability — consistency, robustness, predictability, safety — give teams a concrete language for having the right conversations before they ship. The intern mental model gives non-technical leaders a way to think about AI agent risk that doesn't require a PhD in machine learning.
The enterprises that will be in the 5% next year aren't the ones with the most advanced AI capabilities. They're the ones running the most honest evaluations and building the strongest management infrastructure around the agents they deploy.
The math problem is solvable. It just requires measuring the right things.
Sources: VentureBeat VB Transform 2026 coverage; Cisco enterprise AI data; BCG 2026 AI Value Report; Princeton reliability framework cited by Bryan Silverthorn, Director of AGI Autonomy at Amazon.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.
Continue Reading
- AI Agents Triple Network Traffic: Most Enterprises Aren't Ready
- Shadow AI Agents: 82% of Enterprises Have Unknown Agents Running
- [Why 80% of AI Agents Deliver ROI While Chatbots Don't](/article/why-80-percent-ai-agents-deliver-roi-chatbots-dont)
