AIANALYSISPODCAST
Saxena, who commercialized IBM Watson, says the pilot to production gap is a missing real time check on what agents do, not a model problem, and the 95% figure he cites is publicly challenged.
Manoj Saxena, the founder and former general manager of IBM Watson, has a blunt number for the enterprise AI agent market: 95% of projects stall between pilot and production (Eye on AI, Aug 24 2026). AI agents here are autonomous software systems that take actions and call tools in business workflows, not chatbots that answer questions.
That figure, popular in trade press this year, traces to MIT NANDA's "State of AI in Business 2025" report. Two independent outlets have publicly challenged the methodology (Futuriom, Aug 2025; AgentMode AI), citing a small self-reporting sample and loose definitions of "pilot" and "production." Saxena leans on the same number, and the dispute travels with it.
Saxena's diagnosis points to a specific mechanism: a missing runtime layer that has to decide, in 10 to 300 milliseconds, whether an agent is allowed to take an action, call a tool, or commit an output. The decision runs against a six-tier requirement stack: from UN human-rights frameworks at the top, through industry regulations and company policy, down to per-customer service-level agreements.
A fraud-detection agent about to block a transaction must clear a runtime check in under 300 ms. The check verifies that the action matches the policy tier, that the tool being called is the one the agent was authorized to call, that the output is not exposing regulated data, and that the call fits the customer's specific SLA on false-positive rate. Skip the check, and the agent either freezes for human review or ships an unreviewed action into production. The check is the work that separates a working demo from a deployed system.
A peer-track paper from late 2025, the MAP benchmark study from UC Berkeley, Intesa Sanpaolo, Stanford, and IBM (arXiv 2512.04123v1), measures exactly this layer. It finds that production agents fail for reasons that have little to do with model quality and a lot to do with tool-selection drift, silent partial failure, and unclear authorization scope. The independent result supports the same diagnosis: production agents fail at the governance layer, not the model layer.
A 2026 refresh of public enterprise benchmarks puts the never-reach-production rate at roughly 89% (luizneto.ai), citing surveys from Deloitte, HCLTech, Nasuni, and Kyndryl. Fiddler, an AI observability vendor, gives a wider range of 70 to 95% depending on the segment (Fiddler blog). The numbers vary, but they point the same way: most enterprise agents do not reach production.
The build-versus-deploy gap, Saxena says, is the cost of skipping the runtime layer. A team can build a working agent in a few hours. Getting it into production, with the governance and observability it needs to operate inside a real company's risk envelope, takes six to seven months. The token-consumption paradox sits on the same fault line: per-token prices keep falling, but agentic AI is more expensive than expected, because each action now triggers a chain of evaluation, logging, and re-routing that the per-token math does not capture.
Saxena ties this to a current milestone. Last month, July 2026, he says, automated agent traffic on the public internet first exceeded human traffic, which he analogizes to AT&T's voice-to-data crossover. The milestone, if confirmed by a public measurement, would change what an "AI deployment" means at internet scale. Saxena gave the claim on the podcast; this article could not independently verify it.
TrustWise, the company Saxena runs, sells the layer he names as missing. The "AI Control Tower" runs in live, sidecar, batch, and simulation modes, and TrustWise claims an 83% cost reduction and a 40% safety improvement for customers using it. Those numbers are vendor-supplied and unaudited. Saxena is not an independent analyst; he is the CEO of the company whose product fits the diagnosis. The diagnosis can still be right; readers should weight the vendor conflict.
Saxena previewed a product line he calls "Genesis agents." These agents would surface beneficial hypotheses in domains like fraud detection and revenue leakage, planning 95 moves deep in a search tree. Deep Blue was a 1997 IBM chess computer that beat world champion Garry Kasparov. Saxena's analogy positions the agents as long-horizon reasoners rather than chatbot-sidekicks. The product is not deployed; it is a roadmap claim.
For enterprise builders, the practical read is two-stage. First, build the capability: a working agent in a few hours. Second, build or buy the runtime control layer that turns capability into production. Saxena's diagnosis is most useful as a budgeting tool for the second stage, not as a verdict on the first.