Only About 130 Agentic Vendors Are Real. Here Is the Test HUB International Uses.

Gartner counts thousands of vendors claiming to sell agentic AI and estimates roughly 130 are building the real thing. Sandeep Chennikara opened the fifth Lean Into AI session with that number and two more: Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, and MIT's State of AI in Business found that 95 percent of enterprise AI pilots show no measurable impact on the P&L.

His read on why is not what you expect from a technology director. "Everyone is racing to deploy agents. That is the wrong race."

Sandeep directs enterprise automation and AI engineering at HUB International, one of the largest insurance brokerages in the United States. Kelly Reo of The Ohio State University Center for Operational Excellence hosted. Tyson Heaton of the Lean Enterprise Institute stepped in as facilitator for Steve Pereira on short notice, and asked the questions a Lean practitioner would ask.

The test

Before HUB compared vendors, it wrote a definition. It is automation if code decides what happens next: the path is fixed, a person defined it, and it is reliable and blind to anything new. It is agentic if the model decides what to do, in what order, and with which tools, to reach an outcome you set.

That one distinction is a buyer's test. Gartner calls the alternative agent washing, products rebranded as agents without the substance changing. A definition you wrote yourself costs nothing and survives the demo.

Then HUB built a ladder of six levels with a line drawn across it. L0 is deterministic: robotic process automation pulls policy and renewal data overnight. L1 is assisted: AI drafts the coverage-gap summary, a producer edits and sends. The agentic line sits right here. L2 is a bounded agent that checks each renewing policy for gaps while a producer approves. L3 is a reasoning agent that finds the gap and routes the cross-sell itself. L4 is multi-agent, running renewal, gap analysis, and cross-sell across the book. L5 is an ecosystem, with agents reconciling coverage across carriers and divisions.

The note on his slide is worth repeating: most organizations are not operating where they claim to be.

What the demos leave out

"If I watch YouTube or Instagram, I get the impression that everyone is rocketing to level 4 and level 5 overnight," Tyson said.

HUB's production stack has eight layers: human oversight and control. Governance, guardrails, and policy. Orchestration. Agents and agent building. Identity, access, and security. Integration to systems of record. Data and context that is accessible, governed, and agent-ready. Observability and evaluation across all of it.

Two of those eight are what a demo shows. "The demo is the two blue layers," his slide reads. "The enterprise design is everything around them." Sandeep has talked with Google, Microsoft, and Amazon: "I have seen nobody hit all of them."

Two blockers, one with a number attached

Data first. An agent can only reason over data it can reach, in a form it can use, and in insurance that means policy data trapped in legacy systems, PDFs, and carrier portals. The Fivetran Agentic AI Readiness Index is blunt: 15 percent of enterprises say their data foundation can support agents in production, while 60 percent are already investing millions.

Process second, and this one is closer to home. An agent runs on a process. If that process is undocumented and changes by office and by person, there is nothing stable to hand it. Sandeep did not soften it: "No one ever wants to invest in process. That's the last thing you want to invest in. You just want to automate without ever stabilizing the process, or even defining it to begin with."

He has sat through at least half a dozen conferences where someone said clean up your data. Nobody explained how.

The hard choice, and the third option

Sail to the island: build toward the future state now, while the capability still sets you apart. The risk is quicksand: messy data and unstable process sink the agent.

Right the wrongs: fix the broken data and process before you automate anything. The risk is that you never leave the harbor while disruptors pass you.

Tyson named the parallel without prompting. This is the Lean model line argument, thirty years on.

Sandeep did not pick a side. He picked a scope. Take one high-value workflow. Fix just enough data and process to make it agent-ready, then prove it end to end. "Do not boil the ocean, and do not sail on quicksand."

What we are carrying forward

One more number stayed with me. Most companies see roughly 80 percent adoption on chatbots very quickly. Adoption is not the hard part. "Do you use it as a fancy Google search, or are you changing the way that you're working?"

The most useful thing HUB did was not technical. It wrote down what a word means, then asked honestly where it actually stood. Both are free, and both are available this month.

Try this: place your organization honestly on the six-level ladder this week, then name the one high-value workflow you would make agent-ready first, and the smallest data and process fix that would get it there. Bring it to the next Lean Into AI session; the Center for Operational Excellence is planning it now. Watch this one in full on YouTube — https://youtu.be/uDk8FuPjjko — or find it in the Future of People at Work media library at fpwork.org.

Knowledge Map

Process Keywords: agentic definition, agent washing, six-level autonomy ladder, the agentic line, deterministic automation, bounded agents, reasoning agents, multi-agent coordination, eight-layer production stack, orchestration, observability and evaluation, agent-ready data, model line scoping

Context Keywords / reader pain points: vendor claims that cannot be verified, canceled AI projects, pilots with no P&L impact, policy data trapped in PDFs and legacy portals, undocumented processes that vary by office, high chatbot adoption with unchanged work, siloed OpEx and technology groups, pressure to move before the foundation exists

Application Triggers: If your team says "agent" and means different things, then write one definition before the next vendor call. If you cannot say which level you operate at, then place yourself on the ladder before you buy. If a vendor promises an end-to-end platform, then ask which of the eight stack layers they do not cover. If your data foundation is not ready but the investment is already approved, then scope to one workflow rather than the enterprise. If adoption is high but nothing changed, then measure impact on core processes instead of logins. If your OpEx and technology groups do not speak, then ask for sponsorship or a tiger team, not a coffee.

Related Continuous-Improvement Themes: stability before improvement, standard work as the precondition for automation, model line thinking and where to pilot, scoping to prove before scaling, respect for people in redesigning roles around judgment, honest self-assessment as the first countermeasure

This post was developed from the August 26, 2026 Lean Into AI webinar hosted by The Ohio State University Center for Operational Excellence, featuring Sandeep Chennikara of HUB International in conversation with Tyson Heaton of the Lean Enterprise Institute, and synthesized with Claude AI assistance from the recording, transcript, chat log, session abstract, and Sandeep's presentation deck. The full session is on YouTube at https://youtu.be/uDk8FuPjjko. Statistics are cited as presented; the underlying Gartner, MIT, and Fivetran sources are worth reading directly. Corrections are welcome and credited. It represents ongoing work by the Future of People at Work initiative, a collaboration of Catalysis, Central Coast Lean, GBMP Consulting Group, Imagining Excellence, Lean Enterprise Institute, Shingo Institute, The Ohio State University Center for Operational Excellence, Toyota Production System Support Center (TSSC), and University of Kentucky Pigman College of Engineering.