Buy-in isn't the problem. It never was.
That was the closing argument Jeremy Roberts, Senior Director of Research and Content at Info-Tech Research Group, made to wrap up Info-Tech LIVE 2026 in Las Vegas. Organizations aren't failing to get executives excited about agentic AI. They're failing to execute it. And the gap between those two things is where most projects die.
"Execution will be the differentiator," Roberts said. "Not the biggest budget. Not the flashiest announcement. The companies that are going to be most successful in the agentic era are the ones who execute most effectively."
To make that point concrete, he shared three production results from Info-Tech's own agent deployments — numbers that don't appear in vendor slide decks or analyst projections, because they came from actually building and running agents inside a real organization.
What Three Workflows Returned
The service desk agent handles intake, triage, routing, escalation, and resolution support. Annual savings: $230,000. SLA compliance: 100%.
The HR onboarding agent manages cross-system orchestration, approvals, provisioning, and handoffs — the fragmented, multi-team process that typically takes days and involves more manual coordination than anyone admits. Annual savings: $410,000. Hours saved: 1,200.
The invoice processing agent handles document intake, extraction, matching, exception handling, and audit. Annual savings and value: $2.7 million. Hours saved: 6,800.
Total across three workflows: more than $3.3 million annually and nearly 8,000 hours returned to the organization.
The point Roberts was making wasn't just about the numbers. It was about the pattern. Three workflows in three different domains — IT operations, HR, finance — using the same underlying architecture. "Your IT service desk and your HR onboarding and your invoice process are actually pretty similar," he said. "If you find the commonalities, you can reduce the amount of new work required every time you build the next one."
That's the core principle: each workflow teaches the next. And it's what separates organizations building a composable agentic capability from organizations building one-off automations that don't compound.
How Agents Actually Fail
Roberts didn't stop at the wins. He walked through the failure patterns — the predictable ways agents break in production that most teams discover the hard way.
The first is judgment failure. An agent reaches a decision point where neither available option clears a confidence threshold, so it stalls. It doesn't fail loudly. It just stops. If there's no fallback path designed into the workflow, the whole process hangs waiting for an intervention that may never come.
The second is tool failure. The agent can't interact with an external resource — an API is down, a permission is missing, a schema has changed — and the system breaks at the integration point. In multi-agent environments, one tool failure can cascade.
Both failure modes are designable. You can build confidence thresholds, fallback paths, human-in-the-loop escalations, and retry logic. But you have to anticipate them before you deploy, not after your first production incident.
"Agents fail in predictable ways," Roberts said. "You can design for these things. You can control for them. You can build contingencies."
When to Say No
The other argument Roberts made — and one that doesn't get enough airtime — is the case for restraint. Not every AI proposal deserves a yes.
"It's time to say no to the wrong AI work," he said. "Every now and then something gets put in front of you — a very silly, million-dollar use case — and you need to use your expertise to gently redirect the organization."
That judgment, Roberts argued, is actually a form of value creation. Organizations that chase every AI opportunity end up with fragmented deployments, technical debt, and governance gaps. Organizations that prioritize ruthlessly end up with the kind of composable architecture that compounds.
The architecture itself has three layers. The composable foundation — agents, orchestration, tasks, and workflows. The agentic toolbox — MCP, guardrails, tools, evaluations, data, and integrations. And the deployment capability — design, develop, deploy, and direct. That stack, once built, becomes the infrastructure every subsequent agent runs on.
The Thread Running Through the Conference
Roberts closed by stitching together the week's arguments into five takeaways, and the thread running through all of them was the same one that ran through the entire conference: the technology isn't what's failing. Execution, governance, and organizational design are.
Tom Zehren challenged the 95% AI failure stat on Day 1 and pointed to the gap between business knowledge and technical implementation. Hans Eckman showed developers how the mental models from the previous era of software don't transfer cleanly to agentic systems. Martin Bufi presented evidence from 75+ agents that production success requires tool-first architecture, not prompt-centric builds. Pearl Almeida demonstrated that 52.9% of agents in production aren't monitored or secured. Nysa Zaran showed that data strategy fails when organizations optimize the pipeline instead of the outcome.
Roberts' closing argument was the same one every speaker arrived at from a different direction: understanding the technology is no longer the constraint. Knowing what to do with it — and having the discipline to do it well — is.
"It's not what people say that matters," he said. "It's what you do."
