Every AI panel I’ve sat in this year eventually gets to the same fork in the road: how fast can we move, and what happens when we move too fast? At VARTECH, a panel titled “Building AI Into Your Business” spent most of its time on the second half of that question, and the answer channel partners walked away with wasn’t a warning to slow down. It was a case for building the guardrails first.
The panel – Greg Buzek of IHL Group, Rob Risany of Intel, Lisanne Schliemann of Denali, and Carl Kallen of Lyngsoe – was framed around helping resellers reposition around AI readiness. But the more useful thread, especially for MSPs already fielding client questions about AI oversight, was what happens after the sale: how do you keep AI from quietly creating more work, more risk, or more noise than it removes?
Every department is its own iceberg
Buzek opened with a stat worth sitting with: by his numbers, 90% of organizations are still stuck in basic automation, while up to 48% of available enterprise value sits behind full workflow reinvention. The gap between those two numbers is where most AI initiatives quietly stall, and Buzek’s explanation was blunt. “It’s not an iceberg,” he said. “It’s a million icebergs. Every single operating department inside the business is its own iceberg.” Each one needs its own readiness assessment before AI touches it, which means the guardrail conversation has to start well before implementation, not after something breaks.
Removing work, not shifting it downstream
Denali’s Lisanne Schliemann drew the clearest line between AI that helps and AI that just adds a review step. Early internal AI rollouts, she said, often just moved work around: “It’s cleaning my email, it’s doing a presentation … but all we did was shift work downstream because we still had a layer that was reviewing the work.” The fix, in her telling, wasn’t more automation, it was building in validation and exception-handling up front, while keeping a human accountable for the final result. “The human is still accountable at the end for the results that are going out to our customers,” she said – a standard she measures against cycle time and accuracy, not just output volume.
A confidence score for every output
Intel’s Rob Risany described one of the more concrete mechanisms on the panel: a confidence-rating system that forces a human to actually read what the AI produced before it ships. A rating of 100 means the output came directly from the source material; 80 means the AI inferred it using its own training; 70 means it pulled from an external source it judged reputable. “Anything that doesn’t meet those three criteria, you rewrite and eliminate,” he said. It’s a simple rule, but it converts “trust the AI” into “verify the AI,” which is a much easier standard to actually enforce across a team.
Move fast, but in stages
Lyngsoe’s Carl Kallen made the case for pacing as its own guardrail. Rather than handing a client’s whole workflow to AI at once, he described a crawl-walk-run approach: delegate a narrow, low-risk decision with guardrails attached, measure the result over a defined period, and only then decide whether to extend it. “You have to take it step by step,” he said – treating every expansion of AI’s role as something earned through a measured track record, not assumed on day one.
The takeaway for MSPs
None of this is an argument against moving quickly on AI – every panelist was, if anything, pushing partners to move faster. But the guardrails they described (readiness assessments, human accountability for outputs, confidence scoring, staged rollouts) are the difference between AI that compounds value and AI that compounds risk. For MSPs building or selling AI-enabled services, that’s the pitch worth leading with: not that AI works, but that you’ve built the structure to make sure it keeps working.
