Every leadership team feels the itch that we should be doing something with AI. Competitors post about agents, vendors send demos that look like magic if you don’t squint, and boards ask for a plan. Curiosity is fine. Rushing is how you buy a problem with a subscription attached.
We’ve helped clients pilot tools carefully, and we’ve also been called after a pilot became production because someone demoed it to the CEO. That second path has a cost structure. Nobody puts it on the slide.
Data footprints don’t roll back with the license
Agents often want email, tickets, files, and customer records. Once data leaves a controlled environment or gets indexed somewhere you don’t fully understand, you inherit retention and breach questions. Canceling the tool is easy. Un-sharing history is not. I ask vendors where prompts and outputs live, who can see them, and how deletion works. If the answer is a paragraph of marketing, assume the worst and design permissions like you mean it.
The labor just moves
Automation that “saves time” often creates review work: fixing nonsense outputs, reconciling exceptions, explaining mistakes to customers. If you don’t staff supervision, you didn’t remove labor. You relocated it and paid for software on top. One team we know cut ten minutes off a task and added twenty minutes of cleanup three times a week. The dashboard still said success. The people doing the cleanup were less impressed.
Messy processes become software
When an agent encodes a half-documented workflow, the half-documentation becomes load-bearing, and changing the process later means rework. Vendors rarely lead with that. Independent guidance asks whether the workflow is stable enough to automate at all. Sometimes the right first project is fixing the process without any AI, the same way the right website project is clarifying the offer before anyone picks a framework.
Compliance is a design constraint, not a toggle
In HIPAA- and PCI-adjacent environments, a “helpful” agent can quietly violate access norms or create decisions with no audit trail. You don’t sprinkle compliance on after go-live. Skipping it is a classic path to expensive remediation and awkward meetings.
The wrong problem is the most expensive problem
Teams burn cycles on flashy agents while identity is messy and backups are faith-based. That’s not innovation. That’s misallocation with better branding. Our managed IT background is deliberate here: AI advice sits next to operational fundamentals so you don’t polish a house with a cracked foundation.
A slower path that still moves
Deliberate isn’t frozen. Pick one workflow, constrain data access, define human review, time-box the pilot, measure something that matters, then expand, redesign, or stop. Write down which promises were verified. Future you will prefer a short paper trail over “we moved fast because everyone else was.”
One more thing people under-count is political cost. Once an executive has publicly championed an agent, teams get careful about reporting failure, and bad tools linger because saying so feels like criticizing a strategy. Build an exit into the pilot charter so stopping is a success condition rather than a confession. We write that into the plan on purpose. It sounds bureaucratic, and it keeps adults in the room when the demo glow wears off. Also budget for integration glue, because agents rarely live alone and that “couple of hours” work becomes a quarter more often than anyone admits in the kickoff.
If you feel the rush in your chest, treat that as a signal. The organizations that get value from AI are usually the ones who refuse to get pushed by urgency. We’re happy to help with that kind of practical AI consulting—evaluate, select, implement, govern—without the hype tax.