Three weeks ago I was in a working session on an agentic prototype we’re building. A customer had left negative feedback. The AI caught it. And then the questions started.
How bad is the feedback? Not in absolute terms — that part’s easy. Bad relative to what? This client’s history? Their vertical’s baseline? A score of 6 might be a crisis for a client who’s been a 9 for three years. For a client who onboarded rough and has been climbing, that same 6 might be the first sign things are actually working.
So. Who do we tell?
The account manager seems obvious — until you ask what they should actually do. That depends on what action makes sense to suggest. Which depends on where this client sits in their CX journey, whether they’re in staffing or accounting or AEC, what the contract stage is, what the history looks like. And then: if we suggest something and it gets acted on — what happens next? Right call for this client? Or right call for the average client in this situation?
Those are not the same thing.
Eight questions. All reasonable. None of them the AI’s to answer.
Shreesha Ramdas runs Lumber, an AI-first platform for construction workforce management. He built a system that scans timesheets in real time and catches payroll errors before they become labor board complaints. When I asked how he thinks about AI in a domain where errors have legal and financial consequences, his answer was immediate: “Treat AI as a decision support system with execution capability. Not an autonomous black box.” His system launched at 60% accuracy. It’s at 97% now. Whatever it isn’t confident about gets flagged for a human. Never guessed. Never sent.
Shreesha’s answer came back to me in that working session. Because what we kept running into — with sample data, no real client on the other end — is that detection is the easy part. Meaning-making is the hard part. And meaning-making requires someone who knows the context, carries the relationship, and owns the result when the output is wrong.
Here Is the Failure That’s Coming
An AI agent detects negative sentiment in client feedback. Routes the alert. Someone takes the suggested action. The client — who was at a 6, recoverable, a few good conversations away from a renewal — is now a 3. Nobody caught it because the output looked right and the AI delivered it without hesitation.
Not an AI failure. A process failure. Detection got deployed without the supervision layer underneath it.
Bret Tushaus has been VP of Product at Deltek for sixteen years. He’s watched AEC and professional services firms get this wrong with every technology wave, including this one. “If you have a bad process and you apply AI to it,” he told me, “that bad process may be faster. But it’s still a bad process.”
Every conference talk skips this part. “AI automates tasks” — full stop. No mention of the judgment calls underneath. No mention of who answers when the automation lands wrong.
Nobody Picks This on Purpose
Most firms will end up in the wrong posture. Not because they chose it. Because they didn’t choose at all.
They deploy. Train people on usage. Celebrate throughput. Assume if the output looks right it probably is. And when it isn’t, they find out from the client — which means they find out after the relationship has already taken the hit.
Building the other posture is harder and starts earlier. Before deployment, with a question most teams skip entirely: what are the judgment calls living inside this task, and who owns each one? Accountability gets assigned before the first output goes out. Explicit protocols exist for what the AI handles alone and what gets flagged. Errors show up in the working session, not the client call.
McKinsey named this in April — the new bar for prestige knowledge work is who can supervise the AI, catch its mistakes, and own the client relationship. Landed differently after our working session. Most firms are building the AI. Almost none are building what goes around it.
What Supervision Actually Is
Not reviewing every output — that defeats the purpose.
Shreesha’s team built something more precise. Every decision logged and traceable. Every output explainable — not just what it produced but how the system got there. And when confidence drops, a human gets the call. Whatever falls in that uncertain 3% goes to a person, every time.
But here’s what most PS firm leaders miss: you don’t need a 97% accurate AI system to start this. You need honest answers before any agent touches a client-facing workflow. Which outputs carry real consequence if they’re wrong? Who is accountable for each, by name? When the system surfaces something it shouldn’t handle alone — who makes that call, and is it the same person every time or does it shift by vertical, by client, by stage in the relationship?
If you can’t answer those, you’re in the first posture.
Six months ago, I couldn’t have answered them either. That’s kind of the point.
Building this has changed how I think about the whole category. It is easy to say AI automates tasks. What I’ve learned from a prototype — controlled environment, synthetic data, no real client on the other end — is that a simple task requires a thorough and thoughtful process to work even with a human in the loop. Without one, the bar is a thousand times higher.
Trust is built over years. It takes only a few AI actions to break it.
Lucas Hayden is a product marketing leader at Unanet, a CRM and ERP market leader in AEC. Earlier this year he said something I’ve turned over since: “We still own the craft.” Not as a concession. As a bet. A competitive advantage that holds precisely because the other side keeps automating the wrong things.
The supervision premium isn’t a feature you add to your AI stack. It’s a posture you build before the first mistake reaches a client. After that, it’s damage control.