On a Tuesday morning in October I was at an AI summit put on by our local chamber of commerce, listening to a panel called AI in Action. There were three people on stage. One ran the local operation of an industrial distributor. One was the director of AI at an engineering firm. One was the CEO of a nonprofit that gives families a place to stay while their kids are in the hospital.
They talked a lot about security. Data privacy. Approved tool lists. Safeguards around the tools on those lists. These are people who answer for client data and for families on some of the worst days of their lives, so I understood why.
I was typing notes from their panel into an AI the whole time they talked.
I take the opposite approach. Let’s get to the edge. I think that’s what has resulted in the cutting-edge products we’ve built at Grafe.
Sitting in that room, I also wondered if maybe that’s a problem.
What the edge built
For months we’ve paid an outside agency thousands of dollars a month to handle our search work. Doing it ourselves would have meant hiring someone who knows search and someone who can build software. A company our size doesn’t staff either one for a project like that.
So I built it myself, with AI doing the parts I couldn’t. We now have an internal dashboard that asks ChatGPT, Claude, Gemini and Google’s AI Overviews the questions our buyers and sellers ask, and shows whether they mention us. Every two weeks it sends our marketing lead an agenda: what to write, what to refresh, what to fix. The AI calls behind it cost tens of dollars a month.
That’s one. Our sellers now get a marketing report after their auction showing where the buyers came from and what the marketing did. And we built a data warehouse that ties every bid, every lot and every website visit together, so I can ask a question like “who has bid on walk-in coolers more than once, within 300 miles of this sale?” and get an answer.
Cost and nobody to staff them ruled out all three. Under an approved tool list, none of them would exist.
The burn
Something like what that panel guards against happened to me in March.
A lot of my email gets drafted by AI. On March 12, one of those replies went out under my name before I’d seen it. It assumed something about a meeting that wasn’t true, and it offered to meet in person on the strength of that assumption. I had no chance to catch it.
I don’t remember exactly how I found out. Probably right away, the moment I was like, “Shit. That sent?” Then I got to write the awkward correction email.
After that it became a rule, and the system reads it every time it starts. Before any email goes out, or even gets saved as a draft, I see the full text and say yes. Every email, no exceptions. If I ask for a quick reply, quick means short, not unreviewed.
That wasn’t the only burn. Our catalog skill, the AI workflow that writes up our lots, wrote up items that didn’t exist. The valuation skill didn’t do as good a job early on as it does today, but a person checks every valuation before a client sees one. The one that sorts lots into categories needed the same tuning. The skill that scans bankruptcy filings for us still has defects that break leads it would otherwise have surfaced. All of them take continual refinement and tuning.
We make mistakes on the edge.
Guardrails after
Those mistakes are getting less and less as we push and hit roadblocks, or get burned, which then allows us to refine at the edges.
So the panel and I both have guardrails. The difference is when they get written. They write theirs before anyone touches the tool: which tools are approved, what data can go in. Ours get written after the burn.
That has a cost. Somebody gets the awkward email first.
What it amplifies
The distributor’s general manager on that panel said the line I wrote down first: “We use AI as an amplifier of our field team.”
Later that morning, a university’s vice provost for AI, a statistician, gave the other half. AI makes existing problems worse. It doesn’t fix a broken workforce or broken workflows.
An amplifier doesn’t care what you plug into it.
That’s the real answer to what I wondered in that room. Getting to the edge isn’t the problem. Getting burned and not writing it down is. If the awkward email never turns into a rule, the next one goes out too, and the edge just makes the same mistake faster.
Where the time goes
The summit had an answer for what to do with the time AI frees up. A specialist from a health records company said it plainly: spend it on person-to-person understanding and relationships.
That’s not quite where mine goes. Mine goes to more nuance, more critical thinking, and some more relationships, in that order. Most of that happens inside fringe projects like those dashboards and the warehouse, the ones that were unattainable before.
Some of it does reach people, just not in a way you’d see away from my desk. I’m doing more touch base calls, and more random calls with people that are scheduled. A lot of them get scheduled because the system reminds me who I haven’t talked to in a while.
What I still don’t know
I don’t think the panel is wrong. An organization that holds families’ information on their hardest days probably should write its guardrails first. But I’d ask them one question: what never got built because the tool wasn’t on the list?
And I have one for myself. When I make one of those calls, is that a relationship? Or just a better-kept list?
I don’t know yet.
FAQ
Should a company approve AI tools before people use them? It depends on what you hold. An organization holding families’ or patients’ information probably should write its guardrails first. At Grafe we go to the edge and put the guardrails on what leaves the building: a person approves every email and every valuation.
What happens when you get burned using AI at the edge? You write the rule. After an AI-drafted reply went out without my approval, every email now waits for my yes. The mistakes get fewer because each burn turns into a guardrail.
Where does the time AI saves actually go? For me, more nuance, more critical thinking, and some more relationships. Most of it goes into projects we couldn’t staff or afford before.
Related: The Person-System Trap, How I Actually Use AI, When the AIs Disagree, I Pay Attention.