Look for someone who has actually built or operated an agent that takes action inside a system, not someone who has used AI chatbots well. The distinction matters more than any certification or résumé line, because most of the candidates I’ve spoken with this year are in the early stages of developing how they think about AI. Only a small handful have an actual implementation, and even fewer can comfortably speak to agentic AI use cases and their issues and risks. It ranges from having taken Agentforce training to having built their own MCP integrations. Most are at the early stage of their learning. When looking to hire someone, you want someone who can speak from experience and lay out the issues and risks. You want to hear about their mistakes and what they have learned.
What’s the real difference between someone who’s used AI and someone who’s built with it?
The clearest signal is how someone talks about AI when you’re not asking them to sell you on it. There are two mindsets showing up in the market right now. One treats AI as a chatbot, something you prompt, something that answers questions or drafts an email.
The other treats AI as a doer, something you configure to take action inside a system, read a record, evaluate a condition, execute a step, and hand off or escalate when it hits a limit. That second mindset is the one that actually leads an implementation. The first one can talk about AI fluently and still have never built anything that runs without a human clicking a button first.
How rare is real hands-on agentic AI experience right now?
Rarer than the market suggests. Most people you’ll talk to have read about agentic AI, sat through a vendor demo, or piloted a chatbot layered on top of existing workflows. Fewer have actually stood up an agent that operates inside a live enterprise system with real data and real consequences if it gets something wrong. Fewer still have done it more than once. If someone describes their agentic AI background and every example sounds like a proof of concept that never went to production, that’s worth noticing. Running a tool like Hermes or OpenClaw quickly illustrates agentic AI. It also quickly trains you on token conservation and which models are best suited for the task at hand. Learning frontier models (Codex or Claude Code) and their capabilities is one thing, but finding specific open model (Ollama or LM Studio provided models) capabilities is another. Learning to blend the use of those models for a purpose fit use case is even more training. Take that from the lab or sandbox to the enterprise, and understanding agentic AI, managing it, and knowing how it affects your business stops being optional.
Is running a tool like Agentforce the same thing as being qualified to lead the implementation?
No, and this is where a lot of hiring managers get tripped up. Being able to configure Agentforce, or any agentic AI layer on top of a mid-market ERP or CRM, is a tool skill. Leading the implementation is an architecture and judgment problem: knowing which processes are safe to hand to an agent, which ones need a human checkpoint, how the agent’s actions reconcile with the system of record, and what happens when it’s wrong. Someone can be fluent in the tool and still not be the person you want owning that judgment call.
What should you actually ask in the interview to find out which one you’re talking to?
Ask them to describe a time an agent they built did something wrong, and what happened next. The chatbot mindset tends to answer that question in the abstract, talking about accuracy rates or prompt tuning. The doer mindset answers it with a specific failure, a specific fix, and usually a specific guardrail they added afterward so it couldn’t happen the same way twice. You’re not looking for someone who’s never had it go wrong. You’re looking for someone who’s had it go wrong somewhere that mattered and can tell you exactly what they changed.
Should you hire a dedicated agentic AI lead, or build this into an existing architect role?
For most mid-market companies, the second option makes more sense right now. Leading an agentic AI implementation takes judgment: knowing process risk, data integrity, and where human oversight belongs. That’s the same judgment a strong architect already carries. What’s changed is the expectation that the architect can extend that judgment into agentic tools, not that you need an entirely separate hire. A dedicated AI lead becomes worth it once you’re running agentic AI across multiple systems or business units at real scale. Below that, you’re better served by an architect who’s proven they can operate in the doer mindset, not the chatbot one.
I spent nine years at Salesforce Services, moving from Solution Architect to Engagement Director to RVP, and my job for most of that time was owning delivery and solving problems when implementations went sideways. That’s the same judgment call you’re trying to hire for now, just with agentic tools added to the mix instead of a traditional CRM rollout. I also build agentic systems hands on for my own business, not as a talking point, which is how I’ve been able to tell fairly quickly which mindset someone’s actually operating from when they describe their own work.
