If you want to know whether a task is a good fit for an AI agent, ask yourself a simple question:
“Could I give this to a smart intern?”
Not a unicorn intern. Not a 10-year enterprise architect trapped in a 20-year-old’s body. A real intern that is smart, eager, and capable. They are new to the domain and unsure of your systems. They rely on you for clarity, instructions, and feedback loops. These interns need specific direction about what they are and are NOT allowed to do.
This single framing—“the intern test”—has become extremely useful. It helps leaders understand what agentic technology can realistically take off their plate.
Why the Intern Test Works
AI agents are improving rapidly, but their abilities still cluster around the same kinds of tasks that an intern can crush when given:
- Clear definitions
- Stable procedures
- Well-bounded inputs
- Strong examples
- Reversible, low-risk execution paths
If a task requires deep tribal knowledge, cultural context, or executive judgment, both interns and AI agents struggle. They also struggle when dealing with ambiguous human emotion.
But if a task is structured, inspectable, procedural, and verifiable, both interns and agents can thrive. For extra opinion on the verifiable part see this earlier blog post, “Verifiability as the Central Filter for Agentic/GPT Work.”
This is why the intern test works so well: it forces you to articulate what the task actually is, what “good” looks like, and where human judgment must remain in the loop. And if you have employees who can’t answer those questions for their own workflows? That’s already a signal. It indicates where your process documentation, SOPs (or whatever modern label we’re giving them), and internal enablement are weak.
What Makes a Task “Agent-able”?
Run through the same checklist you would use when handing something to a brand-new intern.
1. Is the task unambiguous?
“Go figure out what our customers think about our product onboarding” → Not internable.
“Summarize the last 50 onboarding NPS comments into the top 3 insights” → Absolutely internable.
Ambiguity kills agentic systems. Clarity fuels them.
2. Does the task have a stable definition of done?
If an intern can know when they’re finished without reading your mind, an agent can too.
3. Is the task reversible or low-blast-radius?
Interns shouldn’t push to production. Neither should agents.
But collecting data, drafting content, generating structured proposals, and synthesizing reports? Perfect.
4. Is there a clear pattern to follow?
Interns learn from examples—so do agents.
Providing 3–5 exemplars is often enough to bootstrap an extremely high-quality agent workflow.
5. Can you quickly verify correctness?
This is the big one.
If you can’t evaluate the output easily, then delegating the task—whether to an intern or an agent—is reckless.
If you can evaluate it, then you’ve just unlocked 90% of agentic value.
Why This Mental Model Helps Leaders
The intern test forces a mindset shift:
Stop thinking about what AI should be able to do. Start thinking about what humans at the earliest stage of competence can do.
This reframes AI from “magic automation” into “scalable talent acceleration.”
It helps leaders:
- Map where agents can immediately create leverage
- Identify which processes need tightening before automation
- Avoid over-promising and under-delivering on AI initiatives
- Recognize when high-risk work still requires human seniority
- Build a crawl-walk-run roadmap that matches their org’s maturity
And importantly for anyone in a Gen-AI implementation role:
It tells you exactly where your AI/Data tools (e.g., Agentforce + Data Cloud) deliver real value today—and where you need to maintain human oversight.
But Here’s the Twist: Agents Are Better Than Interns in One Crucial Way
Interns forget.
Agents don’t.
Once you teach an agent a process, a schema, a style, a checklist, or a playbook, it performs that same pattern with consistency. No drift. No fatigue. No “I misunderstood the assignment.”
This means agents excel at:
- Repeated, high-frequency procedures
- Multi-step tasks with rigid rules
- Summaries, classifications, drafting, extraction
- Transforming unstructured → structured data
- Running sales plays or operational workflows exactly as intended
Agents bring infinite patience and zero ego. You bring domain expertise and judgment. Together, that’s leverage.
How to Use This Model in Practice
Next time you are scoping an agent:
- Pull out the checklist. What would you tell an intern? What examples would you show them?
- Write the instructions as if you’re onboarding a new team member. This doubles as your ARC (Agent Requirements Charter) baseline.
- Define the verification path. If you can’t check it, you can’t delegate it.
- Start small. One task → one pattern → one tight loop.
- Promote the agent over time. Just like a human, give it progressively more responsibility as its competence demonstrates itself.
The Bottom Line
AI agents aren’t replacing judgment.
They’re replacing the busywork that prevents judgment from happening.
If you want to surface the best use cases for agentic tech, consider areas like inside sales, marketing ops, customer success, HR, and finance. Stop looking for the biggest, most glamorous problems.
Start with the intern test.
If you’d trust a smart intern to do it with good instructions, a few examples, and a feedback loop, then you can almost certainly build an agent to do it. The agent will operate more consistently and more scalably.
Leave a Reply