How I'd Pick the First Task to Hand to an AI Agent (With a 20-Line Scoring Script)

AI agents are the big topic this month, but most people I talk to outside tech haven't tried one yet. The most common mistake I see is picking the tool first and the task second.

As a developer, I'd reverse that. Here's the small framework I use to choose a first task, plus a script you can run.

Chatbot vs. agent, in one line

A chatbot answers when you ask, and then you do the work. An agent takes a goal, plans the steps, uses tools (email, calendar, a spreadsheet, a store), does the task, and reports back.

That difference matters because it changes the risk. An agent can act, so a bad choice of task costs more than a bad chatbot answer.

Why the gap exists

McKinsey's State of AI survey (Nov 2025) reported that 62% of organizations were experimenting with AI agents, but only 23% had scaled them. Lots of pilots, few things in daily use. A small team that picks one narrow job and makes it reliable can get ahead of that.

Score your tasks, don't guess

List everything you repeat each week. Score each task from 1 to 5 on four things:

• Frequency: how often it happens
• Time cost: how long it takes
• Clear rules: how consistent the steps are
• Safety: how easy it is to catch and fix a mistake

Add the four scores. A high total means a good candidate. One rule overrides the total: if safety is 1 or 2, wait, however high it scores elsewhere.

The script

tasks = [
    {"name": "Sort and label incoming email",   "frequency": 5, "time_cost": 4, "clear_rules": 4, "safety": 5},
    {"name": "Answer repeat customer questions", "frequency": 5, "time_cost": 4, "clear_rules": 4, "safety": 3},
    {"name": "Send refunds",                     "frequency": 3, "time_cost": 3, "clear_rules": 4, "safety": 1},
    {"name": "Write monthly newsletter draft",   "frequency": 2, "time_cost": 4, "clear_rules": 3, "safety": 4},
]

MIN_SCORE = 16
MIN_SAFETY = 3

for t in tasks:
    t["score"] = t["frequency"] + t["time_cost"] + t["clear_rules"] + t["safety"]
    t["ready"] = t["score"] >= MIN_SCORE and t["safety"] >= MIN_SAFETY

for t in sorted(tasks, key=lambda x: x["score"], reverse=True):
    flag = "START HERE" if t["ready"] else "wait"
    print(f'{t["score"]:>2}  {flag:<10}  {t["name"]}')

Output:

18  START HERE  Sort and label incoming email
16  START HERE  Answer repeat customer questions
13  wait        Write monthly newsletter draft
11  wait        Send refunds

Refunds never make the list because of the safety rule, even though the task is frequent. That's the point.

Then run it like a developer would

1. Test on old data first. Run the agent on about 20 past cases and compare with what a person actually did. Most failures are fixed by sharper instructions, not a different tool.
2. Pilot with approval on. Every action needs a human OK at first. Track how often outputs are approved unchanged, edited, or rejected.
3. Minimum access. Give it only the accounts it needs. Never a shared admin login.
4. Keep a "never" list. For example: never issue refunds, never promise delivery dates, never invent a policy.
5. Read the logs and keep a kill switch. If you can't see what it did or turn it off quickly, pick another tool.

Measure it

monthly value = (hours saved per week × 4.3 × value of an hour) − tool cost

Record your baseline before you start. Without it you can't tell whether the agent helped.

If you want the longer version

I'm a developer who works with AI, and I put this into a short beginner-friendly guide for non-technical small business owners: 7 starter agents, a 7-day launch plan, prompt templates, and a launch checklist. It's my first ebook, so I'd really like honest feedback.

The AI Agent Playbook for Small Business

What's the first task you'd hand to an agent? Tell me in the comments and I'll tell you how I'd score it.