The Agent Idea Filter: 5 Questions to Ask Before You Build an AI Agent

Why filter before you build

The most expensive mistake in this field is not a bug; it is a month spent building an agent nobody needed. Students arrive with "an agent that runs my business", clients arrive with "we want AI", and both are sincere and both are unbuildable as stated. The filter is the half-hour that turns either into something a person can build and someone will use. Sardar Ghaffar uses it in the first module of every course and on every agency discovery call, in the same five questions, because the questions are the same whether the builder is a student or a company.

The five questions

Score each from 1 to 5. Write the score down — the point is to compare ideas, not to feel good about one.

Q1Is it repeatable?

Does the task happen more than once — every day, every week, every time a customer does something? A one-off job is not worth an agent; you would spend longer building it than doing it. The best ideas are the ones someone is bored of. 1 = happens once · 5 = happens every day.

Q2Is the input and output clear?

Can you say, in one sentence, what goes in — a message, a form, a question — and what comes out — an answer, a booking, an order in the kitchen? If you cannot, it is not one idea yet; it is two or three mixed together. Split it and score the pieces separately. 1 = vague: "help with marketing" · 5 = "customer sends an order → agent confirms the total and emails the kitchen".

Q3Does it save real time or make real money?

Would a real person be relieved if this stopped landing on their desk? Would a business pay for it? This is the question that separates a project from a product — and it is the one to ask a client on the discovery call, in exactly these words. 1 = nice to have · 5 = someone is doing this by hand for hours every week, right now.

Q4Is it safe to automate?

Nothing in the task needs a human judgement that carries legal, medical, financial or safety risk if the AI gets it wrong. Reminders, bookings, orders and FAQs are safe; diagnoses, contracts and large payments are not agent-only work. Safe also means reversible: if the agent errs, can a human undo it? 1 = a high-stakes judgement call · 5 = low risk, reversible if wrong.

Q5Is it doable with Claude and tools?

Can you name the tools it would need — send an email, check a calendar, read a spreadsheet, look up a menu? If the tools exist or can be built, it is doable with what the course teaches. If it needs robotics, real-time video or something nobody has an API for, park it. 1 = needs something exotic · 5 = text in, tool calls, text out.

Scoring

• Average 4 or more (20+ out of 25): build it.
• Average 3 to 4: usually one question is dragging it down. Rewrite the idea to fix that question — narrow the input and output, add a human approval step for safety — and score again.
• Below 3: park it. Not wrong, just not yet.
• Any 1 on safety: not agent-only work, whatever the total. Redesign so a human approves the risky step.

The worksheet

Five ideas in, one out. Copy it, fill it, keep it — the rejected ideas are next year's.

AGENT IDEA FILTER — score 1 (low) to 5 (high)

#  Idea (one line: who sends what in → what comes out)   Rep  I/O  $$  Safe  Doable  Total
1  ________________________________________________     __   __   __   __    __     __
2  ________________________________________________     __   __   __   __    __     __
3  ________________________________________________     __   __   __   __    __     __
4  ________________________________________________     __   __   __   __    __     __
5  ________________________________________________     __   __   __   __    __     __

Chosen idea:        ______________________________________________
Who does it serve?  (one person or one type of customer — named, not "everyone")
                    ______________________________________________
What does "done"    (the exact output the agent produces when it succeeds)
look like?          ______________________________________________

Ideas that pass

• A café agent that takes WhatsApp orders and confirms the total — the course's own project.
• A tutor's agent that answers "when is my next class?" and reschedules from a calendar.
• A clinic's agent that qualifies a lead, books a slot and sends a reminder.
• A freelancer's agent that drafts a proposal from a client's brief.

Ideas that fail — and why

• "An agent that manages my whole business" — not one clear input and output — it is ten ideas.
• "An agent that gives legal advice" — safety score too low for agent-only work.
• "An agent for one event next month" — not repeatable enough to be worth building.
• "An agent that does my marketing" — vague in and out; split into "write three post drafts from this product page" and score that instead.

After the filter

A chosen idea with a named user and a written success condition is the first line of its 7-Part Job Description: the Role and the Goal are already decided. The "doable with tools" answer becomes the Steps. The safety answer becomes the Rules and the escalation line. That is why the filter is Module 1 and the job description is Module 2 of AI Agent Mastery: one feeds the other, and the agent you build for the rest of the course is the one that passed.