Why / What / How
A three-question filter for choosing which AI agent initiatives are worth pursuing, before you touch a tool.
Why this paper exists
Most AI agent projects are chosen badly, and they are chosen badly in a predictable way. Someone sees a tool, gets excited, and starts a project around the tool. Six months later there is a working demo that solves a problem nobody important had. The budget is spent and the business is no different.
This paper gives you a filter to stop that before it starts. It is three questions, asked in order: Why, then What, then How. The order is the whole point. Non-technical leaders tend to jump straight to How, because How is where the shiny tools live. Asking Why first kills the weak ideas while they are still cheap. This is the same discipline as starting from the customer and working backwards, applied to your portfolio of AI bets.
At a glance
Before you fund an agent initiative, answer three questions in sequence.
Why are we doing this at all? It has to advance a real strategic goal, remove a real customer friction, or create real business value. If you cannot name which, stop here.
What is it, really? An internal productivity booster, an AI in the customer journey, or a product you sell. The honest answer changes everything downstream.
How will we build it? Which tools, build or buy, what it costs, and how you keep it responsible.
Most failed projects skipped straight to How. Most good ones earned their way down from Why.
Why first
Start with the reason the work should exist, and be strict about it. There are only three answers that count. The initiative advances a strategic goal the leadership team already cares about. It removes a friction a real customer actually feels. Or it creates business value you can describe in plain terms: revenue, cost, risk, speed. "It uses AI" is not a Why. "Our competitor has one" is not a Why. "It would be impressive in the board deck" is not a Why.
The reason to be this strict is that AI agents are unusually good at producing impressive demos for weak reasons. The technology is seductive, the first prototype is quick, and the gap between "looks remarkable" and "matters to the business" is wide and easy to miss. A clear Why is the thing that closes that gap, and if you cannot write one in a sentence the team would agree with, you have just saved yourself a quarter.
What it really is
Once the Why holds, get honest about what you are actually building, because there are three very different shapes and people blur them constantly.
An internal productivity booster helps your own people do their work faster. It is the lowest risk and the easiest to justify, and it is where most organizations should start. The audience is your staff, the bar is "does it save real time," and a rough version is fine.
An agent in the customer journey acts where your customers experience you. The stakes jump, because now a mistake is not an internal annoyance but a customer-facing one. The bar rises to reliability, tone, and trust, and the oversight has to rise with it.
A product is something you build for others to use and, usually, to pay for. This is a different business entirely, with its own economics, support burden, and roadmap. Calling an internal tool a "product" is how teams accidentally sign up for years of obligation they never scoped.
Naming the shape correctly tells you how much rigor, reliability, and ongoing investment the thing actually needs. Most disappointment comes from building one shape and expecting another.
How to build it
How comes last, and only now, because now you know what you are building and why. This is the practical layer, and it has four parts worth deciding on purpose.
Which tools. The platform and the model matter less than people think and should be chosen to fit the job, not the other way around.
Build or buy. Most use cases are better bought or assembled than built from scratch, and the instinct to build everything in-house is where a lot of time disappears.
Cost. Agents have a running cost per task, not just a license, and a use case that makes sense at a hundred runs a day can stop making sense at a hundred thousand. Do that arithmetic early.
Responsible AI. Decide up front where the data goes, what the agent is allowed to do, and how you would explain its decisions. Retrofitting this is painful, and in regulated work it is the difference between shipping and not.
How is where most teams want to start. It is the right set of questions and the wrong place to begin, because a perfect answer to How for a use case that fails Why is just an efficient way to waste a budget.
What this changes for the leadership team
The filter changes what you say yes to, and more importantly what you say no to early. A team that runs Why, What, How on every proposed agent will kill its weak ideas in a meeting rather than a quarter, will build the right shape for the right audience, and will stop confusing motion with progress. The discipline is not anti-ambition. It is what lets you spend your ambition on the few agents that will actually move the business, instead of spreading it across a science fair of demos.
How to use this
Run the three questions on your current list of AI agent ideas, in order, and stop at the first one that fails. When you get to Why, do not accept an answer that uses the word AI in it; make the team state the strategic goal, the customer friction, or the value in business language. When you get to What, push past the comfortable label; an idea everyone is calling an "internal tool" is sometimes a customer-facing commitment in disguise, and it is cheaper to notice that now. Only spend real time on How for the ideas that survived the first two questions. You will usually find your list gets shorter and your remaining bets get better.
What to do this week
Take the AI agent ideas currently floating around your organization and put each one through the three questions on a single page: the Why in one sentence of business language, the What as one of the three shapes, and a first honest pass at How. Anything that cannot clear Why comes off the list. What remains is a portfolio you can actually defend, and a much shorter set of things to build well rather than a long set to build badly.
Frequently asked questions
Is this only for AI agents, or for any AI project? It works for any AI initiative, and for plenty of non-AI ones. It is sharpest for agents because agents are the easiest to start for the wrong reason and the most expensive to run without a clear purpose.
What if the Why is real but small? A small but real Why is fine, especially for an internal booster where the cost and risk are low. The filter is there to stop initiatives with no real Why, not to demand that every one be transformational. Match the size of the investment to the size of the Why.
Who should run this, and when? Leadership, before any budget or build decision, and again as a checkpoint before anything customer-facing goes live. It is a portfolio tool, so it belongs with the people who own the portfolio, not only the team building a given agent.
How does this connect to managing the agent once it is built? Why, What, How decides whether and what to build. Once you decide to build, Managing AI Agents Like Teammates covers how to run the agent well: the job, context, autonomy, tools, and oversight it needs.