The AI Filter: Four Questions, Three Examples
A test against the AI trap: four questions, worked through three typical initiatives, plus the one-page form for your next project.
What this guide is for
You have a list of AI initiatives. Probably too long. Probably nobody has ever tested it against anything other than enthusiasm.
This guide gives you a test that runs in an hour and needs no tooling. Four questions, one form, three worked examples.
It will not tell you what to build. It tells you what not to build, and that is the more expensive half of the decision.
The starting point, in one number
We analysed 4,161 coded sales conversations from twelve companies. Most common objection: “no need”, 971 times. It produced zero meetings.
No model, no agent and no better subject line changes that. It is an audience problem dressed up as a tooling problem, because the tooling problem is more comfortable to solve.
That is exactly what the filter is built against.
The four questions
1. Does it bring more revenue at a lower CPL or CAC? Not more activity. Not more leads. More revenue, or the same revenue for less.
2. Does it save money against your current solution? Against the real bill: licences plus working time plus maintenance. Not against the vendor’s list price.
3. Does this process need to exist at all? If it disappeared tomorrow, would anyone notice? Who? How?
4. And is it a bottleneck you have now? Not one you might have in two years. Now.
Four times no means no resources. Not later, not a small pilot, not a side project. No.
The filter is deliberately hard. A soft filter lets everything through, and that describes most roadmaps.
Three worked examples
The three below are archetypes, not client cases. They appear on almost every list. The numbers are example calculations for you to replace with your own.
Example A: the AI SDR that writes first-touch messages
| Question | Answer | Reasoning |
|---|---|---|
| 1. Revenue / CAC | unclear | More messages only lower CPL if the reply rate holds. Check that before scaling. |
| 2. Savings | yes | Replaces writing time, which is measurable: messages per week times minutes per message. |
| 3. Process needed | yes | First touch is part of selling. |
| 4. Bottleneck now | depends | If your reps spend under 20% of their time writing, it is not one. |
Verdict: conditional. Measure for one week how much time actually goes into writing. Under 20% and you are solving a problem you do not have.
Example B: the website chatbot
| Question | Answer | Reasoning |
|---|---|---|
| 1. Revenue / CAC | no | A chatbot generates no traffic. At best it changes what happens to traffic you already have. |
| 2. Savings | no | Under 50 enquiries a month you save nothing. You only move the work. |
| 3. Process needed | no | If it disappeared tomorrow, nobody would notice. That is the answer. |
| 4. Bottleneck now | no | The bottleneck almost always sits earlier: too few people arrive. |
Verdict: do not build. Four times no. The chatbot is the standard example of a project that looks good, can be planned, and changes nothing.
Example C: lead scoring that weights signals
| Question | Answer | Reasoning |
|---|---|---|
| 1. Revenue / CAC | yes | The same selling time spent on better contacts lowers CAC without extra effort. |
| 2. Savings | yes | Less time in contacts who never buy. |
| 3. Process needed | yes | Prioritisation already happens, just on instinct. |
| 4. Bottleneck now | yes, if your list is bigger than your capacity | Otherwise no: at 30 contacts a month you do not need a model. |
Verdict: build, provided question 4 is a yes. And start small: three signals are enough for version one.
The form
Copy this for every initiative on your list. One page, not a deck.
Initiative: ____________________________________________
1. More revenue at lower CPL/CAC? yes / no
Measured how: _______________________________________
Today: ______ Target: ______
2. Saves money vs. current solution? yes / no
Cost today (licence + time + upkeep): ______ /month
Expected cost after: ______ /month
3. Does the process need to exist? yes / no
Who would notice if it went: ________________________
4. Bottleneck we have NOW? yes / no
Evidence: ___________________________________________
Decision: build / conditional (condition: __________) / do not build
Next review on: __________Two rules for the form:
An empty line is a no. If you do not know today’s number, you cannot claim an improvement.
“Conditional” needs a condition with a date. Otherwise it is a yes with a guilty conscience.
The session: one hour, four steps
Minutes 0–10. Every live and planned AI initiative onto one list. Including the ones somebody works on on the side. Especially those.
Minutes 10–40. One form per initiative, five minutes maximum. Anyone arguing longer already knows the answer.
Minutes 40–50. Anything with four times no gets stopped, not paused. Paused means it keeps consuming attention.
Minutes 50–60. Give the freed capacity to a single initiative that scored four times yes. Not two.
What to measure afterwards
Delete one metric from your weekly report: reply rate. In our dataset the campaign angle with the highest reply rate booked zero meetings.
Use instead: meetings, meetings per 1,000 contacts, and effort per meeting. A reply is an intermediate step, not an outcome.
Four principles that stay
The bottleneck decides, not the roadmap. Everything outside the bottleneck is occupation.
Building feels like progress. It looks like work, it can be planned, and as long as building continues nobody asks about the outcome.
An empty cell is a no. Not a “we still need to find out”.
Count what pays. Meetings, pipeline, revenue. Everything else merely feels good.
The short version
Four questions. One hour. Four times no means do not build.
The hard part is not the method. It is stopping an initiative that already has work in it. That is exactly what the filter is for: so the decision hangs on the questions and not on who proposed it.