Impossible-to-Really-Hard AI Opportunity Filter
Prioritize newly feasible problems where AI can create disproportionate leverage
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 97%
The filter begins with problems the organization previously could not solve because the necessary intelligence, personalization, content production, or unit economics were unavailable. Teams reassess those problems against current AI capabilities and retain the ones that have shifted from impossible to merely difficult. They then rank the candidates by business leverage, practical feasibility, and the likelihood of reaching a reasonable initial solution within a few months. The selected project is treated as an iterative system rather than a one-time implementation: launch a bounded version, measure its effect, identify failure modes, and improve it. This concentrates investment where AI changes what the business can do, rather than merely making an existing task marginally faster.
Origin
Kevin Scott, Microsoft’s CTO, framed the underlying idea as tackling problems that have moved from impossible to really hard; Marketing Against The Grain applied it to business AI strategy.
Core principles
- 01Target problems whose feasibility has recently changed
- 02Prefer high-leverage business outcomes over novel technology
- 03Expect an initial solution within months rather than years
- 04Improve the first workable solution through iteration
How to run it
- 1
Recover the impossible list
Identify valuable customer or operational problems the organization previously rejected as infeasible. Record the specific constraint that made each one impossible.
Pro tip Interview functional leaders for problems they stopped proposing because of cost or scale.
Watch out Do not fill the list with low-value tasks merely because they are easy to automate.
- 2
Test changed feasibility
Assess whether current models, tools, data, and integrations can now overcome the original constraint. Retain problems that are difficult but plausibly solvable.
Pro tip Use a small technical spike to test the riskiest capability.
Watch out A compelling demo is not evidence that the full workflow is feasible.
- 3
Rank the leverage
Compare candidates by expected growth, customer value, cost reduction, and strategic differentiation. Choose opportunities whose payoff justifies sustained iteration.
Pro tip Include the value of serving customers who were previously uneconomical to serve.
Watch out Do not prioritize novelty over measurable business impact.
- 4
Build a reasonable first solution
Create a bounded implementation that can produce real evidence within a few months. Define success and failure measures before launch.
Pro tip Constrain the first pilot to one audience, channel, or workflow.
Watch out Avoid waiting for a perfect autonomous system.
- 5
Iterate from evidence
Study errors, user behavior, and business results, then improve the system repeatedly. Expand only after the bounded use case works.
Pro tip Maintain an error taxonomy to guide each iteration.
Watch out Do not scale an unreliable pilot simply because the underlying technology is exciting.
In the wild
A software company cannot economically give every visitor a customized human demonstration. It pilots an AI guide for one product line, trains it on approved product material, measures answer quality and qualified conversions, and improves weak responses before expanding coverage.
→ A formerly uneconomical service becomes a scalable pre-sales experience.
A SaaS company previously could not manually inspect every call, email, and support ticket from customers who left without giving a reason. It uses AI to infer recurring causes, validates a sample, and sends the findings to product teams.
→ Unknown churn is converted into actionable product evidence.
Common mistakes
Automating whatever is easiest
Easy automation may save time but miss the distinctive leverage created when AI makes a previously impossible outcome feasible.
Treating really hard as a multiyear moonshot
The intended opportunities should support a reasonable solution within months, followed by continued iteration.
Scaling before validating reliability
An impressive prototype can conceal data, integration, and quality failures that become expensive at scale.
Is it for you?
Best for
It is best for leaders choosing a small portfolio of consequential AI initiatives.
Not ideal for
It is not ideal for routine improvements that conventional software can already deliver cheaply.
From the transcript
“you should take problems that used to be impossible and are now just really hard.”
“within a couple months, you can have a reasonable solution here, right? And then you're gonna iterate and build on that solution as you figure…”
“invest in the time and energy to make things that were used to be impossible possible today, because that's where you get a lot of…”
From the episode
Steal HubSpot’s Go To Market Strategy With Ai For 2025