What Was Impossible That's Now Cheap?
Core question: When a technology drops the cost of something by a factor of 1000, the interesting move isn’t “do the old thing faster.” It’s: what could nobody ever afford to do before that’s now trivial? The new question is the strategy.
The articulation
Benedict Evans, May 2026: contrast the two prompts —
- Old, automatable question: “Listen to every call and tell me if the customer sounds strange.”
- New, previously-impossible question: “Listen to a million calls every day — what do you notice?”
The first question existed before AI; AI just makes it cheaper to answer. The second question was not askable before. The cost was prohibitive, and so the question never made it into anyone’s roadmap.
Same pattern for any deterministic question that becomes probabilistic-tractable:
- “Show me the data Bob’s talking about on this call.” (was: impossible during the call)
- “What were the top concerns on customer calls in the last six months?” (was: ten analysts × eight weeks)
- “What pricing changes would improve our churn?” (was: never asked because nobody could answer)
The third one is the most interesting because it has no deterministic answer — only an LLM could attempt it. It’s a question that did not exist as a question in the prior era.
How to apply
- List the questions your team would love to answer but doesn’t bother asking because the cost or labor would be absurd. (“If I had ten analysts for a month, what would I have them do?”)
- For each, run the cost-collapse test: if I could ask this question for $5 and get an answer in 30 seconds, what would I do with the answer?
- The winners are questions whose answer changes a decision that’s currently being made on gut or pattern-matching.
- The biggest winners are questions that nobody is asking yet because the cost has been infinite for so long that they fell out of the discourse.
Why this beats “automate the existing process”
Automating the existing process is the Absorb stage — useful, table-stakes, but symmetric (everyone does it).
Inventing new questions is the Innovate / Disrupt stage — asymmetric, defensible, transformative. The Innovate stage is where Spotify (“get all the music there is”) beat iTunes (“don’t have to buy the CD”). iTunes asked the old question with new technology. Spotify asked a new question.
The 1997 test
“Imagine asking ‘What will be changed by the internet?’ in 1997” — Benedict Evans, channeling Yogi Berra
Most answers in 1997 were “online versions of existing things.” The real answers — search-as-default, social graphs, smartphones, cloud as substrate — were unaskable as questions at the time. The same is almost certainly true now for AI. The dominant 2030 use cases are not on anyone’s roadmap in 2026.
That’s not a reason to wait. It’s a reason to ask questions that don’t yet have product categories.
Failure modes
- Asking only the old questions faster. “What if our existing dashboard had AI?” produces an absorb-stage product. Useful, but not strategy.
- Asking questions with no decision attached. A million-call analysis that doesn’t change pricing, hiring, or product roadmap is a parlor trick.
- Confusing volume for novelty. “Analyze 10x more data” is still the old question. Novelty is in the shape of the question, not the size of the input.
Cross-References
- Companion frames: Was-Cost-of-Task-Your-Moat, Task-vs-Job, Absorb-Innovate-Disrupt
- Related patterns: AI-Planning-Inflection-Point, Augmentation-Over-Automation