Ride the AI Value Wave
Treat today's AI capability as the worst it will ever be and expand where it's most efficacious
- Difficulty
- Moderate
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 88%
Prasanna's answer to the 'AI is overhyped / nothing works' debate: the value of AI changes every day and is rising, so the mistake is asking a static 'where's the value?' The move is to be adaptable — deploy AI today where it's already most efficacious, plan for where it'll be efficacious tomorrow, and slowly expand the frontier as capability grows, rather than throwing tools at your hardest problems and concluding they don't work.
Origin
Prasanna's own framing at Block; the 'this is the worst it will ever be / now the baseline' articulation is co-developed with host Lenny Rachitsky in the conversation.
Core principles
- 01The value of AI is not fixed — it changes every day and is growing in capability
- 02Today's performance is the floor, not the ceiling: 'this is the worst it will ever be'
- 03Deploy where it's most efficacious now (greenfield, non-technical self-serve, UI test automation), not on your hardest legacy code
- 04Match effort to the capability frontier and expand outward as it rises
- 05Senior-human strengths (architecture, race conditions, orchestration, portfolio judgment) are where AI still underperforms — lean into those as humans
How to run it
- 1
Assess today's frontier honestly
Identify where AI already outperforms and where it underperforms your people. Greenfield codebases and non-technical self-serve tools = high gain; giant complex legacy codebases and deep architectural work = low gain.
Pro tip 'Green fields codebase or a app for a new platform' shows the most aggressive gains; complex existing code does not.
Watch out Don't 'just throw these tools at their giant code bases and hoping good things will happen' — that's how skeptics conclude AI doesn't work.
- 2
Deploy where it's most efficacious first
Put AI on the wins available now: let non-technical teams build their own small tools, automate UI testing, run background CI agents on vulnerabilities and bug tickets.
Pro tip Enterprise-risk-management self-serve tools compressed 'weeks of work into hours' — look for internal-apps-team backlog that teams can now self-build.
- 3
Plan for tomorrow's value
Treat every current limitation as temporary. Build roadmap and expectations around where capability will be, not just where it is.
Pro tip Re-evaluate the frontier continuously — the baseline moves under you.
- 4
Reserve humans for portfolio and depth judgment
Keep senior humans on architecture, orchestration, race conditions, and cross-company 'is this even worth doing' judgment, where AI still fails.
Watch out AI 'isn't able to have this sort of portfolio judgment or judgment across a global sense of what's important.'
In the wild
Block's enterprise risk management team used Goose to build a whole self-service enterprise-risk system themselves instead of waiting for an internal apps team to schedule it on a Q2 roadmap.
→ 'Compressing like weeks of work into hours,' turning a queued backlog item into same-day delivery.
Very AI-forward engineering teams using Goose daily self-report 8–10 hours saved per week; across the whole company (support, legal, risk, engineering) Block estimates it is trending toward 20–25% of manual hours saved.
→ Roughly a quarter of an average employee's manual time saved — which Prasanna frames as 'just the start' and 'the worst it will ever be.'
Common mistakes
Asking a static 'where's the value?'
Treating AI value as fixed makes you miss that it 'is changing every day'; you must plan for tomorrow's capability, not just measure today's.
Testing AI on your hardest problem first
Throwing tools at giant legacy codebases produces weak results and false disillusionment; start where it's already efficacious and expand.
Is it for you?
Best for
Product and engineering leaders deciding where to pilot AI and how to set expectations amid the hype-vs-skeptic debate
Not ideal for
Teams needing guaranteed, static ROI before any adoption, or work dominated by deep architectural/orchestration problems where AI still underperforms
From the transcript
“the truth is the value is changing every day. So you need to be adaptable and look at what the value is today and plan…”
“this is the worst it will ever be”
“the companies that aren't feeling the success in AI are trying to just throw these tools at their giant code bases and hoping good things…”
“we're probably trending towards 20 to 25% of manual hours saved”
From the episode
How Block is becoming the most AI-native enterprise in the world
Dhanji R. Prasanna