ExpertsFMHow is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor)
ACQ2 by Acquired · August 18, 2025
5 Selected Lessons:
framework · 2 reasoning_pattern · 1 prediction · 1 heuristic · 1
Three-tier map of the AI software stack
Source
ACQ2 by Acquired — How is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor) (August 18, 2025) · at 28:11 ▸ · at 30:12 ▸
When this applies
when mapping any emerging technology ecosystem into infrastructure, tooling, and application layers to decide where to build or invest
The takeaway
Segment an emerging tech stack into infrastructure-builders, tool/pickaxe providers, and application-layer composers, and choose which layer to compete in based on your capital intensity tolerance and differentiation source.
Why it works
Each layer has different economics and competitive dynamics: foundation models compete on scarce research talent and capital-intensive training; tool providers sell picks-and-shovels; applied/vertical companies compose models into customer-specific value and are 'downstream' of the capital-intensive layers.
One level up
In any gold-rush-style technology wave, map the stack before choosing a position — the economics differ wildly by layer.
Where it stops holding
This stops working when the layers collapse or vertically integrate (e.g., a foundation model company builds its own vertical apps), erasing the competitive distinctness between layers.
Detail
procedure:
- Identify foundation/frontier model companies (scarce research talent, capital-intensive, license/lease their models)
- Identify tool providers (data labeling, vector databases, RAG infrastructure — the 'pickaxes')
- Identify vertical AI application companies (compose models into end-customer experiences, e.g. Sierra for customer service, Harvey for legal)
- Decide which layer matches your capital structure and differentiation strategy
benchmarks: {}
rationale: Just as a SaaS company wouldn't build its own data centers, an applied AI company shouldn't try to build its own foundation model — rent the scarce capital-intensive layer and differentiate downstream.
In their words
I think there's basically kind of three big categories of AI software companies. First are the foundation and frontier model companies
Outcome-based (resolution) pricing for AI agents
Source
ACQ2 by Acquired — How is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor) (August 18, 2025) · at 15:41 ▸ · at 39:32 ▸ · at 42:34 ▸ · at 48:52 ▸
When this applies
when pricing any product/service that fully completes a task autonomously, as opposed to merely assisting a human to do it
The takeaway
When a product fully substitutes for a human task rather than assisting with it, price on verified outcome completion rather than usage or access, because usage-based pricing doesn't capture or incentivize the actual value delivered.
Why it works
Because agents 'get the job done' rather than just help a human be marginally more productive, seat-based or per-message pricing doesn't map to value delivered; charging only on full task completion (no human involvement) aligns vendor incentive with customer outcome and forces vendor accountability for actual results.
One level up
Pricing model should match the nature of the value exchange: assistance tools can charge for access/usage, but agency-delivering (task-completing) tools should charge for completed outcomes, shifting risk to the provider and forcing genuine value delivery.
Where it stops holding
This stops working when task 'completion' is ambiguous or hard to verify objectively, when the provider can't bear the cash-flow risk of unpaid failed attempts, or when customers prefer predictable fixed costs over variable outcome-based billing.
Detail
procedure:
- Ask what the product is actually doing (assisting vs fully completing a task)
- Reject pricing proxies that don't map to value (seats, messages, tokens, conversations)
- Apply value-based pricing principles, pricing against value delivered
- Define 'resolution' strictly as full completion without human involvement
- Only charge when that resolution bar is met
benchmarks: {"example threshold":"Sierra's bet target was exceeding 80% full-resolution rate within a year, which was exceeded by a significant margin"}
rationale: This deeply aligns incentives: customers lean on the agent as much as possible because it delivers both better experience and savings, and the vendor is motivated to build the most capable agents because it only gets paid on success.
In their words
We arrived at what we call outcome based pricing or resolution based pricing where we only charge our customers when their agent successfully completes the task that it set out to do.
Scarce-to-plentiful lens for judging transformational technology
Source
ACQ2 by Acquired — How is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor) (August 18, 2025) · at 19:05 ▸ · at 22:18 ▸
When this applies
when trying to judge whether any new technology is a true societal inflection point versus an incremental improvement
The takeaway
Test a technology's transformational claim by asking what scarce resource it is converting to abundance and how central that resource was to daily life and identity.
Why it works
Technologies that convert something previously scarce (energy, food, intelligence) into something plentiful restructure what people spend their time and identity on, which is a qualitatively different kind of change than mere efficiency gains.
One level up
The size of a technology's societal impact scales with how foundational the resource it democratizes was to human existence, not with its technical sophistication.
Where it stops holding
This stops working when the resource being made 'plentiful' was never truly scarce or central to identity/livelihood, making the abundance largely irrelevant to societal structure.
Detail
approach: Ask: what was scarce before that this technology makes plentiful, and how deeply was that scarcity tied to identity, livelihood, or daily life structure?
signals:
- resource previously consumed large fractions of human time/attention (food, energy)
- resource is tied to personal identity (intelligence, expertise)
- abundance changes what people spend time on, not just how efficiently they do existing tasks
rationale: Energy and food scarcity-to-abundance transitions reshaped society profoundly; intelligence may follow the same pattern because people tie identity, status, and livelihood to it.
In their words
The first principles way I think about it is what are you making plentiful now that was scarce before and how does that impact society?
Fast, self-serve deployment is required when revenue is gated by successful go-live
Source
ACQ2 by Acquired — How is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor) (August 18, 2025) · at 48:52 ▸ · at 52:18 ▸
When this applies
when a vendor's revenue model is contingent on a customer achieving a working deployment, making deployment speed a direct revenue lever
The takeaway
If your revenue model is tied to a customer achieving a working outcome, treat deployment speed and ease as a core product requirement, not just a UX nicety, because it directly gates your cash flow.
Why it works
Because payment only starts once the product is live and working, slow or IT-gated implementation cycles directly delay vendor revenue — so the business model creates a built-in incentive to make deployment both fast (weeks not months) and accessible to non-technical business teams (no-code tools), bypassing the slow requirements-handoff loop between IT and business teams.
One level up
Let the structure of your revenue model drive which operational capabilities (speed, self-service, cross-team enablement) you must excel at, rather than treating them as separate concerns from pricing.
Where it stops holding
This stops working when deployment complexity is inherently high regardless of tooling (e.g., deep systems integration required) or when the customer's internal processes (legal, security review) are the bottleneck rather than the vendor's tooling.
Detail
rule: Make product setup fast (days to a few weeks) and accessible to non-technical business teams via no-code tools when revenue depends on successful go-live
threshold: typically live in 'a small handful of weeks,' as low as two to three weeks for agile firms, up to a couple months for traditional companies with internal gates
rationale: If it's complicated or slow, you're not earning money till it's live and successful — so speed to delivery and ease of setup directly protect revenue.
In their words
So we focused on a couple of things. We're typically going live in a small handful of weeks. It can be as low as two or three weeks for an agile firm.
Slow, uneven societal absorption of AI despite rapid technical capability
Source
ACQ2 by Acquired — How is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor) (August 18, 2025) · at 22:18 ▸
When this applies
when predicting how fast any breakthrough technology will actually diffuse through an economy with regulatory, cultural, and structural rate-limiters
The takeaway
A technology's capability ceiling and its real-world diffusion rate are governed by different constraints; expect diffusion to lag capability due to institutional rate-limiters.
Why it works
Even if the underlying technology (intelligence) becomes generally available, downstream absorption is gated by non-technology constraints — regulation, cultural resistance, institutional process — the way drug discovery speed doesn't equal faster clinical trials.
One level up
When forecasting disruption, separate 'what the technology can do' from 'how fast society's non-technology bottlenecks allow it to matter.'
Where it stops holding
This stops working when the technology bypasses institutional gatekeepers entirely (e.g., consumer-direct adoption with no regulatory or cultural chokepoint), in which case diffusion can match capability growth.
Detail
approach: Model technology impact as capability (fast-moving) multiplied by absorption rate (slow, gated by regulation/culture/institutions)
signals:
- regulatory approval cycles
- cultural resistance to displaced identity/roles
- sector-specific absorption capacity differing (healthcare vs. software)
rationale: Generalized super intelligence doesn't guarantee rapid real-world therapies or economic shifts because societal rate-limiters remain independent of the technology's raw capability.
In their words
I always use the analogy like you can invent a lot of drugs, but if clinical trials still take a long time, you're not necessarily going to get new therapies rapidly.