Three questions have dominated every investment committee conversation we have had at Milestone AI Ventures over the past eighteen months:
- Will B2B software evolve the way it did during the transition from on-premise to cloud — or will it be displaced entirely by LLMs and other foundation models integrating up the tech stack?
- Will a new set of AI-native startups dominate, as happened in the cloud era? Or will incumbents defend themselves more effectively this time around?
- If the former, will the highly-funded first movers win — or will an as-yet-unknown cohort of fast followers take their place?
These are not academic questions for us. They shape every sourcing decision, every check we write, and every piece of portfolio advice we give. We do not have definitive answers, but we have a framework — and we think the framework matters as much as the conclusions.
The starting point is the Gartner Hype Cycle. Are we currently in the "Inflated Expectations" phase, where enthusiasm has outpaced reality, or are we entering the "Slope of Enlightenment," where genuine productivity gains begin compounding into durable business outcomes? There are strong arguments on both sides, and we have learned to resist picking a single position.
The Case for "Slope of Enlightenment": Real Productivity Gains Are Accumulating
The most compelling evidence that AI has crossed from hype into genuine utility is found in operational data that is increasingly hard to dismiss. The numbers, frankly, are striking:
- Engineering teams are coding approximately 55% faster when using AI code-completion tools, according to GitHub's research on Copilot's impact on developer productivity.
- Physicians are spending roughly 40% less time on clinical documentation, freeing capacity for actual patient care.
- Customer support teams have reduced average call-handling time by 35%, with corresponding improvements in satisfaction scores.
- Pharmaceutical competitive-intelligence workflows have seen 30% efficiency gains, compressing multi-week research cycles into days.
- Insurance underwriting and claims-processing teams report 40% reductions in cycle time across both functions.
These are not anecdotes from friendly early adopters. They are measured outcomes from production deployments at scale. And the revenue trajectories of AI-native companies reinforce the picture: Anthropic reported annualized revenue increasing 3— in five months; Cursor hit approximately $500M ARR within three years of founding; Lovable reached roughly $100M ARR just eight months after launch. Growth at these rates is historically unprecedented.
There is a structural argument as well. The data-plus-workflow-plus-distribution moat may be more durable in this technology cycle than in previous ones. Companies that get into production workflows first accumulate real-world usage data, tighten feedback loops between users and product, and build distribution advantages that are genuinely difficult to displace — even for well-funded followers.
The Case for "Inflated Expectations": The Warning Signs Are Equally Real
Against that optimistic picture, we hold a set of countervailing concerns that we think the market is underweighting.
Macro valuation concentration is historically a leading indicator of correction. When price-to-earnings ratios of the top ten companies separate dramatically from the rest of the S&P 500, history suggests a narrow group of stocks is driving outsized gains — and that a correction typically follows. We are watching those spreads carefully.
Some valuations have disconnected from any plausible near-term commercial reality. Thinking Machines Lab reportedly raised a $2B seed round at a $12B valuation. Safe Superintelligence was valued at approximately $30B despite being pre-product. Figure AI reached a reported $39B valuation. We hold deep respect for the founding teams involved — serial track records command legitimate premiums. But valuations at these magnitudes embed assumptions that require extraordinary outcomes to justify, and extraordinary outcomes are, by definition, rare.
The production-readiness gap remains larger than the press coverage suggests. Most enterprise AI projects stall at the pilot stage. The HBR's "AI experimentation trap" analysis and NTT DATA's research finding that 70–85% of GenAI deployment efforts fail to meet ROI expectations are consistent with what we hear in our own portfolio diligence. Data quality, skills gaps, and integration complexity are not problems that another foundation model release will solve by itself.
The growth narratives bear closer inspection. The "10— engineer" and "AEs closing $1M per month" headlines often resolve, on closer examination, into 9-9-6 work weeks, 5—+ burn ratios, prosumer growth models with structurally high churn, and — most troublingly — circular revenue dynamics where hyperscalers and AI infrastructure providers fund the customers who in turn purchase their services. This circular financing creates the appearance of demand amplification that makes revenue look cleaner than it is.
Different Layers of the Stack Face Different Risk Profiles
One of the most important analytical moves we have made is to stop treating "AI" as a single market with a single risk profile. It is a stack. Bubble risk, moat potential, and incumbent staying-power vary dramatically by layer — and the application layer faces the most acute near-term risk.
Two dynamics are particularly threatening to AI application layer startups. First, foundation model providers are actively expanding up the stack — OpenAI's general task-completion agents and Anthropic's autonomous coding capabilities are the opening moves of a much larger platform push. Expect this to accelerate dramatically over the next eighteen to twenty-four months.
Second, CIOs — who spent the last decade managing the SaaS sprawl created by a "bring your own app" culture — are now reacting with a strong preference for building on foundation models rather than procuring point solutions. The centralized data lake + foundation model architecture is what the hyperscalers are recommending, and a generation of CIOs who lived through the complexity of integrating forty-seven SaaS tools are receptive to the pitch. The implication for application-layer startups is that their traditional buyers are increasingly disposed to build rather than buy.
Where We See Relative Safety: The Vertical AI Thesis
Vertical AI occupies a structurally different position, and it is where we at Milestone AI Ventures are finding the most compelling risk-adjusted opportunities.
Vertical markets are less attractive targets for foundation model expansion. The smaller TAMs that historically made vertical software less interesting to large-cap technology companies now provide a degree of protection against foundation model providers who are optimizing for the largest possible addressable markets. A general-purpose agent that handles horizontal tasks at scale will not prioritize the regulatory compliance requirements of a specific vertical — the economics do not justify the engineering investment.
Vertical AI companies face lower valuation overhang. The largest venture funds — many managing multi-billion-dollar pools for the first time — need investments with the potential to drive meaningful fund-level returns. That requirement pushes them toward horizontal and functional software with massive TAMs, leaving vertical opportunities to smaller, more patient investors. The result is that vertical AI companies tend to raise at more rational valuations, which means the bar for a good outcome is considerably lower and the risk of catastrophic valuation implosion is reduced.
CIO "build" tendencies are industry-specific. Technology-sector CIOs may have the engineering talent to support an internal-build mentality. CIOs in financial services, healthcare, manufacturing, and regulated verticals generally do not — and they prefer vendors who already understand their compliance requirements. This creates genuine purchase intent for well-positioned vertical AI companies that horizontal competitors cannot easily replicate.
The Fast-Follower Question: A Contrarian Investment Thesis
The dominant narrative in AI investing rewards first movers. Raise the biggest round, attract the best engineers, define the category, own the PR narrative, and win customers by being the company most likely to dominate the space. The logic is internally coherent, and it has driven a remarkable concentration of capital into a small number of high-profile companies.
We are skeptical of this consensus, and the skepticism is grounded in historical evidence rather than contrarianism for its own sake.
Google was not the first search engine. The iPhone was not the first smartphone. Salesforce was not the first cloud CRM. Zoom was not the first video conferencing platform. Workday was not the first cloud HCM system. The pattern of fast followers displacing first movers is so consistent across technology cycles that it should at minimum complicate the reflexive preference for the biggest, earliest round.
The academic literature is equally pointed. First movers bear the full cost of demand creation — educating the market, establishing category norms, absorbing the learning costs of figuring out what works — while fast followers study their actions and copy the strategies that work while avoiding the ones that do not. This learning-curve dynamic is especially pronounced when the technology is new, user needs evolve rapidly, and the product-market fit signals are still noisy. High valuation multiples awarded to first movers embed expectations that are historically associated with lower long-run success rates, because they induce scaling behaviors that are difficult to reverse when the underlying unit economics prove weaker than assumed.
The dot-com parallel is instructive here, even if imperfect. The names that dominated the press and fundraising scenes in 1999 and 2000 — Pets.com, Webvan, Kozmo, eToys — ended up very differently from the companies that emerged in the "trough of disillusionment" to define the internet economy: Google, Amazon, Salesforce. We are not predicting that today's AI first movers will follow the former path. But we are paying close attention to what the fast followers building today — with less capital, more realistic expectations, and the benefit of watching what the pioneers got wrong — are capable of.
Our Investment Framework for the Post-AI Software Landscape
Pulling these threads together, here is how we at Milestone AI Ventures are thinking about B2B software investment in the current environment:
Vertical specificity over horizontal ambition at the early stage. We are biased toward companies building for industries with genuine regulatory complexity, specialized data requirements, and buyers who cannot easily replicate the solution internally. Healthcare, financial services, legal, manufacturing, and life sciences all qualify. These markets reward deep domain expertise in ways that generalist competitors find difficult to match.
Healthy skepticism toward first-mover premium in valuation. When evaluating follow-on investments or secondary positions in existing AI companies, we scrutinize the degree to which current valuations embed first-mover assumptions that history suggests are unreliable. A company valued at 50— ARR because it is "first" in a category deserves significantly more diligence than the headline would suggest.
Long-term conviction in fast followers entering the trough. We believe the most durable investment opportunities in the current cycle will emerge from companies that start building in the next twelve to twenty-four months — after the market has separated genuine utility from hype, and when valuations have reset to levels that allow for investor-friendly outcomes without requiring extraordinary exits.
It is early. The honest answer to the title question — will B2B software thrive or die in the post-AI era? — is that we do not know. But combining historical patterns with current market dynamics, we are constructing our portfolio accordingly: overweight vertical AI, skeptical of first-mover premiums, and patient for the fast-follower cohort that will define the next chapter.
Sarah Chen is a Partner at Milestone AI Ventures, where she leads enterprise software and vertical AI investments. She previously worked in growth strategy at a series of B2B SaaS companies and holds an MBA from the Wharton School. The views expressed here are her own and do not constitute investment advice.