The Last Mile Is Real, But the Moat May Be Shrinking

I read this piece from Judy Rider at Crunchbase News, featuring NEA partner Tiffany Luck on how vertical AI startups can build durable moats. Luck makes a compelling case: the last mile of AI adoption—the gap between a general model's output and a finished work product—is where defensibility lives. She points to portfolio companies like August for legal due diligence and Samaya AI for equity research as examples of startups that own the end-to-end workflow and deliver artifacts enterprises actually trust.

As a seed investor, I find this framework elegant but incomplete. The article is written from a VC's vantage point: Fortune 500s, forward-deployed engineers, and multimillion-dollar contracts. That's a valid slice of reality, but our internal data at PainSignal—which tracks thousands of real-world AI adoption problems from the ground up—suggests the picture is more complex. The last mile is real, yes, but the moat may be narrower than Luck implies, and the real opportunities might lie in places VCs aren't looking.

Take the SMB segment. Our dataset logs over 1,204 distinct problems from businesses with fewer than 100 employees in industries like retail, construction, and food service. These problems—things like inventory reconciliation, shift scheduling, invoice matching—are mundane, painfully manual, and have virtually zero AI solutions available. Average severity: 3.9 out of 5. Builders, this is a goldmine. Yet Luck's entire framing assumes enterprise buyers with engineering resources. The SMB last mile is wide open, and it doesn't require deploying engineers on-site.

Then there's the trust problem. Luck touches on enterprise concerns about data provenance and auditability, but we see a more visceral form of skepticism: frontline workers who simply don't trust AI. Our dataset contains 34 problems explicitly about trust or skepticism, with workers reporting they "don't trust AI to make decisions" or "prefer manual processes." Average severity: 3.7 out of 5. This isn't a feature request—it's a psychological barrier. Startups that design for trust—transparent outputs, user-controlled overrides, gradual automation—can differentiate where technical accuracy alone won't.

Now, the biggest challenge to the moat thesis: horizontal models are improving fast. Luck says most general tools act as "research co-pilots, excellent at taking a user from 0% to 80%." But our data shows that users report general models like GPT-4 now handle up to 70% of domain-specific tasks—contract redlining, code review, financial analysis—but still miss the final 30%. That 30% gap is the moat. But if model capability continues to improve, that gap narrows. We track 47 problems tagged 'model capability' where users are surprised by how much a general model can already do. The implication: startups must build defensibility through proprietary workflows and data, not just task-specific accuracy. A model fine-tuned on legal documents today could be commoditized tomorrow.

Luck argues we haven't yet seen our way of working change—we're still using the same UIs. But our data challenges that, too. We've logged 212 problems related to AI-native interfaces: workers in logistics and healthcare already interact via voice-first or conversational UIs, often forced by new tools that disrupt existing workflows. The friction isn't just about integration—it's about reimagining how work gets done. Startups that treat UI as an afterthought miss the chance to build habits that lock in users.

None of this diminishes Luck's core insight: owning the end-to-end workflow is powerful. Our data reinforces that, with 473 problems tagged 'workflow integration' averaging severity 4.1 out of 5 across healthcare, legal, and manufacturing. These are gaps general models can't fill because they lack domain context, data access, and process knowledge. The moat is real—but it's not static. It's subject to erosion from better models, competing startups, and shifting user expectations.

For founders and investors, the takeaway is twofold. First, don't overlook SMBs—they're hurting and underserved. Second, build for trust and workflow lock-in, not just accuracy. The last mile is where the value lives, but the path is littered with friction that general models are racing to eliminate. The startups that win will be the ones that make the last mile feel like part of the journey, not a detour.