Vertical Agents and the "Last Mile" of Domain Expertise

Table of Contents

• What "Last Mile" Actually Means for an AI Agent
• Why the First 80% Is a Commodity
• Where the Last Mile Hides
• How Vertical Agents Capture Tacit Knowledge
• The Reliability Tax of the Last Mile
• Pricing the Last Mile: From Seats to Outcomes
• The Build-vs-Buy Tension
• Insights Most People Overlook
• References

What "Last Mile" Actually Means for an AI Agent

In telecom, the "last mile" is the stretch of cable from the trunk network to your house. It's a small fraction of the total distance but a large fraction of the cost and complexity, because it has to deal with every messy local condition: the trees, the easements, the one house at the end of a gravel road.

Vertical agents have the same shape. The "trunk network" is the frontier model plus the obvious workflow: read the document, extract the fields, draft the response. Any competent team can wire that up with an off-the-shelf model and a retrieval layer. The last mile is everything between a plausible draft and an answer a domain professional will actually sign their name to.

Consider a prior-authorization agent in healthcare. Pulling the diagnosis codes and the requested procedure from a chart is the trunk. Knowing that this particular payer rejects this code combination unless you attach a specific clinical note, that the appeal window is 14 days not 30 for this plan type, and that the phrase "medically necessary" needs different supporting language for a commercial plan versus Medicare Advantage, that's the last mile. It's not in the model's training data in any reliable form. It lives in the heads of the people who file these claims fifty times a day, and in the scar tissue of past denials.

This is the central truth of the vertical-agent thesis: the value isn't the agent's ability to act, it's the agent's ability to act correctly within a specific domain's rules, including the rules nobody bothered to write down.

Why the First 80% Is a Commodity

It's worth being honest about how little the easy part is worth, because a lot of GaaS startups are quietly building on top of it and calling it a moat.

The first 80% of almost any vertical workflow is now buildable by a small team in weeks. The model providers keep pushing capability and cost in the customer's favor, Anthropic, OpenAI, and Google have all made agent-building primitives like tool use and structured output close to commoditized, and the a16z analysis of where AI value accrues has argued for years that durable margin tends to land with whoever owns the proprietary data and the customer relationship, not whoever wires up the model.

What this means in practice: if your "legal contract-review agent" just summarizes clauses and flags unusual terms, you are one frontier-model release away from being a feature. A horizontal coding-assistant or document-AI platform can absorb that capability into its general product and give it away. The 80% is a race to the bottom because everyone is running the same race with the same engine.

The teams that survive treat the 80% as table stakes, necessary, unglamorous, and not where they spend their differentiation budget. They spend it on the 20% that's specific, defensible, and miserable for a generalist to replicate.

Where the Last Mile Hides

If you want to find the last mile in any vertical, look for the places where current human experts disagree, hesitate, or reach for a colleague. Those are the seams where tacit knowledge lives. A few recurring patterns:

Exception handling. The happy path is documented; the exceptions are tribal. An accounting close runs smoothly until there's an intercompany elimination that doesn't net to zero, and the controller knows from experience it's almost always the same subsidiary's currency translation. That instinct is the asset.

Regulatory interpretation. Regulations are written; their application is negotiated. Two compliance officers at two banks will read the same rule and reach different thresholds for filing a suspicious-activity report. The agent that encodes the defensible interpretation, and can show its reasoning to an examiner, is doing last-mile work.

Counterparty behavior. In procurement, insurance claims, or sales, the other side has patterns. A seasoned claims adjuster knows which body shops inflate estimates and by roughly how much. That pattern recognition isn't in a manual; it's in the data trail of thousands of past claims.

Format and tone landmines. In regulated and high-stakes writing, a denial letter, a patent claim, a clinical note, the difference between acceptable and litigable is often a single hedge word. Experts internalize these landmines. New hires step on them for a year.

The common thread: none of this is reliably present in the public corpus the base model learned from. McKinsey's research on the economic potential of generative AI repeatedly lands on the same point, the value concentrates in deep, function-specific workflows, not in general-purpose text generation. The deeper and more specific the workflow, the longer the last mile, and the more defensible the agent that masters it.

How Vertical Agents Capture Tacit Knowledge

Knowing where the last mile lives is different from capturing it. This is the actual engineering and operations problem behind a serious vertical agent, and there are roughly four mechanisms that work.

Learning from corrections, not just data

The most underrated asset a vertical-agent company has is the stream of human corrections. Every time a paralegal edits the agent's contract redline, every time a coder rejects the agent's pull request, every time a controller overrides a journal entry, that's a labeled example of the last mile in action. Companies that instrument these corrections and feed them back into evaluation suites and fine-tuning datasets compound a lead that pure model access can't buy. This is the proprietary-workflow-data moat that the rest of this cluster keeps circling back to.

Encoding domain logic as guardrails, not prompts

A prompt that says "be careful about payer rules" does nothing. What works is hard-coded domain logic wrapped around the model: deterministic checks, lookup tables of payer-specific requirements, validation that the output conforms to the format a downstream system will accept. The model handles the fuzzy reasoning; explicit rules catch the known landmines. The best vertical agents are mostly software with a model in the loop, not a model with some software around it.

System-of-record integration

The last mile often isn't knowledge at all, it's access. The agent can't decide correctly if it can't see the patient's full history, the deal's full email thread, the property's full transaction record. Deep integration into the system of record is what turns a clever drafting tool into something that operates with the context a human expert has. This is also why integration depth keeps showing up as the real defensibility story in vertical AI.

Expert-in-the-loop design

Early on, the agent should be wrong in ways a domain expert can catch and fix cheaply. The product design that wins routes the genuinely hard 20% to a human, captures that human's decision, and shrinks the share of cases needing escalation over time. The agent earns autonomy; it isn't granted it on day one.

The Reliability Tax of the Last Mile

Here's the part that trips up teams who underestimate the domain: the last mile is exactly where reliability gets expensive, and reliability is what customers are actually paying for.

An agent that's 95% correct on a customer-support reply is a useful product. An agent that's 95% correct on a medication dosage, a tax position, or a contract indemnification clause is a liability. The acceptable error rate isn't a product decision; it's set by the domain, and in regulated verticals it can be brutal.

This creates a counterintuitive economics. The closer you get to full last-mile coverage, the more expensive each marginal point of accuracy becomes, more evaluation data, more edge-case handling, more human review, more domain experts on payroll. The reliability curve flattens hard near the top. A vertical-agent company's real skill is knowing which last-mile cases must be nailed (the ones with legal or clinical consequence) and which can stay in the human-escalation bucket without killing the value proposition.

It's also why agent security and reliability deserve their own line in the budget rather than being treated as polish. A vertical agent that confidently produces a wrong-but-plausible regulatory filing does more damage than one that abstains. The mature pattern is calibrated uncertainty, the agent that knows when it doesn't know and routes accordingly, which is far harder to build than raw capability, and far more valuable in a regulated vertical.

Pricing the Last Mile: From Seats to Outcomes

The last mile is also what justifies the pricing shift that defines the GaaS model. If you're selling the commodity 80%, you're stuck competing on per-seat SaaS pricing against everyone else who licensed the same base model. If you own the last mile, you can credibly price on outcomes, per claim adjudicated, per contract reviewed, per authorization secured, because you're selling a result a customer used to pay a salaried expert to produce.

Outcome pricing only works when the agent reliably crosses the last mile, because outcome pricing puts the vendor on the hook for being right. You can't charge per successfully filed prior authorization if a meaningful slice of your filings get denied for reasons a human would have caught. In that sense, last-mile mastery and per-outcome economics are the same achievement viewed from two angles: one technical, one commercial. The agents that can charge for outcomes are exactly the ones that solved the domain; the ones still charging per seat usually haven't.

The Build-vs-Buy Tension

A natural question for any buyer: if the last mile is what matters, and the last mile lives in our people and our data, why buy a vertical agent at all? Why not build?

It's a fair challenge, and the honest answer is "it depends on whether the workflow is core to your identity." A regional bank's compliance process is differentiated and risk-laden enough that it may justify building, or at least heavily customizing. A dental practice's insurance-billing workflow is painful but not strategically unique, buying a vertical agent that has already crossed that last mile across thousands of practices beats reinventing it.

The deciding factor is usually data scale. A single company sees its own last-mile cases. A vertical-agent vendor serving a hundred companies in the same niche sees a hundred times the edge cases, denials, exceptions, and corrections, and that aggregated view lets it cross the last mile in places no single customer ever could. That cross-customer data network effect is the strongest version of the vertical moat, and it's the reason the services-to-software flip is happening: agencies and specialist firms with deep domain reps are realizing their accumulated judgment is the product, and an agent is the delivery vehicle.

Insights Most People Overlook

The last mile is a moving target, not a finish line. Domains change, new regulations, new payer rules, new contract norms. A vertical agent isn't a product you ship once; it's a maintained relationship with a domain. Teams that treat last-mile coverage as a one-time data-collection sprint get quietly overtaken by teams that built the machinery to keep relearning the domain every quarter. The moat is the update loop, not the snapshot.

Horizontal platforms will absorb the 80% on purpose, to starve you of revenue while you fund the 20%. This is a deliberate strategy, not an accident. The commodity layer gets cheaper and better for free, which sounds great until you realize it's compressing the part of your product that pays the bills. The defensive move is to make sure the customer experiences your value as the last mile, the correctness, the integration, the accountability, not as the drafting, which they'll soon expect for nothing.

The best last-mile data is the data customers are embarrassed by. Denials, rejections, overrides, escalations, the cases that went wrong. Companies instinctively hide these. But the failure trail is precisely the labeled dataset of the domain's hard edges. The vertical-agent company that gets contractual rights to its customers' correction stream, and treats it as a crown-jewel asset, is collecting something a better-funded competitor literally cannot buy.

Sometimes the right last-mile decision is to refuse. A surprising amount of domain expertise is knowing which work not to do automatically. An agent that escalates the genuinely ambiguous 5% with a crisp summary is often more valuable than one that automates 100% with a hidden error rate. Buyers eventually figure out that confident wrongness costs more than honest escalation, and they reward the vendors who built for it.

Last-mile depth and total addressable market trade off against each other. The deeper you go into one vertical's last mile, the narrower your market. This is the uncomfortable arithmetic behind every "we'll start vertical then expand horizontal" pitch. The expansion is hard precisely because the thing that made you defensible, deep, specific domain mastery, doesn't transfer to the next vertical, where a different incumbent has already walked a different last mile.

References

• Andreessen Horowitz: Who Owns the Generative AI Platform?
• McKinsey: The Economic Potential of Generative AI, The Next Productivity Frontier