Build vs. Buy for Vertical Agents: A Decision Framework That Survives Contact With Reality

Table of Contents

• Why This Decision Looks Easy and Isn't
• What You're Actually Buying When You Buy a Vertical Agent
• The True Cost of Building (The Line Items Nobody Quotes)
• The Five Real Buy Conditions
• The Four Conditions That Actually Justify Building
• The Hybrid Path Everyone Forgets
• A Scoring Framework You Can Run This Week
• Pricing Models Change the Math
• Insights Most People Overlook
• References

Why This Decision Looks Easy and Isn't

A vertical agent is an AI system tuned for one job in one industry: reviewing commercial leases, drafting prior-authorization appeals, closing the books for a mid-market manufacturer, triaging tier-1 IT tickets. Because the model underneath is a commodity you can rent by the token, a lot of leadership teams reach the same conclusion in the first meeting: "We have engineers. We have an API key. Why pay a vendor a per-seat or per-outcome markup when we can wrap the model ourselves?"

I've watched that reasoning play out enough times to recognize the shape of the regret. The demo gets built in a weekend and looks magical. Six months later the same team is staffing a small standing army to keep it from hallucinating into a compliance incident. The model was never the hard part. The model was the part that was already solved for you.

The build-vs-buy decision for vertical agents is really a question about which 90% of the work you want to own. The generative core is cheap and getting cheaper. Everything wrapped around it, the domain logic, the system-of-record integrations, the evaluation pipeline, the human-in-the-loop escalation, the audit trail a regulator will demand, that's the expensive, durable, unglamorous 90%. This piece is one node in a larger map of why vertical agents beat horizontal platforms, and the moat question and the build-vs-buy question turn out to be two sides of the same coin: whoever owns the workflow data and the integrations owns the outcome.

What You're Actually Buying When You Buy a Vertical Agent

When a vendor sells you an "AI accounts-payable agent," the model is maybe 5% of what you're paying for. Strip the marketing away and you're buying four things that are genuinely hard to reproduce.

Encoded domain expertise. A good vertical agent has thousands of hours of practitioner judgment baked into its prompts, its guardrails, its retrieval logic, and its eval set. The vendor learned which exceptions matter, which field on an invoice trips up 8% of documents, which phrasing a claims adjuster will reject. You're buying their scar tissue. This is the last-mile domain expertise problem, and it is brutal to rebuild from scratch because most of it isn't written down anywhere, it lives in the support tickets the vendor has already absorbed.

Pre-built integrations. The agent already talks to NetSuite, or Epic, or Salesforce, or whatever system of record runs the workflow. Those connectors are maintained, versioned, and tested against the vendor's own customers' edge cases. Integration depth is increasingly the real defensibility in this category, which means it's also the thing you're least equipped to match on your own timeline.

A working evaluation harness. This is the most underrated line item. Any serious vertical agent ships with an eval suite that catches regressions when the underlying model changes, when a prompt is tweaked, or when a customer's data drifts. Anthropic's own guidance on building agents stresses that reliable agents require rigorous, continuous evaluation rather than vibes-based testing, and that harness is a product in itself.

Liability and compliance posture. In regulated verticals the vendor often carries some of the risk, SOC 2, HIPAA BAAs, errors-and-omissions coverage, documented audit trails. Build it yourself and that risk is entirely yours.

The True Cost of Building (The Line Items Nobody Quotes)

When teams estimate a build, they price the obvious stuff: a couple of engineers, model API costs, some prompt engineering. Then reality sends an invoice.

The honest cost of an internal vertical agent includes the eval and observability infrastructure (often a bigger codebase than the agent itself), the data pipeline to feed it your system-of-record context, the human-review tooling for the cases it can't handle autonomously, and the ongoing maintenance treadmill every time a foundation model gets deprecated or a vendor API changes a schema. McKinsey's research on enterprise AI adoption keeps surfacing the same pattern: organizations systematically underestimate the operational and change-management costs of scaling AI, and the gap between a proof of concept and a production system is where most internal builds quietly die.

Then there's the part nobody enjoys: the long tail of edge cases. The 80% solution is a weekend. The 95% solution is a quarter. The 99% solution, the threshold where a vertical agent is actually trustworthy enough to run with reduced supervision, can be a multi-year program. A vendor amortizes that tail across hundreds of customers. You amortize it across exactly one: yourself.

And opportunity cost is real. Every engineer maintaining your homegrown contract-review agent is an engineer not working on the thing your company actually sells. Unless agent-building is what your company sells, building usually means taxing your best people to reinvent a wheel that three startups are already iterating on full-time.

The Five Real Buy Conditions

Buy when most of these are true. In my experience, that's the large majority of cases.

The workflow is standardized across your industry. If your AP process, your prior-auth flow, or your lease abstraction looks roughly like everyone else's, a vendor has already solved it better than you will. Standardization is the vendor's advantage and your signal to buy.

You need it in production this quarter, not next year. Buying compresses time-to-value from quarters to weeks. If the business case is time-sensitive, and the resolution-rate arms race in support shows how fast these categories move, waiting on an internal build can mean losing the window entirely.

The vertical is regulated and you don't want the liability. Healthcare, financial services, legal. Let a specialist who's already cleared the compliance bar carry that weight. This connects directly to how vertical agents win regulated industries: the regulatory moat protects the buyer as much as the vendor.

You lack a standing ML/agent platform team. If you can't staff eval engineering, observability, and prompt maintenance as permanent functions, you cannot operate a serious agent in production. Buying rents that capability.

The pricing is outcome-aligned. Per-resolution or per-outcome pricing means the vendor only wins when you do. That alignment is hard to replicate internally, where every failed agent run still costs you salary and tokens regardless of whether it produced value.

The Four Conditions That Actually Justify Building

Building isn't always wrong. It's right under a narrow and specific set of conditions, and notice that "we're a special snowflake" is not on the list.

Your workflow is your competitive moat. If the way you underwrite, route, or price is genuinely proprietary and central to how you win, you may not want it living inside a vendor's product (and trained, however indirectly, against your competitors' usage). When the agent is the strategy, owning it can be worth the cost.

You have proprietary data no vendor can access. A horizontal vendor trains on the general case. If you sit on a uniquely rich dataset, decades of your own outcomes, building can produce an agent meaningfully better than anything buyable. This is the proprietary-workflow-data moat argument, and it's the strongest genuine case for building.

No credible vendor exists yet. Some verticals are too small or too new for a startup to have shown up. If you're early to a niche and the market map is empty, building may be the only option, though be honest that a vendor may appear in twelve months and outpace you.

You're going to become the vendor. The services-to-software flip is real: agencies and specialist firms are turning their internal agents into products. If your build is a path to a new revenue line, the calculus inverts entirely, now the engineering cost is product investment, not overhead.

The Hybrid Path Everyone Forgets

The framing of "build vs. buy" is a false binary that conveniently ignores the option most mature teams actually land on: buy the platform, build the differentiation.

You can license an agent framework or a vertical platform that handles the undifferentiated heavy lifting, model orchestration, the eval harness, the standard connectors, and then build only the thin layer of proprietary logic that's genuinely yours. This is the "buy the boat, customize the cabin" approach, and it's where the smart money increasingly sits. Andreessen Horowitz has written about how the application layer is where vertical AI companies capture durable value precisely because the orchestration underneath is commoditizing, which means you can rent the commodity and concentrate your scarce engineering on the part that compounds.

Configuration-over-code is the version of this that most non-developer operators should reach for first. A lot of modern vertical agents are deeply configurable without touching their internals: you tune the rules, the escalation thresholds, the integrations, and the brand voice through settings, getting most of the customization benefit of a build with almost none of the maintenance burden. If you can get 90% of the way there by configuring a bought product, the remaining 10% almost never justifies a from-scratch build.

A Scoring Framework You Can Run This Week

Skip the six-week consulting engagement. Score your situation on five axes, 1 to 5, and let the numbers argue.

1. Workflow uniqueness. 1 = identical to industry standard, 5 = genuinely proprietary and central to how you win.
2. Vendor maturity. 1 = several credible, well-funded vendors exist, 5 = nobody serves this niche.
3. Internal capability. 1 = no ML/agent platform team, 5 = a seasoned standing agent team with spare capacity.
4. Time pressure. 1 = needed in production now, 5 = no deadline, strategic horizon.
5. Regulatory/liability exposure. 1 = heavily regulated, want the risk transferred, 5 = low stakes, comfortable owning the risk.

Add it up. Under 12: buy, and don't agonize. 12 to 18: pursue the hybrid path, buy the platform, build the thin differentiating layer. Over 18: building may be justified, but pressure-test it against the opportunity cost before you commit a team. The framework isn't magic; its value is forcing you to put a number on "internal capability," which is the axis teams flatter themselves on most.

Pricing Models Change the Math

One factor warps the entire calculation: how the vendor charges. The GaaS market is rapidly moving from per-seat toward industry-specific value capture and per-outcome pricing, and that shift changes build-vs-buy in your favor as a buyer.

Under per-seat pricing, a high headcount could make building look cheaper at scale, the classic "we'll grow into the savings" argument. But under per-outcome pricing, you pay only for value delivered, which means the vendor absorbs the cost of failed runs, model upgrades, and the entire edge-case tail. A built agent has no such mercy: every failed run, every regression, every 2 a.m. page is on your payroll whether or not it produced anything. When you model the genuinely loaded cost of ownership, not the napkin version, outcome-based buying wins far more often than the build-it crowd expects. Run the real numbers, including the eval team and the maintenance treadmill, before you let a per-seat price tag talk you into a build.

Insights Most People Overlook

The model is the cheapest part, and that's precisely why building tempts you. Cheap, accessible models create the illusion that the whole system is cheap and accessible. It's a cognitive trap: the visible cost (tokens) is trivial, so people anchor on it and ignore the invisible costs (evals, integrations, the edge-case tail) that are 20x larger. The easier the model gets to call, the more dangerous the build-it instinct becomes.

Buying can build a deeper moat than building. Counterintuitive but true. If you buy a great vertical agent and feed it your proprietary workflow data, you accumulate the operational advantage while the vendor maintains the commoditizing plumbing. You get the system-of-record advantage without paying to maintain the boring middle. Building from scratch can actually leave you with a worse, slower-moving system than a competitor who bought and configured well.

"We'll build it to keep our data private" is often backwards. Reputable vendors now offer stronger data isolation, zero-retention contractual terms, and compliance certifications than most internal teams can stand up on their own. The homegrown system frequently has weaker security posture because nobody on the team specializes in it. Privacy is sometimes a reason to buy, not build.

The build decision has a half-life. A vertical you build for today because "no vendor exists" can be eaten by a well-funded startup in a year, the risk that a horizontal platform or specialist vendor eats your custom build is real and accelerating. Any build justified by "the market's empty" should carry an explicit reassessment date. Empty markets in agentic AI do not stay empty.

Your maintenance burden compounds while a vendor's compounds in your favor. Every model deprecation, every API schema change, every new edge case adds to your internal backlog forever. The vendor faces the same forces, but they amortize the fix across their whole customer base and ship it to you for free. Over a three-year horizon the lines diverge sharply, and they diverge against the builder.

References

• Building Effective Agents, Anthropic Engineering
• The State of AI, McKinsey & Company (QuantumBlack)
• AI in the Enterprise: The Application Layer Advantage, Andreessen Horowitz