Sierra and Decagon both closed financing rounds at $4.5 billion valuations in 2026. PolyAI completed its strategic enterprise focus around the same time. The three companies represent the vertical-prebuilt corner of the voice and conversational agent market: customer-service-specific products that arrive with industry-tuned models, pre-built integrations to CRM and ticketing platforms, regulated-vertical compliance defaults, and operating-procedure patterns derived from running tens of thousands of customer-service flows. The build-versus-buy decision an architect faces in 2026 — should we adopt a vertical voice agent product, or should we build one on Vapi, Retell, or Bland? — is no longer the abstract question it was two years ago. The vertical platforms have published case studies, the build-on-platform stacks have shipped at scale, and the unit economics of each path are knowable. This lesson is the honest comparison: where vertical-prebuilt wins, where custom-built wins, and how to calculate the build-versus-buy point for a 100,000-call-per-month workload.

The Three Vertical Platforms of 2026

Sierra, Decagon, and PolyAI each occupy a slightly different corner of the vertical voice and conversational agent market. Understanding which corner is which is the first step to evaluating fit.

Sierra — the conversational AI platform

Sierra was founded by Bret Taylor and Clay Bavor (former Google and Salesforce executives) and raised at a $4.5 billion valuation in early 2026. The platform's positioning is conversational AI for customer experience — both voice and chat, with shared underlying intelligence. Sierra customers include large consumer brands (Sonos, WeightWatchers, Casper, SiriusXM) and enterprise services companies. The product ships with industry-tuned models for customer service, vertical-specific knowledge integration, and a deployment framework Sierra calls Agent Development Lifecycle (ADL).

Sierra's differentiator is the conversational quality. The platform invests heavily in conversation design — the artifacts the Sierra delivery team produces look like a film script, not a chatbot configuration. Sierra also publishes a metric the company calls "outcomes" rather than deflection: the percentage of calls that produced the customer's intended outcome, measured against the customer's success criteria. Outcomes is a more demanding metric than deflection (deflection captures "agent handled it"; outcomes captures "customer got what they wanted") and reflects Sierra's customer-experience positioning.

Decagon — the AI agent platform for enterprise customer service

Decagon raised at a $4.5 billion valuation in early 2026. The platform's positioning is enterprise customer-service agents — primarily chat with growing voice support — built on a configurable Agent Operating Procedures (AOPs) framework. Decagon customers include large fintech, e-commerce, and B2B SaaS companies.

Decagon's Agent Operating Procedures pattern is the differentiator. AOPs are explicit, named, versioned procedures the agent follows for specific case categories. An AOP for "refund request" lists the verification steps, the policy checks, the system calls, the escalation triggers, the response format. The agent does not free-form-decide how to handle the case; it follows the procedure. The pattern is closer to traditional contact-center scripting than to fully autonomous agents, and it is the pattern customer-service operations leaders adopt easily.

The AOPs pattern has implications. It scales with the customer's procedure library — Decagon customers run 100-500 AOPs in production after one year of adoption. It compounds quality (every AOP is testable in isolation; quality regressions are localized to specific AOPs). It constrains autonomy (the agent does not "think outside the procedure," which is exactly what regulated and high-stakes flows want).

PolyAI — the voice-first contact-center platform

PolyAI started in 2017 as a voice-first conversational AI company and matured through enterprise contact-center deployments. The 2026 PolyAI is the most voice-native of the three vertical platforms — chat is a secondary modality, voice is the primary product. PolyAI customers include large hotel chains, financial services, telco, and retail.

PolyAI's differentiator is voice-specific maturity. The platform has solved the voice-specific problems (interruption handling, accent robustness, contact-center integration, IVR replacement) longer than the chat-first platforms. PolyAI's deployment pattern emphasizes the voice channel as a contact-center transformation rather than a chatbot extension; the typical PolyAI deployment replaces an IVR plus offloads a percentage of agent-handled calls.

The Vertical-Prebuilt versus Custom-Built Trade-off

The decision facing an architect: adopt a vertical platform (Sierra, Decagon, PolyAI) or build a custom voice agent on a developer-platform foundation (Vapi, Retell, Bland)?

What vertical-prebuilt gives you

• Faster time to first production call. Sierra and Decagon implementations measured in 8-16 weeks for an enterprise rollout. Custom builds typically run 4-6 months to comparable scope.
• Pre-built integrations. Vertical platforms ship with first-party integrations to Salesforce, HubSpot, Zendesk, Genesys, NICE inContact, and the major CRM and ticketing platforms. Custom builds re-do the integration work.
• Vertical knowledge baked in. The model has seen customer-service flows for thousands of similar enterprises. The prompt engineering, the eval set construction, the failure-mode awareness are not starting from scratch.
• Operating-procedure libraries. Decagon's AOP library, Sierra's flow patterns, PolyAI's contact-center conventions are starting points. The customer extends and customizes rather than building from blank.
• Compliance preparation. Vertical platforms have been through customer security reviews hundreds of times. The SOC 2, HIPAA, GDPR, and PCI-DSS documentation is mature.
• Managed updates. Model upgrades, prompt improvements, framework changes happen behind the platform. The customer's engineering team does not own the upgrade path.

What custom-built gives you

• Cost flexibility. At high call volume, the per-call economics of custom-built can be 30-60% lower than vertical-prebuilt. The cost flexibility is the most common reason large operators choose custom-built.
• Architectural control. The customer chooses the LLM, the STT, the TTS, the orchestration, the observability stack. Swapping any layer is a configuration change, not a vendor negotiation.
• Differentiation. The agent's behavior is exactly what the customer designs. Brand-sensitive voices, unusual flow patterns, vertical-specific intelligence the platforms do not have can be built.
• Data ownership. Call transcripts, model traces, eval data live in the customer's infrastructure. The customer's data does not contribute to the vendor's model.
• Engineering capability building. The team that builds the custom voice agent builds organizational capability that compounds. The team that buys does not.

The fundamental trade-off

Vertical-prebuilt trades flexibility and unit economics for speed, integration depth, and managed operation. Custom-built trades speed and managed operation for flexibility and unit economics. The fork in the road is whether the customer's value comes from the voice agent being faster to ship and well-managed, or from the voice agent being uniquely tuned and cost-optimized.

The 2026 honest answer: most enterprise contact centers under 100,000 calls per month are better served by vertical-prebuilt. Most operators above 500,000 calls per month are better served by custom-built. The 100,000-to-500,000 range is where the calculation gets interesting.

The 100K-Call-Month Build-versus-Buy Calculation

The 100,000-call-per-month workload is the canonical mid-size contact-center workload. At this scale the unit economics of buy and build are close enough that the decision turns on factors beyond cost.

Vertical-prebuilt cost model at 100K calls/month

Typical pricing for Sierra, Decagon, or PolyAI at 100,000 calls per month:

• Platform fee. $30,000-50,000 per month for the licensed seats and platform access. The seat count depends on the customer's operational staffing.
• Per-call charge. $0.40-0.80 per call (3-minute average) all-in. The range reflects whether the customer is on a base or premium tier.
• Implementation. $200,000-500,000 one-time for the initial deployment (8-16 weeks of vendor delivery work). Amortized over a 3-year contract, $5,500-13,900 per month.
• Total monthly. $75,500-143,900 all-in for 100,000 calls per month, including amortized implementation. Approximately $0.76-1.44 per call.

Custom-built cost model at 100K calls/month

Typical economics for a Vapi-based custom voice agent at 100,000 calls per month with 3-minute average call duration:

• Platform per-call cost. $0.12-0.20 per minute all-in = $0.36-0.60 per call. Monthly: $36,000-60,000.
• Engineering FTE. Custom-built requires 2-3 dedicated engineers ongoing (1 voice/ML engineer, 1 platform/integration engineer, 0.5-1 ops/SRE). Loaded cost approximately $25,000-40,000 per month.
• Integration work amortized. Initial build of CRM/ticketing integrations, eval set construction, observability setup. Approximately $150,000-300,000 one-time, amortized over 2 years: $6,250-12,500 per month.
• Compliance burden. Customer's security team handles the SOC 2 / HIPAA / GDPR documentation. Approximately $5,000-10,000 per month equivalent in security-team time.
• Total monthly. $72,250-122,500 all-in. Approximately $0.72-1.23 per call.

What the calculation actually shows

At 100,000 calls per month, vertical-prebuilt and custom-built land within 10-15% of each other on total cost. The decision is not cost-driven at this scale. The decision is driven by time-to-market, by engineering capability the customer has or wants, by integration depth, and by the long-run growth trajectory.

If the workload is expected to grow to 500,000+ calls per month in 18 months, custom-built saves significantly at the higher scale (unit economics gap widens as fixed costs amortize over more calls). If the workload is expected to stay at 100,000 calls per month for several years, vertical-prebuilt is the time-saving choice.

Where the cost calculation breaks down

The honest economics include factors harder to put on a spreadsheet:

• Engineering team retention. Custom-built requires the team that built it. If the voice engineer leaves, the customer's exposure is significant. Vertical-prebuilt insulates against personnel risk.
• Vendor lock-in versus platform-vendor lock-in. Vertical-prebuilt is one vendor; custom-built has 4-7 vendors (LLM, STT, TTS, platform, observability, compliance tooling, integration). The lock-in profile is different but not necessarily lighter.
• Innovation speed. Frontier model improvements roll out faster to platforms than to custom builds. The custom-built team owns the upgrade work; the vertical-platform customer waits for the vendor.
• Customization ceiling. Vertical platforms can be customized to a point. Beyond that point — unusual voice behaviors, vertical-specific intelligence the platform doesn't support — the customer is constrained.

The Agent Operating Procedures Pattern from Decagon

Decagon's Agent Operating Procedures (AOPs) framework is the single most-borrowed pattern in 2026 vertical voice and chat agents. Even custom-built teams have adopted variants of the pattern because it solves problems that fully autonomous prompting does not.

What an AOP is

An AOP is an explicit, named, versioned procedure the agent follows for a specific case category. The structure:

• Trigger condition. The case category that triggers the AOP. "User requests a refund." "User asks about appointment availability." "User reports a billing dispute."
• Verification steps. The identity and context checks the agent performs before proceeding. "Verify identity by account number plus zip code." "Verify last transaction matches user's stated refund subject."
• Policy checks. The business rules the agent applies. "Refund eligible if purchase within 30 days AND not already refunded AND not flagged for fraud review."
• System actions. The tool calls and side effects. "Call refund-issue API with transaction ID. Log to ticketing system. Update customer record."
• Response patterns. The language the agent uses to communicate progress and outcomes. "Inform customer of approval/decline. Provide expected timing. Confirm next steps."
• Escalation triggers. The conditions under which the agent escalates. "Refund amount above $500 = manager approval. Customer explicitly disputes policy = human handoff. Customer escalates emotionally = empathetic handoff."
• Eval cases. The eval set specific to the AOP. Golden, edge, adversarial, escalation cases for the specific procedure.

Why the AOP pattern works

Three reasons.

• Testability in isolation. Each AOP has its own eval cases. Regressions are localized. The team can iterate on one AOP without re-validating the entire agent.
• Operational comprehensibility. Contact-center operations leaders understand procedures. They can read the AOP, propose changes, and validate against their domain knowledge. The pattern fits the customer's existing operating model.
• Constrained autonomy. The agent does not free-form-decide how to handle a refund. It follows the procedure. The constraint reduces the surface area for hallucination and aligns with regulated-industry compliance expectations.

Adopting the pattern in custom builds

The AOP pattern is implementable on any LLM-based agent platform. The customer constructs an AOP library (start with 5-10 highest-volume case categories, expand to 50-200 over time), encodes each AOP as a structured prompt segment (or as a separate tool the agent calls), maintains a per-AOP eval set, and tracks per-AOP metrics in production.

The investment is comparable to Decagon's hosted version (the AOP library is the work, whether the platform is Decagon or custom). The trade-off is that custom-built teams take longer to converge on the right AOP structure; Decagon's experience operating thousands of AOPs across customers has produced patterns the platform encodes by default.

When Vertical-Prebuilt is Clearly the Right Choice

Four patterns where the architect should not even seriously consider custom-built.

Compliance-bound regulated vertical with under 200K calls/month

Healthcare, financial services, insurance — the verticals where compliance documentation is the gating concern and the workload is moderate. PolyAI, Sierra, and Decagon have done the security reviews. The customer's security team adopts the platform faster than they can adopt the custom build.

Short time-to-market constraint

A 90-day deadline to ship a voice agent (board commitment, regulatory pressure, competitive response). Custom-built takes 4-6 months minimum. Vertical-prebuilt is the only path that hits the deadline.

Customer-experience-led differentiation

The CX leader is the project sponsor; the engineering organization is not the value-creating function. The customer values the conversational quality, the integration depth, and the managed operation more than the cost optimization.

Standard customer-service flows

The customer's flows are common — appointment scheduling, balance inquiry, refund processing, status check. The customer is not inventing the wheel. Vertical platforms have these flows productized; the customer customizes rather than builds.

When Custom-Built is Clearly the Right Choice

Four patterns where the architect should not even seriously consider vertical-prebuilt.

High-volume operator above 1M calls/month

The unit economics gap at scale is decisive. Vertical platforms' per-call pricing does not amortize fixed costs the way custom platforms can. A 1M-call-per-month operation saves $400,000-800,000 per year by going custom.

Highly differentiated agent behavior

The customer's value comes from the agent doing something other agents do not. A brand-sensitive voice. A vertical-specific intelligence the platforms do not have. An unusual flow pattern. Custom-built is the only path.

Strong engineering organization with capability-building intent

The customer's engineering team is sophisticated, wants to own the platform, and has capacity to maintain it. Custom-built compounds organizational capability; vertical-prebuilt does not.

Data residency or sovereignty constraints

The customer requires call data to stay in specific infrastructure (on-prem, customer-controlled cloud account, specific geographic region). Vertical platforms ship with defaults; custom-built can match any constraint.

The Hybrid Pattern — Platform Foundation, Vertical Overlay

A 2026 pattern increasingly seen at sophisticated operators: build on Vapi or Retell, but adopt vertical-platform-style operating procedures, conversation design artifacts, and managed-update patterns.

What the hybrid looks like

The custom platform provides the technical foundation — LLM choice, TTS choice, observability, integration. The team adopts the AOP pattern from Decagon, the conversation-design artifact discipline from Sierra, the voice-channel maturity practices from PolyAI. The result is a custom-built agent that operates with vertical-platform discipline.

Why the hybrid works

The technical platform choice and the operating-pattern choice are independent. The vertical platforms' biggest contribution is not the technology stack (the underlying LLMs, STT, TTS are largely the same). It is the operating discipline — AOPs, conversation-design artifacts, managed iteration cycles. That discipline is portable to custom builds.

The trade-off

The hybrid demands more from the team. Adopting the AOP pattern requires investment in procedure library construction, per-AOP eval sets, operations leader collaboration. The platform does not give the customer the procedures for free; the customer builds them. The compensating advantage is that the procedures are exactly tuned to the customer's business, not industry-average defaults.

Key Takeaways

• Sierra ($4.5B valuation), Decagon ($4.5B valuation), and PolyAI are the three vertical-prebuilt voice and conversational agent platforms in 2026. Sierra leads on conversational quality and outcomes-as-metric. Decagon leads on Agent Operating Procedures pattern. PolyAI leads on voice-native contact-center maturity.
• Vertical-prebuilt wins on time-to-market (8-16 weeks vs 4-6 months), pre-built integrations, vertical knowledge, operating-procedure libraries, compliance preparation, and managed updates.
• Custom-built wins on cost flexibility (30-60% lower per-call at scale), architectural control, differentiation, data ownership, and engineering-capability building.
• At 100,000 calls/month, vertical-prebuilt ($75,500-143,900/month) and custom-built ($72,250-122,500/month) land within 10-15%. The decision is not cost-driven at this scale.
• Below 100K calls/month and compliance-bound regulated vertical: vertical-prebuilt almost always wins. Above 500K calls/month: custom-built almost always wins. The 100K-500K range is where the calculation gets interesting.
• Decagon's Agent Operating Procedures (AOPs) pattern is the most-borrowed concept in 2026 vertical agents. An AOP has trigger, verification, policy checks, system actions, response patterns, escalation triggers, eval cases. The pattern is portable to custom builds.
• Vertical-prebuilt is the right choice when: compliance-bound regulated vertical under 200K calls/month, short time-to-market constraint (under 90 days), CX-led differentiation, standard customer-service flows.
• Custom-built is the right choice when: above 1M calls/month, highly differentiated agent behavior required, strong engineering organization with capability-building intent, data residency or sovereignty constraints.
• The hybrid pattern — platform foundation (Vapi or Retell) with vertical-style operating procedures (AOPs), conversation-design artifacts, managed iteration cycles — combines technical control with operating discipline. Demands more from the team but tunes more precisely to the customer.
• The cost calculation breaks down on engineering team retention risk, vendor lock-in versus platform-vendor lock-in trade-off, innovation speed (vertical platforms get model upgrades faster), customization ceiling. The numbers on the spreadsheet do not capture the full economics.