Build vs buy, with the receipts
A demo-grade copilot is a weekend project. A production-grade agent — one that acts on your API for authenticated end users, in their language, without leaking data between tenants — is not. Here is the math, with named sources.
What production-grade actually includes
The gap between the weekend demo and something you can put in front of customers is exactly the part that never shows up in the demo:
- End-user authenticationThe agent must act as the signed-in user — token verification, scopes, step-up confirmation. Teams report the OAuth takes longer than the tool logic.
- Tenant isolationEvery query, cache, log, and metric scoped per customer — and provable, because your buyers' security reviews will ask.
- Action safetyWrite and destructive actions need server-enforced confirmation gates. An agent that deletes without asking generates the tickets it was meant to prevent.
- Multilingual, including RTLYour end users do not all read English. Prompts, UI, and error states in every market language, with right-to-left layouts that actually mirror.
- Evals and regressionsModel and prompt changes need an evaluation harness, or every upgrade is a gamble taken in production.
- Security patching and upkeepStreaming protocols, MCP spec revisions, dependency advisories — the maintenance starts the day the demo ends.
The published failure data
- 95%
- of enterprise GenAI pilots fail to reach production.MIT NANDA, Aug 2025
- 17% → 42%
- abandonment of AI initiatives, year over year.S&P Global, 2025
- 40%
- of agentic AI projects forecast to be canceled by 2027.Gartner, Jun 2025
- 57%
- of buyers expect AI ROI within three months.G2, 2025
Industry consensus puts the build threshold near ~1M conversations per year — below it, buying is the defensible call. Most product teams sit well below it.
The in-house math
An in-house build starts at one AI engineer ($200K/yr loaded, if you can hire one) plus the months before anything customer-safe ships. Syncanix starts at $299/mo, runs the same week, and every allowance is published.
Three ways vendors charge for the same conversation
Pricing model shapes matter more than list prices: the same end user, the same month, three very different bills.
| Model | How it bills | A heavy user’s month | Bill predictability |
|---|---|---|---|
| Per resolution | Each AI-resolved conversation billed — Intercom publishes Fin at $0.99 per resolution. | 12 resolved conversations bill 12 times — $11.88 for that one user at the published rate. | Scales with conversation volume; retry storms and clarification loops show up on the invoice. |
| Per conversation (custom contracts) | Enterprise per-conversation contracts — Decagon does not publish pricing. | Depends on the contract; volume terms are negotiated per deal. | Predictable only after a negotiation; not self-serve. |
| Compute units (Syncanix) | A compute unit (CU) is a weighted measure of what your assistant does — each message, tool call, and the underlying tokens draw CU, so usage drives cost, not headcount. Tiers from $299/mo, each with a published included-CU allowance. | A heavy user who chats and triggers tools draws more compute units than a casual one — you pay for usage, not headcount. Weights are published and shown on a live gauge; at the limit the assistant pauses by default (overage is opt-in at $0.012/unit), never a surprise invoice. | Computable on a napkin — your monthly usage against a fixed included-CU allowance, shown on a live meter. |
Published list prices as of mid-2026; the arithmetic is illustrative. Resolution rates and conversation depth vary by product — this compares pricing mechanics, not outcomes.