Horizontal platforms vs custom agents
The short version: if your need is standard ticket deflection, buy a platform. You will not beat platform economics with a custom build, and we’ll say so before you spend anything.
Last updated:
Buy a platform when
- Your volume is text — email, web chat, common questions
- Your knowledge base is already decent
- You’re deflecting enquiries rather than completing transactions
- Your systems are mainstream SaaS with existing connectors
- You want it live next week
Intercom Fin, Zendesk AI, HubSpot Breeze and Copilot Studio all do this well.
Current platform pricing (2026)
| Platform | Per-resolution | Seat cost (monthly) | Notes |
|---|---|---|---|
| Intercom Fin | $0.99 per outcome | $39–$139/seat | 50-outcome monthly minimum ($49.50 floor) |
| Zendesk AI | $1.50–$2.00 per AR | $19–$149/seat | 5–15 free ARs/agent/mo, then $1.50 committed or $2.00 pay-as-you-go |
| HubSpot Breeze | Included in plan | $20–$150/seat | Credit-based, varies by hub |
Sources: Intercom pricing, 2026; Zendesk pricing, 2026; Getmacha comparison, 2026.
The exact point platforms stop being enough
- When the agent has to complete the transaction, not describe it. Deflection is answering “what’s your returns policy.” Completion is checking live availability in your PMS, holding the room, applying the rate rule and confirming. Platforms read documentation; they don’t reliably act inside your systems of record.
- When your domain knowledge isn’t in a knowledge base. It’s in an ERP, a permit system, a CMMS, or the head of someone who’s been there fifteen years. Nothing to index means nothing to deflect from.
- When voice is your primary channel. Most platforms are text-first with voice added. If most of your volume is calls, that’s a different architecture.
- When you’re regulated. Where a wrong answer creates liability, deterministic logic and approval gates matter more than conversational fluency.
- When your systems are unusual. Platform connectors cover mainstream SaaS. They do not cover your terminal operating system, your historian, or the mining ERP you’ve customised for a decade.
Cost comparison, honestly
| Platform | Custom | |
|---|---|---|
| Time to live | Days to weeks | 4–12 weeks |
| Cost model | Per resolution, $0.99–$2.00 | Build + usage + retainer |
| Standard workflows | Excellent value | Poor value |
| Non-standard workflows | Hits a wall | The reason to build |
| Integration depth | Connector library | Whatever your API allows |
| Ownership | Vendor’s platform | Your infrastructure |
The hybrid nobody mentions
Most businesses should run both. Platform for high-volume standard contact; custom agents for the transactional and operational workflows the platform can’t reach. They coexist, and the platform handles the volume that makes custom work uneconomic.
If a vendor tells you it’s either/or, they sell one of the two.
Frequently asked questions
We already have a platform. Is this a replacement?
Usually a layer above it. Keep the platform for what it does well.
How do we know we’ve hit the wall?
Escalation rate is the tell. If your platform is deflecting a healthy share but escalations are all the same few transactional requests, those are the custom builds.
What about Copilot Studio specifically?
If you’re in a Microsoft tenant it’s often the right home, and we build there. It’s a platform with real extensibility rather than a closed product.
Will you tell us to buy a platform instead?
Yes, when it’s true. It happens regularly during audits.
If we’re the wrong choice, we’ll say so
The audit includes the disqualification list. Several clients have taken the document and built it themselves, which is a fine outcome.