Build vs. buy: the AI gateway decision framework
Innorise Engineering · 28-05-2026 · 6 min read
The decision matrix
| Factor | Build (custom gateway) | Buy (Portkey, Helicone, etc.) | |--------|------------------------|-------------------------------| | Setup time | 4–8 weeks | 1–2 days | | Monthly cost | Infra only (~£2k) | £500–£5k+ | | Provider lock-in | None | Low–Medium | | PII/Compliance | Full control | Vendor-dependent | | Custom routing | Unlimited | Feature-gated | | Team expertise needed | Senior backend + ML ops | Low |
When to build
- Strict data residency (healthcare, finance, gov)
- Multi-tenant with per-tenant cost ceilings
- Custom routing logic (fallback chains, semantic routing)
- You already have the infra team
When to buy
- Speed to market > control
- Single-tenant, standard use cases
- No dedicated ML ops capacity
- Budget < £50k/year
The hidden costs
Buying: vendor price increases, feature gaps you can't fix, data egress fees, compliance audits.
Building: 2–3 senior engineers × 6 weeks = £150k+ before first request. Ongoing: model updates, security patches, observability.
Our default for clients
Start with a managed gateway (Portkey/Helicone). Migrate to custom when you hit a hard constraint: data residency, per-tenant kill-switches, or routing logic the vendors don't support. The migration path is clean if you abstract the gateway interface from day one.
Have a system like this to build?
We architect and ship platforms where correctness and automation matter.
Start a project