Should you buy an AI agent platform or build custom?
A decision framework for packaged AI agent platforms versus custom agents in your existing stack: signals, total cost, lock-in, and the exceptions test.
Written by Northstar
Northstar is an AI agent systems studio. Alex leads engineering and product systems; Jordan leads operations and workflow fit. We ship production agents inside tools teams already use.
Alex Morgan · LinkedIn · Northstar
Direct answer
Buy a platform when your workflow is genuinely standard, the vendor's security posture covers you, and your team will operate the product. Build custom - usually agency-built inside your existing stack - when your tools, rules, and exceptions are the actual product and a package would force workarounds at every step. The honest test is not the feature matrix but who handles your exceptions - that is where agent projects live or die.
The real question: who owns the exceptions
Every workflow has a happy path and a long tail of exceptions: the missing field, the refund above threshold. Platforms are built for the median customer's happy path; your exceptions either fit their configuration surface or they do not.
- If roughly 90% of your cases fit the platform's model and the rest escalate cleanly to a human queue, buying is rational.
- If exceptions are a large share of volume, or escalation needs custom logic across your tools, bending the platform costs more than building the workflow directly.
Ask any platform vendor to walk your ten ugliest real cases through their product live; a demo with your data beats the feature matrix.
Decision scorecard
Score each signal; the column with more marks tells you which path to price first.
| Signal | Points toward platform | Points toward custom |
|---|---|---|
| Process shape | Standard, common across your industry | Your rules and exceptions are the moat |
| Tool footprint | Lives mostly in one system the platform integrates well | Orchestrates across CRM, inbox, docs, internal APIs |
| Approval gates | Vendor's built-in approvals fit your risk rules | You need custom gates per action type and role |
| Compliance | Vendor certifications cover your obligations | Data residency or audit rules the vendor cannot meet |
| Exit | Clean export, portable configuration | You want to own prompts, evals, and logic outright |
Mixed scores are normal; the common healthy outcome is a hybrid: a platform for the commodity slice, custom agents for the workflow that differentiates you.
Total cost, compared honestly
Sticker prices mislead in both directions. Compare total first-year cost per path. Figures below are typical market ranges, not Northstar quotes, and vary widely by vendor and scope.
| Cost category | Platform path | Custom path |
|---|---|---|
| Entry | Per-seat or per-agent license, often $50-500 per seat monthly | Fixed pilot, commonly $10,000-50,000 for one workflow |
| Setup | Configuration, sometimes paid onboarding | Included in a properly scoped pilot |
| LLM usage | Usually bundled, capped, or marked up in the license | Your own API keys; tens to low hundreds of dollars monthly per moderate-volume workflow |
| Ongoing | License forever, per seat, repriced at renewal | Optional defined-scope retainer for monitoring and iteration |
| Changes | Wait for the vendor roadmap or pay for workarounds | Scoped changes to a system you own |
| Exit | Migration project, possible data and prompt loss | You keep code, prompts, evals, and docs |
Two costs get forgotten. Platform buyers forget the internal labor of operating the product and handling everything it cannot do. Custom buyers forget ongoing ownership: monitoring, eval upkeep, prompt updates when models change - budget a retainer or internal time for it.
Buy a platform when
- The workflow is a commodity and speed matters more than fit.
- The vendor's security posture beats what you would build: SSO, audit logs, certifications.
- Your exceptions map cleanly to the platform's escalation model.
- The exit is clean: data exports and rebuilds elsewhere without a hostage negotiation.
- You have an operator: a platform without an internal owner becomes shelfware.
Build custom when
- The agent orchestrates across systems you already run: CRM, inbox, documents, internal services.
- Approval gates are specific: who approves which action at what threshold, with an audit trail in your format.
- The workflow should be invisible inside existing tools rather than another tab your team must adopt.
- The logic is intellectual property you want to own, version, and test with your own evals.
- Compliance requires control a multi-tenant product cannot give: data residency, retention rules, model choice.
Custom does not mean from scratch: credible builds assemble proven components - orchestration frameworks, your existing tools, managed model APIs. The build vs buy analysis covers where Zapier-class tools fit in that spectrum.
The trap: buying a platform to avoid process work
The most common failure is not choosing the wrong vendor; it is buying a platform to skip defining the process. The undefined process resurfaces as an unstaffed exception queue, spreadsheet workarounds, and a renewal meeting where usage numbers embarrass everyone. Whichever path you take, the sequence is the same: map the workflow, define the gates, write the acceptance test, then choose the tooling.
What an agency does on each path
- Platform path: workflow mapping before purchase, configuration, gates within the product's limits, acceptance testing, change management.
- Custom path: full systems design - orchestration, integrations, gates, logging, evals - plus documented handoff so you own the result.
- Either path: an honest recommendation should sometimes be "buy the platform"; an agency that only ever answers "custom" is selling capacity, not judgment.
How Northstar fits
Northstar compares both paths inside a paid audit against your real cases and quotes a fixed pilot only for the path the evidence supports. When custom wins, we prefer implementing inside tools you already run. Start with an audit.
FAQ
Yes, and it is really custom with a lower license bill and a higher ownership bill: you avoid per-seat pricing but take on hosting, upgrades, security patching, and upstream debugging. It fits teams with real engineering capacity; for everyone else it becomes abandonware unless the time is budgeted honestly.