The question comes up every week: "Should we build a custom AI agent or just use something that already exists?"
Most of the time, the answer is "buy." Most of the time, that's also the wrong answer for the team asking.
Here's the thing — the decision isn't about technology preferences. It's about five specific questions, and the answers to those questions produce a clear output. This post walks through the framework.
What "buy" actually means
When people say "buy," they usually mean one of three things:
SaaS tools with AI features baked in. Intercom Fin, Zendesk AI, HubSpot's AI assistant, Freshdesk Freddy. These are support or CRM platforms that added an AI layer on top of their existing product. You configure them, you don't build them. Monthly seat pricing typically runs $50–$150 per agent seat, plus whatever you already pay for the base platform.
No-code AI agent platforms. Tools like Voiceflow, Botpress, or Relevance AI. You wire together flows, connect to your data sources, and deploy. More flexible than pure SaaS, but you're still constrained to what the platform supports. Pricing varies — Relevance AI starts at around $19/month for light use, up to $500+/month at scale.
AI middleware with pre-built connectors. Zapier AI, Make (formerly Integromat), or n8n with LLM nodes. If your workflow is mostly "trigger → LLM call → action," these can get you live in days. Useful for internal ops automation, less useful for customer-facing agents that need nuanced context.
What you do not get from any of these: control over the underlying model behavior, the ability to use your proprietary data as a first-class context source, custom personas, or freedom from per-seat pricing as you scale.
What "build" costs
The honest range for a production AI agent built by an experienced team is $15,000 to $80,000, with most projects landing between $20K and $45K.
Where you land in that range depends on:
- Integration complexity. Connecting to two REST APIs is different from connecting to a legacy ERP with no documented API.
- Data preparation. If your knowledge base is a pile of PDFs and tribal knowledge in people's heads, you have a data problem before you have an AI problem. Data prep can add $5K–$15K.
- Compliance requirements. HIPAA, SOC 2, or financial regulation adds architecture work, logging requirements, and testing time.
- Conversation scope. A narrowly scoped agent that handles one type of request (e.g., "answer questions about our pricing") costs a fraction of a general-purpose agent that routes across ten workflows.
Ongoing costs after launch are real: API fees (typically $200–$1,500/month depending on volume), monitoring infrastructure, and iteration time. Budget 15–20% of the build cost per year for maintenance.
The five questions
1. How custom is the workflow?
Describe your agent's job in one sentence. If that sentence sounds like "answer customer support questions," an off-shelf tool probably handles it. If it sounds like "review incoming purchase orders against our supplier contracts, flag discrepancies above $500, and route to the right approver based on department and dollar threshold," you have a custom workflow.
The test: can you configure an existing tool to do this within two hours, without writing code? If yes, buy. If no — or if you're not sure — keep reading.
2. Do you have proprietary data?
The value proposition for building custom almost always involves data you own that no vendor has. Proprietary product documentation. Internal process knowledge. Historical transaction records. Customer interaction history.
If your agent's quality depends on accessing this data accurately and keeping it within your control, buying creates a ceiling on performance. Intercom Fin can only be as good as what Intercom's platform can ingest. A custom RAG-based agent built against your own data has no such ceiling.
Ask yourself: if a vendor's AI model gives a wrong answer because it didn't have access to our internal documentation, is that an acceptable outcome? If the answer is no, you're building.
3. Is compliance required?
HIPAA, SOC 2 Type II, PCI-DSS, GDPR with data residency requirements, FedRAMP — any of these changes the calculus.
Most SaaS AI tools process data on vendor infrastructure. If your use case involves protected health information, financial data subject to regulation, or anything with contractual data residency requirements, check vendor compliance documentation before assuming they cover you. Many do have compliance certifications; verify they apply to the AI feature specifically, not just the base product.
If your compliance requirements are strict and the vendor's coverage is ambiguous, you'll want to self-host or use a cloud-native build where you control the data flow end to end.
4. How many seats, and how fast will that grow?
Run the math on per-seat pricing. At 50 users, $100/seat/month is $5,000/month — $60K/year. A custom build at $30K is break-even in six months.
At 10 users, that math doesn't work. At 10 users, you haven't even started to amortize the build cost.
The crossover point depends on your specific vendor pricing, but once you're above 100–150 seats or expect to reach that within 12–18 months, buying usually costs more over a three-year horizon than building. Especially when you factor in that vendor pricing tends to go up, not down.
5. What's the maintenance plan?
Building is not a one-time cost. Someone has to own the agent after launch. That means:
- Reviewing conversation logs for failures
- Updating the knowledge base when your product changes
- Adjusting prompts when behavior drifts
- Monitoring API costs and model performance
If you have a developer or technical team member who can own this part-time, maintenance is manageable. If you're expecting to "set it and forget it," you should buy — or at least hire for this before you build. An unmaintained custom agent degrades over time as your data gets stale and model providers update their underlying models.
Decision matrix
| Situation | Recommendation | |---|---| | Standard workflow, no proprietary data, under 50 seats | Buy (Intercom Fin, Zendesk AI, etc.) | | Standard workflow, no compliance needs, fast timeline | Buy or use a no-code platform | | Custom workflow, proprietary data, 100+ seats | Build custom | | Strict compliance (HIPAA, SOC 2), data residency | Build custom or self-host | | Proprietary data is the moat, brand matters | Build custom | | Early stage, under 10 users, validating the concept | Buy or prototype with existing tools | | Existing tool getting expensive, workflow is mature | Migrate to custom |
One more consideration: the hybrid path
You don't have to choose once and stick with it. A reasonable path for many businesses:
- Start with a SaaS tool to validate that the AI use case actually solves the problem users have.
- Once you hit volume, workflow complexity, or data requirements that the tool can't handle, build a custom agent to replace it.
This avoids committing $30K to a problem you haven't yet confirmed is worth solving. The downside is migration effort later, but that's a good problem to have — it means the use case proved out.
The bottom line
Buy if: your workflow is generic, your data isn't proprietary, your seat count is low, and you don't have someone to own post-launch maintenance.
Build if: your workflow is custom, your data is proprietary, your compliance requirements are strict, or per-seat pricing is already painful or will be soon.
The mistake most teams make isn't choosing wrong — it's not running the numbers before they decide. Take 30 minutes, answer the five questions above honestly, and you'll have a defensible answer you can explain to your board.