The decision isn't "should we add AI chat to our SaaS product?" Most products will. The decision is "should we build it ourselves or drop in Intercom Fin or Zendesk AI?"
That decision has different answers at different stages of your company, and most teams get it wrong in predictable ways.
What you actually get from off-shelf tools
Three vendors dominate this space for SaaS products:
Intercom Fin. Intercom's AI agent runs on GPT-4 and can be configured to handle support queries using your help center content. Setup is fast — if you already use Intercom, you can have Fin live in a few hours. It handles conversation handoff to human agents gracefully, has decent analytics, and the brand experience is polished. Pricing: Fin charges $0.99 per resolved conversation on top of your existing Intercom subscription. At 1,000 resolutions/month, that's $990/month on top of your base plan.
Zendesk AI. Similar positioning to Fin — answer bots powered by your help center articles, routing to agents, ticket deflection. Zendesk AI is bundled into higher-tier Zendesk plans or available as an add-on. Their "Agent Copilot" features add AI assistance for human agents. Pricing varies significantly by plan; AI features typically add $50–$100 per agent per month.
Freshdesk Freddy AI. Freshdesk's AI layer offers answer bots, auto-tagging, and sentiment analysis. More affordable than Zendesk at lower seat counts. Freddy Self Service starts at around $29/month for basic bot functionality.
What all three share:
- Your conversation data is processed on their infrastructure and may be used to improve their platform's models
- Customization is limited to what their configuration panel exposes
- You cannot build custom workflows that go beyond their integration layer
- Brand experience is constrained by their UI framework
- Per-conversation or per-seat pricing that scales linearly with your usage
The "your data trains their model" point deserves particular attention. Intercom's and Zendesk's terms of service give them rights to use interaction data for product improvement. For most SaaS companies this is an acceptable tradeoff. For companies in healthcare, finance, or any domain handling sensitive data, it may not be.
What you get from building custom
Full control over the stack. You own the entire pipeline: retrieval, prompt design, model selection, response formatting, and the UI. If you want the chatbot to call your internal APIs, check a user's subscription status before answering, and route differently based on their plan tier — you can do all of that.
Your data stays yours. Customer conversations, support context, and product usage data flow through your own infrastructure. Nothing is shared with a third-party platform for any purpose.
Custom tone and persona. Your chatbot can sound exactly like your product. If your brand is direct and technical, it's direct and technical. If your documentation uses specific terminology your users expect, the agent uses it. This is not achievable with off-shelf tools beyond surface-level configuration.
No per-conversation pricing. Once built, your ongoing cost is the LLM API usage — which scales with token count and volume, not per-resolution fees. At high volume, this is significantly cheaper than per-conversation pricing.
Deep product integration. A custom AI agent can look up user account details, check subscription status, pull order history, read from your product database, and take actions — not just answer text questions. Off-shelf tools can connect to APIs via integrations, but the depth and control is limited.
Cost comparison at three scales
Let's model a SaaS support chatbot that handles ~60% deflection rate (meaning 60% of queries are resolved by the AI without human intervention).
Assumptions:
- 500 total monthly support queries per 100 users
- Average 2,000 tokens per conversation (input + output)
- Build cost for custom: $30,000 (production MVP)
- Custom ongoing: $400/month (GPT-5 mini API + infrastructure)
100 users
| Option | Monthly queries | Monthly cost | Annual cost | |---|---|---|---| | Intercom Fin | 500 | ~$300 Fin + $300 base plan = ~$600 | ~$7,200 | | Zendesk AI | 500 | ~$300 AI add-on + $250 base = ~$550 | ~$6,600 | | Custom (amortized over 3 years) | 500 | $1,233 build + $400 ops = ~$1,633 | ~$19,600 |
At 100 users: off-shelf wins, clearly. The custom build cost amortized over three years is 2–3x the off-shelf cost per year.
1,000 users
| Option | Monthly queries | Monthly cost | Annual cost | |---|---|---|---| | Intercom Fin | 5,000 | ~$3,000 Fin + $400 base = ~$3,400 | ~$40,800 | | Zendesk AI | 5,000 | ~$2,000 AI + $500 base = ~$2,500 | ~$30,000 | | Custom (amortized over 3 years) | 5,000 | $1,000 build + $1,200 ops = ~$2,200 | ~$26,400 |
At 1,000 users: custom is competitive and often wins. The $30K build cost is now amortizing to $833/month over 3 years — and total cost is below or comparable to off-shelf options, while giving you more control.
10,000 users
| Option | Monthly queries | Monthly cost | Annual cost | |---|---|---|---| | Intercom Fin | 50,000 | ~$29,700 Fin + $500 base = ~$30,200 | ~$362,400 | | Zendesk AI | 50,000 | ~$15,000 AI + $1,500 base = ~$16,500 | ~$198,000 | | Custom (amortized over 3 years) | 50,000 | $833 build + $4,000 ops = ~$4,833 | ~$58,000 |
At 10,000 users: custom wins decisively. Per-conversation pricing at this volume is punishing. A custom build at $30K—even with $4,000/month in API costs at scale—is 3–6x cheaper than off-shelf options per year.
Note: these are illustrative estimates. Your actual numbers depend on your specific pricing tier, negotiated contracts with vendors, and query volume. Intercom offers volume discounts at enterprise scale, which narrows the gap somewhat but doesn't eliminate it.
The limitations of per-seat and per-conversation pricing
Off-shelf AI chatbot pricing is designed for a world where your support volume scales with customer count in a predictable way. But as your product grows more complex and customers use it more heavily, two things happen:
-
Volume grows faster than seats. A 10x growth in users often produces more than 10x growth in support conversations, especially during growth phases when your user base is less experienced.
-
Pricing tiers don't anticipate usage spikes. A viral growth moment, a bug that triggers a wave of support tickets, or a feature launch that generates confusion can produce billing surprises at $0.99/resolution.
Custom systems aren't immune to cost surprises (API costs can spike with volume), but you control the infrastructure and can set rate limits, implement caching, and optimize aggressively in ways you cannot with a vendor platform.
When to delay building custom
The cost comparison above shows custom winning at 1,000+ users, but there's a caveat that the numbers don't capture: you need enough conversation volume to build reliable evaluations before you build custom.
Evaluations — test cases that tell you whether your chatbot is performing correctly — require real user questions. If you have fewer than 500 unique conversation examples, your eval set is too thin to be trustworthy. You'll ship a system that tests well on your 50 hand-crafted examples and fails on real traffic.
Before building custom:
- Have at least 3–6 months of real support conversation logs
- Understand which question categories make up 80% of your volume
- Be able to write acceptance criteria for what a "good" answer looks like for your top 20 question types
If you're pre-1,000 users and don't yet have this conversation data, start with an off-shelf tool. Use it. Log everything. Build the pattern recognition that will inform your custom build's design. Then, when you hit the volume thresholds that make the economics of custom favorable, you'll have both the business case and the training data to do it right.
The bottom line
Buy (Intercom Fin, Zendesk AI, etc.) if: you're under 500–1,000 users, you don't have 6+ months of conversation logs, your workflow is standard support Q&A, and per-conversation pricing isn't yet painful.
Build custom if: you're approaching or past 1,000 active users, per-conversation costs are compounding, you need deep product integration, your data handling requirements don't fit third-party terms, or your brand experience requires control off-shelf tools can't provide.
Hybrid path: use off-shelf to collect data and validate the use case, then build custom once the volume and patterns justify it. The transition is more work than staying on one platform, but it avoids either paying too much for off-shelf at scale or under-building because you started too early.