Build vs Buy: In-House Support Bot vs Custom AI Agent
It's not a two-way choice
The build vs buy support bot decision usually gets framed as "licence a SaaS bot or build our own." There's a third option in the middle that most e-commerce teams end up at: a custom agent built for you, on your stack and your APIs, without starting from a blank repo. Knowing which of the three you need starts with what your tickets actually are.
Option 1: buy a SaaS support bot
The AI features bundled into helpdesk platforms. Fast to switch on, maintained for you, and decent at deflecting informational tickets from your help centre.
What it costs you:
- Generic actions. It can answer from your articles; it usually can't run your return-eligibility rules against this order and issue the refund.
- Your workflows don't quite fit. Escalation, tagging, and routing bend to the platform's model, not yours.
- Pricing that scales with success. Per-resolution or per-seat fees that climb with volume — and spike in peak season, exactly when volume matters most.
- Limited control over money-action safety. You inherit the vendor's guardrails and audit trail, whatever they are.
Option 2: build in-house from scratch
Full control, no licence fees, and a multi-quarter project with a long maintenance tail: the model integration, tool schemas, an evaluation harness, an escalation console, and someone on call for all of it.
Worth it only if support automation is close to core business and you have the engineering capacity to own it indefinitely.
Option 3: a custom agent on your stack
An agent built around your order and returns APIs, your policies, and your existing helpdesk for escalation — but using proven components rather than inventing them. You get real transactional actions and control over the escalation logic without carrying a from-scratch codebase.
On our support agent project this was the shape: an LLM agent with function calling over the client's order and returns APIs, confidence-based escalation into the agents' own console, built on their infrastructure — response time down 60%, roughly two-thirds of conversations auto-resolved, CSAT steady at 4.5/5.
Signals you've outgrown "buy"
- Transactional volume. A real share of tickets are returns, refunds, order changes — actions, not answers — and they need your business rules.
- Peak-season economics. Per-resolution pricing at Black Friday volume is a large, forecastable bill.
- CSAT sensitivity. You need escalation tuned to your standards, not a vendor default.
- You already have the APIs. The expensive prerequisite is done; a custom agent is now mostly integration and policy work.
The middle path in practice
Keep a SaaS bot or help-centre search for pure FAQ deflection, and add a custom agent for the transactional tickets where your data and rules matter. You're not ripping out what works; you're covering the part it can't reach.
A quick way to place yourself
Four questions, roughly in order of weight:
- What share of tickets need an action on your systems (return, refund, order change) rather than an answer? High share pushes toward a custom agent.
- Do you already have APIs for those actions? No — build the APIs first; that's the real project, and until it's done every option is limited to answers.
- Can you staff an escalation queue and review money actions? No — stay with a SaaS bot and keep the agent answer-only.
- Is support automation close to core, with engineering to own it? Only "yes" here justifies building fully in-house instead of a custom agent on proven components.
Where "buy" stays the right answer
- Mostly informational tickets. No transactional volume to justify integration work.
- Low overall volume. Build and maintenance don't pay back against a small queue.
- No capacity to own it. A custom agent you can't staff decays. A maintained SaaS bot doesn't.
- No usable APIs yet. Fix or build those first; that's a separate project and the real blocker.
FAQ
Is a SaaS support bot just a stopgap? No — for informational-heavy support it can be the right permanent answer. Building is a response to transactional volume and control needs.
Can we start with buy and move later? Yes, and it's common. Run the SaaS bot, instrument which tickets it can't resolve, and use that data to scope a custom agent for the gap.
How long does a custom agent take? The model is rarely the long pole — tool integration, schema discipline, and the escalation console are. It scales with how many systems the agent has to touch.
ISTRALLEN builds custom support agents on your stack for teams that have outgrown a generic bot; see AI for E-commerce.