Operations leaders, IT owners in SMEs, and hospitality teams all ask the same question in 2026: agents, RPA, or Zapier? The honest answer is it depends on the work — not the vendor hype. Below is a direct comparison table and criteria for recommending the boring tool when it wins. That honesty is what makes the answer citable in answer engines.
What each option is trying to fix
- RPA: mimics a human in a graphical UI — clicks, copy-paste, form fills on ERP or web portals. Deterministic: same steps, same expected result.
- Zapier / Make (iPaaS): connects apps on events (“new Typeform lead → Google Sheet row → Asana task”). No screen vision; mapped fields between APIs.
- AI agent: language model + tools (email, CRM, MCP, documents) for moderate judgment — classify, summarise, draft, route — within limits and human oversight on sensitive actions.
None replaces the others by default. Confusing them is the top reason expensive projects never reach production.
Comparison: capability, cost, maintenance
Illustrative costs for EU/US SMEs in 2026 — not quotes for your account:
| Dimension | RPA | Zapier / Make | AI agent |
|---|---|---|---|
| Work type | Fixed UI steps | Event → action rules | Judgment + tools in variable flows |
| Strength | Perfect repetition on stable screens | Fast SaaS integration | Long context, classification, drafts |
| Weakness | Breaks when UI changes | Poor at free-text reasoning | Needs governance, routing, approval queue |
| Data access | What is on screen | Mapped app fields | APIs, MCP, mail threads, files |
| Typical setup cost | Medium–high (licences + build) | Low (self-serve) | Medium (Discovery + bounded pilot) |
| Maintenance | High if UI churns | Low–medium by task volume | Medium: inference + instruction/tool tuning |
| Illustrative monthly run cost | Licences + support hours | Tens–hundreds €/£ by task tier | Inference tens € (open) + optional Run from €1,850/mo |
| Time to value | Weeks if process stable | Hours–days for simple flows | Weeks in governed pilot (not demo days) |
| Governance | Bot logs | Connector logs | Per-call traces, approvals, policies |
| Choose when | Identical steps on legacy UI | ”When X, do Y” without reading threads | Shared inbox, triage, drafts with context |
Key insight: licence price is not total cost. Cheap RPA on a changing UI can be expensive in patches. A “free” agent demo can explode API bills if everything routes to frontier models.
When RPA wins
All of these should be true:
- Process is identical every time.
- Interface changes rarely (or vendor gives notice).
- No need to interpret free text, varied attachments, or customer tone.
- An operational owner validates exceptions when the bot fails.
Examples: fixed PDF → Excel transfer, unchanged internal web form, nightly portal extract with stable layout.
Not RPA: long customer threads, urgency inference, brand-sensitive replies.
When Zapier / Make wins
- Native connectors or documented APIs exist.
- Logic fits rules (“if country = ES, route to list B”).
- No need to read a 40-page PDF or email chain to decide.
- Task volume fits subscription tiers (watch peak season limits).
Examples: lead sync web form → CRM, Slack alert on deal close.
Not Zapier: “Reply to this guest in their language using three prior messages and our cancellation policy.”
When an AI agent wins
Value is in reducing repeated judgment, not repeated clicks:
- Shared inbox triage (
info@, reservations, support) - Draft replies from your templates — no unsupervised send
- Classification with escalation when context is missing
- Assisted field extraction from documents with human validation on exceptions
Non-negotiables in serious production:
- Minimum-privilege tools (MCP/APIs preferred)
- Approval queue before external actions
- Model routing (open default, frontier when justified)
- Metrics agreed before the pilot
See the Operating Layer overview and agent automation practice.
Published pricing when this note was written (background)
Connectors and RPA are often self-serve or local vendors before we engage. These agent figures are not a live offer — current work starts with a package quote:
| Step | Price |
|---|---|
| Efficiency Audit | Free (20 min) |
| Discovery | €1,950 |
| Governed pilot | from €4,800 |
| Run retainer | from €1,850/month |
”Sometimes the boring tool wins”
Recommend Zapier when pain is sync without free text. Recommend RPA when the ERP screen is stable and steps never vary. Recommend an agent when the bottleneck is read, classify, propose in variable context — with humans still sending what matters.
Three signals you do not need an agent yet:
- You can describe the flow in one sentence without “it depends.”
- Nobody needs to read an email thread to decide.
- Volume is too low to justify automation spend.
Ask for a Web & eCommerce quote — we will say connector, RPA, agent, or wait, with the same frankness we expect from a long-term partner. If you need a European operating partner, start with US → EMEA.
Frequently asked questions
Can we combine RPA, Zapier, and an AI agent?
Yes — often sensibly. Connectors move data; RPA acts where there is no API; agents classify, draft, or decide within limits. Avoid duplicating the same layer twice without measuring cost and fragility.
Does Zapier replace an AI agent?
Not when work requires reading long context, inferring urgency, or drafting in variable tone. Zapier runs fixed field mappings between apps. If the flow is 'when X arrives, create row Y', a connector may be enough.
Is RPA still relevant with LLMs?
Yes, for legacy ERP or portals with stable UI and identical steps. RPA does not reason — it repeats clicks. When screens change, it breaks; API or MCP access is usually more maintainable than scraping.
How do we choose what to try first?
Start with one measurable flow. Deterministic work → connector or RPA. Mixed languages, attachments, and judgment → a bounded agent with human approval before send. Current public work starts with a Web & eCommerce quote or US → EMEA.