The drag-and-drop demo takes ten minutes. Wire up a flow, drop in an API key, embed a widget — you’ve got a working bot before lunch. That’s the whole pitch of a no-code AI agent builder, and for a first prototype it delivers exactly what it promises.
What it doesn’t show you is month nine: an invoice that’s triple what the pricing page implied, and an export button that turns out not to exist. Visual builders sell speed. They rarely explain how the meter runs once you’re past a demo, or what happens to your flows, prompts, and knowledge base the day you want to leave.
This is a segment guide to that market — how the five common billing models actually charge you at volume, what you can and can’t take with you, and when it’s time to graduate off the canvas. Full disclosure up front: we build a self-hosted chat agent that sits at one end of this migration path. That’s not a reason to distrust the pricing math below — it’s just why you should check the vendor numbers yourself before you commit.
What Counts as a No-Code AI Agent Builder
”No-code AI agent builder” gets used as one category on review sites. It isn’t one. Three distinct product shapes get lumped under that label, and each one meters, breaks, and locks you in differently.
Visual flow builders are the classic no code chatbot builder: a canvas, drag-and-drop nodes, branching logic drawn as boxes and arrows. Voiceflow, Landbot, and Botpress anchor this category — see our head-to-head breakdowns on how Voiceflow compares and where Botpress differs. These tools are built for conversation design, not general automation. They’re the closest thing to a purpose-built visual chatbot builder on the market.
Agent and workflow orchestrators — n8n, Zapier Agents, Stack AI, Relevance AI, Lindy — are platforms to build AI agents that chain steps across many services: pull a record, call a model, write to a CRM, send a Slack message. The bot builder framing is almost incidental; these are automation platforms that happen to include an LLM step.
Hosted RAG bots — Chatbase, Dify, Flowise — are an AI agent creation platform narrowed to one job: ingest documents, answer questions against them, embed a widget. Less flexible than an orchestrator, faster to stand up than a custom RAG pipeline.
Why the taxonomy matters: a flow builder’s failure mode is a canvas too tangled to read. An orchestrator’s failure mode is a workflow cap you hit from background automation, not customer traffic. A hosted RAG bot’s failure mode is a knowledge base you can’t tune once answer quality plateaus. Pricing pages don’t sort by this. Your renewal invoice will. For a broader ranking across every category of agent tool, not just no-code, see our roundup of the best AI agent tools, or browse the full getagent.chat blog archive.
How No-Code AI Agent Builders Actually Bill You
The sticker price on a pricing page is close to meaningless. What matters is the meter behind it — the one unit of activity that increments your bill. Five metering models cover almost the entire market, and they respond to growth in completely different ways.
The number of teams adopting agentic AI is climbing fast enough that this stops being hypothetical. PwC’s September 2025 survey found 42% of enterprises expect enterprise-wide agentic AI adoption within 6-12 months, against 17% that had already deployed it company-wide. A lot of new subscriptions are about to meet these meters for the first time.
| Metering model | What fires the meter | What makes the bill spike |
|---|---|---|
| Per-seat | A licensed editor or admin logs in | Headcount growth, not usage — you pay more as your team grows, even with flat traffic |
| Per-credit | Each LLM call, weighted by which model answered | Switching to a better model multiplies the credit cost of the exact same conversation |
| Per-conversation | A session with two or more end-user messages | High-volume, low-value chit-chat and bot testing count the same as a resolved support case |
| Per-workflow-execution | One full automation run, regardless of step count | Background jobs — polling, retries, scheduled syncs — burn executions with zero customer traffic behind them |
| Per-task / per-action | One atomic step a workflow or agent performs | Multi-step automations get billed per step, so complexity is expensive even at flat conversation volume |
Notice what none of these meters track directly: value delivered. A per-seat model penalizes team growth. A per-credit model penalizes quality upgrades. A per-execution model treats a two-step flow and a twelve-branch flow identically, but penalizes anything that fires in the background. Before you sign, ask one question the sales page won’t answer for you: what specific action, in your actual bot, increments this meter — and how many of those does a normal week generate?
What the Major Platforms Charge in 2026
Numbers below come straight off each vendor’s public pricing page as of this writing. Pricing on these platforms changes often — treat this as a snapshot, and check the vendor page before you budget against it.
| Platform | Billing model | Entry price | Self-hosting |
|---|---|---|---|
| Chatbase | Per-message-credit | Hobby $32/mo (700 credits); Standard $120/mo (4,000 credits); Pro (15,000 credits). Auto-recharge $40 per extra 1,000 credits | No |
| n8n | Per-workflow-execution | Cloud Starter €20/mo (2,500 executions); Pro €50/mo (10,000 executions) | Community Edition free; self-hosted Business €667/mo billed annually |
| Lindy | Per-task, plus add-ons | Plus $49.99/mo; Pro $99.99/mo. Tasks roughly $0.01–$0.10 each; each phone number +$10/mo | No |
| Relevance AI | Per-action + vendor credit pass-through | Pro $19/mo annual (2,500 actions + $20 vendor credits); Team $234/mo annual (7,000 actions + $70 credits). Overage $80 per 1,000 actions | No |
| Dify | Per-message-credit | Professional $59/mo; Team $159/mo. Credits don’t roll over; bring-your-own-key on paid tiers | Community Edition free, open source |
| Zapier | Per-task | Professional $29.99/mo (750 tasks); Team $103.50/mo (2,000 tasks). AI Agents is a separate paid add-on, price not published | No |
| Voiceflow | Hybrid credit + seat | Pro around $60/mo (~10,000 credits); Business around $150/mo (~30,000 credits), extra seats ~$50/mo each | No |
| Botpress | Per-conversation (since May 2026) | Reportedly a pay-as-you-go entry with a small included AI spend and a few hundred incoming messages; post-change tier detail is thin publicly | Yes — open-source core |
| Landbot | Per-chat | Starter around $46/mo (~500 chats, ~100 AI); Pro around $105/mo | No |
| Flowise | Per-prediction | Starter around $35/mo (10,000 predictions); Pro around $65/mo (50,000 predictions) | Community Edition free |
| Stack AI | Per-run | Free tier, then custom enterprise quotes (consolidated mid-2026) | Enterprise only |
Read the pattern, not just the row. Every self-serve entry tier looks affordable in isolation. The gap between “affordable” and “what you’ll actually pay” is entirely a function of which meter you’re on and how your bot’s traffic maps to it — which is the next section.
The Scale Math: What 10,000 Conversations Costs
Take a mid-size support bot doing 10,000 conversations a month — a realistic number for a company with a few hundred thousand monthly site visitors. Here’s what four different metering models do with that same traffic.
Credit-based (Chatbase-style): assume roughly three AI responses per conversation on average — some are one-line FAQ answers, some run five or six turns. That’s about 30,000 credit-consuming responses. At the cheapest model tier (1 credit per response), you’re already at 30,000 credits — more than seven times the Standard plan’s 4,000-credit allotment, and double the Pro tier’s 15,000. The overage alone, in $40-per-1,000-credit blocks, adds up fast before the base subscription is even counted.
Now apply the credit-multiplier trap. Swap the cheap model for a Claude Sonnet-class model because answer quality matters for support — that’s 3 credits per response instead of 1. Same 10,000 conversations, same 30,000 responses, but now 90,000 credits. Conversation volume didn’t move. The bill tripled because the model got better.
Execution-based (n8n-style): if the bot runs as one workflow per conversation, 10,000 executions sits right at n8n’s Pro cap. Add the background jobs a real support bot needs — scheduled KB re-syncs, retry logic, webhook relays — and you’re over the cap from automation traffic alone, with no overage billing and a hard upgrade wall.
Task-based (Zapier-style): a support flow with intake, CRM lookup, and response formatting as separate steps turns each conversation into roughly three tasks — 30,000 tasks against a 2,000-task Team-tier cap.
The constant across all four: the underlying LLM API call costs about the same wherever it runs. A GPT-4o-mini or Claude Haiku token bill doesn’t change because you called it from n8n instead of your own server. What changes is the platform’s toll on top of that call — and at 10,000 conversations, the toll is usually the bigger number. For the model-cost side of this math in more detail, see our deeper pricing breakdown.
The Five Layers of Lock-In
”Vendor lock-in” undersells what’s happening. It’s not one thing you’re locked into — it’s five, and they compound.
Model lock-in. You can only use the models the platform has integrated and priced into its credit table. Wanting a different model isn’t a dropdown change — on credit-metered platforms it’s a different multiplier on every future bill, and on locked-routing platforms it isn’t possible at all. We cover this trade-off in more depth in our piece on running multiple LLM providers.
Orchestration lock-in. Your flow logic lives inside the platform’s proprietary canvas format. No git diff, no code review, no branch-and-merge. A junior teammate can’t open a pull request against a drag-and-drop flow — they can only screenshot it and describe the change in Slack.
Data lock-in. Conversation transcripts, lead records, and analytics sit in the vendor’s database. Where export exists at all, it’s usually a CSV dump, not a portable schema you can load somewhere else without rebuilding the mapping by hand.
Knowledge lock-in. Chunking strategy, embedding model, reranking — the whole retrieval pipeline behind a hosted RAG bot is a black box. If answer quality plateaus, you have no knobs to turn. See our notes on what actually drives retrieval quality for what those knobs look like when you do have access to them. Chatbase is the sharpest example here — no documented export, no BYOK, no retrieval tuning; our Chatbase comparison goes through the specifics.
Governance lock-in. Audit logs, access roles, data residency, incident response — you inherit whatever the vendor built, on whatever timeline they built it. If your buyer asks where the data physically sits, the honest answer is “wherever the vendor’s infrastructure is,” not “wherever we decide.”
Can You Export Your Agent?
This is the question that never comes up in a sales demo, and the one that matters most the day you want to leave. Two things determine how painful an exit is: whether your flow logic exports in a usable format, and whether you can point the platform at your own LLM key instead of its routing.
| Platform | Flow export |
|---|---|
| n8n, Dify | Clean JSON, importable elsewhere |
| Botpress | Exportable, developer-friendly (open-source core) |
| Voiceflow, Flowise, Stack AI | Partial / API-only / non-standard |
| Chatbase, Landbot, Relevance AI, Lindy, Zapier | No documented export |
| Platform | Bring-your-own LLM key |
|---|---|
| n8n, Flowise, Dify, Botpress | Full support |
| Voiceflow, Relevance AI | Partial, or paid-tier only |
| Chatbase, Zapier Agents, Lindy, Landbot, Stack AI (non-enterprise) | Locked to vendor routing, or undocumented |
The pattern is not subtle. The platforms with roots in open-source or developer tooling — n8n, Dify, Flowise, Botpress — treat portability as a feature. The platforms built as closed hosted products treat it as an afterthought at best. Neither approach is wrong for what those products are optimizing for. But if you’re choosing a no-code AI agent builder for anything beyond a six-month pilot, this table should weigh as much as the pricing table did.
One caveat: this reflects publicly documented behavior at the time of writing. Vendors change export and BYOK support without much announcement — verify against current docs before you build a migration plan around any single row here.
Signs You’ve Outgrown the Visual Canvas
None of these signs mean the tool failed. They mean you’ve grown past the job it was built for.
- Conditional logic sprawling past the canvas. When a flow needs a wiki page to explain, and scrolling replaces reading, the visual metaphor is costing you more than it’s saving.
- You need a model the platform doesn’t route to. A newly released or fine-tuned model shows up, and there’s no path to use it without waiting on the vendor’s integration roadmap.
- Per-seat cost is rising with headcount, not usage. You’re paying more every time you hire, regardless of whether conversation volume moved at all.
- Compliance or data-residency questions show up in an RFP. A prospect asks where the data lives and who can access it, and the honest answer routes through a vendor’s infrastructure you don’t control.
- You want to version flows in git. Code review, rollback, and branch-and-merge are how your engineering org already works — the canvas can’t participate in that process.
- Retrieval quality has plateaued. Answers are stale or shallow and you can’t touch chunking, embedding choice, or reranking to fix it — you can only rewrite the source documents and hope.
- You’re treating the monthly upgrade prompt as routine. Hitting the execution, task, or conversation cap every cycle and clicking “upgrade” without a second thought is a sign the meter, not your usage, is now driving the roadmap.
Any one of these is a nudge. Two or three together are a business case for moving off the canvas — not necessarily urgently, but deliberately, on your own timeline instead of the vendor’s.
The Migration Path Off a No-Code AI Agent Builder
Some of what you built carries over cleanly. Some of it doesn’t, and pretending otherwise is how migrations blow their timeline.
What carries over: your prompts (they’re just text), the source documents behind your knowledge base (PDFs, docs, URLs — the raw material, not the vendor’s processed index of it), your conversation transcripts (useful as an evaluation set for the new system), and your intent list (the catalogue of what users actually ask, which took real effort to build and has nothing platform-specific about it).
What you rebuild: the flow logic itself — treat the visual canvas as a spec you’re reading, not code you’re porting, since embedding formats and canvas structures aren’t portable between vendors. The retrieval configuration — chunking, embeddings, and reranking need to be rebuilt in the new system rather than migrated, because no two platforms chunk or embed the same way. And every integration, one at a time.
A realistic sequence: export everything exportable now, while the subscription is still active — several of the platforms in the table above have no export tooling, so this step has a shelf life. Re-ingest your source documents into the new retrieval pipeline and re-tune chunking against your own content. Rebuild the highest-traffic flow first, not the most complex one, so you get a real quality comparison fast. Reconnect integrations by usage volume. Run both systems in parallel against a held-out sample of real transcripts until the new one matches or beats the old one’s answer quality. Then cut over gradually — a percentage of traffic at a time, not a hard switch.
None of this is a weekend project, and no vendor’s migration guide will tell you that honestly. Budget real engineering time, not a sprint. If you’re building the replacement yourself rather than moving to another hosted product, our guide to building your own bot covers the DIY side of that decision in detail.
When No-Code Is Still the Right Call
This isn’t a case against no-code AI agent builders. Most teams that reach for one made the right call, and plenty should never leave.
Prototyping. Validating whether an AI agent solves a real problem before you commit engineering time is exactly what a visual canvas is for. Ten minutes to a working demo is a genuine advantage, not a gimmick.
Non-technical teams. If nobody on the team writes code and nobody’s hiring an engineer for this, a no code chatbot builder is the only realistic path to shipping anything at all.
Sub-1,000 conversations a month. At that volume, almost every metering model in this piece stays inside its entry tier. The lock-in risk is real but the financial risk isn’t — you’re not exposed to the scale math from the section above.
Campaign bots with a shelf life. A promotion, an event, a seasonal launch — something that runs for eight weeks and gets retired doesn’t need portability. Nobody migrates a bot they’re about to delete.
No ops capacity. Self-hosting anything means someone owns uptime, patching, and backups. If that capacity doesn’t exist on your team, a managed platform’s operational overhead is worth paying for, meter and all.
Industry surveys suggest SMB chatbot adoption plans are around 64% for 2026, up from roughly 38% in 2024 — most of that growth is teams in exactly these categories, and a no-code AI agent builder is the correct default for them, not a compromise.
The Self-Hosted Endpoint
For teams that have actually hit the ceiling described above — not hypothetically, but in a real invoice or a real RFP question — a self-hosted option is one answer, not the answer. It trades a monthly meter for infrastructure you own and operate yourself, which is a real trade, not a free upgrade.
We build one of these: AI Chat Agent, currently at v1.8.1. It’s a one-time €79 purchase — no monthly fee, full source code, lifetime updates — deployed via Docker Compose across five services (Node server, React admin, PostgreSQL with pgvector, Redis, Nginx). That structure directly answers three of the lock-in layers above.
Model lock-in: it connects to five providers — OpenAI, Anthropic Claude, Google Gemini, OpenRouter, or any OpenAI-compatible endpoint including Groq, Ollama, or a self-hosted model — and you switch providers in the admin panel without migrating data. Knowledge lock-in: retrieval is hybrid dense-and-lexical search with an LLM reranker and configurable chunking (512 tokens, 50-token overlap, markdown-aware), so the pipeline is something you can actually tune rather than a black box. Governance lock-in: self-hosting puts the data in infrastructure you already control and audit, a different posture than trusting a vendor’s compliance page.
It’s tested with 1,622 automated tests, encrypts stored provider keys with AES-256-GCM, and ships a widget under 26KB gzipped with Shadow DOM isolation. It also handles the things a support bot needs day to day — operator live takeover mid-conversation, lead capture with alerts to email, Telegram, or webhook, and multi-bot management with isolated data per bot.
None of this makes sense for a prototype or a campaign bot — see the section above. It’s built for the team that’s already done the scale math and the export math, and landed on “we’d rather own this.” For a fuller comparison of the two operating models, see self-hosted versus SaaS chatbots.
Frequently Asked Questions
What is a no-code AI agent builder?
A no-code AI agent builder lets you assemble a working chatbot or agent on a drag-and-drop canvas instead of writing code, usually with an embeddable widget at the end of it. The label covers three different product shapes: visual flow builders like Voiceflow, Landbot and Botpress; agent and workflow orchestrators like n8n, Zapier Agents and Lindy; and hosted RAG bots like Chatbase, Dify and Flowise. Each one meters usage and locks you in differently, so the category label matters less than which shape you picked.
How much does a no-code AI agent builder cost?
Self-serve entry tiers mostly land between €20 and $60 a month — n8n Cloud Starter at €20, Chatbase Hobby at $32, Landbot around $46, Dify Professional at $59, Voiceflow around $60. The sticker price is the smaller half of the answer: what you actually pay depends on the meter behind it, and overage is where bills spike, at $40 per extra 1,000 credits on Chatbase or $80 per 1,000 actions on Relevance AI.
Can you export your chatbot flows if you switch platforms?
It depends entirely on the vendor, and it is rarely mentioned in a demo. n8n and Dify export clean JSON you can import elsewhere, Botpress is exportable off its open-source core, while Voiceflow, Flowise and Stack AI offer partial, API-only or non-standard export — and Chatbase, Landbot, Relevance AI, Lindy and Zapier have no documented export at all. Export everything exportable while the subscription is still active, because that option disappears the day you cancel.
What is the difference between no-code, low-code, and self-hosted AI agents?
No-code means the vendor’s canvas, the vendor’s meter and the vendor’s infrastructure, with flow logic living in a proprietary format you can’t diff or code-review. The low-code middle ground is the open-core platforms — n8n, Dify, Flowise, Botpress — which still give you a visual builder but add JSON export and bring-your-own-key support. Self-hosted means you run the stack yourself, trading a monthly meter for infrastructure you own, operate and patch.
When should you stop using a no-code builder?
The signals are concrete: conditional logic that needs a wiki page to explain, a model the platform won’t route to, per-seat cost rising with headcount instead of usage, data-residency questions in an RFP, wanting flows in git, retrieval quality that has plateaued, or clicking the monthly upgrade prompt as routine. Any one of those is a nudge rather than an emergency. Two or three together are a business case for moving off the canvas on your own timeline instead of the vendor’s.
Can you use your own LLM API key with a no-code builder?
n8n, Flowise, Dify and Botpress support bring-your-own-key fully; Voiceflow and Relevance AI support it partially or on paid tiers only; Chatbase, Zapier Agents, Lindy, Landbot and non-enterprise Stack AI are locked to vendor routing or leave it undocumented. This matters more than it sounds, because on credit-metered platforms the model you choose is a bill multiplier — swapping a cheap model for a Claude Sonnet-class one at 3 credits per response triples your credit burn on identical conversation volume.
If you’re still on the visual canvas, that’s probably the right place to be — go build the thing, ship the pilot, revisit this once the invoice or the export question actually arrives. If you’ve already read that invoice twice this quarter, take a look at the live AI Chat Agent demo and see whether the self-hosted model fits your stack. When you’re ready, get AI Chat Agent for a one-time €79 and stop budgeting around a meter.