A Berlin fintech migrated to Intercom Fin and migrated back three weeks later. The lesson: chatbots answer, agents act, and most businesses confuse which one they need. A founder-written guide.
In March 2026, a Berlin fintech with 40 employees migrated their support from a 4-person human team to Intercom Fin. Three weeks later they migrated back.
The reason was not that Fin was bad. Fin handled 60% of their tickets correctly — close to Intercom's published resolution numbers. The problem was the bill. At €0.99 per resolution, the 60% Fin was handling cost €4,200 a month. The four humans had been costing €3,800.
They had bought an AI agent. What they had wanted was a chatbot.
The distinction matters more in 2026 than it ever has, because the same vendors who sold "chatbots" in 2023 are now selling "AI agents" — at four to ten times the price. Some businesses need that. Most do not. The fastest way to waste money in customer support this year is to confuse the two.
What a chatbot actually is
A chatbot answers. It takes a question, looks at what it knows, and writes a reply. Modern ones use large language models so the replies sound human. Older ones followed scripted trees — if the user asks A, route to B. The architecture is different. The job is the same. Answer questions.
The economics of a chatbot are flat. You pay for the platform, you train it on your help center, and the cost per conversation is roughly zero. Whether 100 people ask about your refund policy or 100,000 do, the platform cost barely moves. That is why Chatbase can charge $40 a month for unlimited messages on the entry tier and stay profitable. The work was done at index time.
Chatbots are good at FAQ deflection. They are mediocre at anything more.
What an AI agent actually is
An AI agent acts. It can answer a question — but more importantly, it can do the thing the question is about. A chatbot tells the user how to book a meeting. An agent books the meeting. A chatbot tells the user how to update their shipping address. An agent updates it. A chatbot tells the user that filler is €350. An agent qualifies the lead, books the consultation, sends the calendar invite, and writes the lead into the CRM with a score of 7 out of 10.
The internal architecture is different. A chatbot is a single LLM call: input → response. An agent is a loop: input → reason → call a tool → check the result → decide → respond. The "tools" are APIs into the things the agent can do — your calendar, your CRM, your Shopify catalog, your Stripe account.
The economics are not flat. Agents are now mostly priced per resolution: Fin charges around $0.99, Decagon and Sierra charge enterprise rates somewhere north of that. The unit cost scales linearly with use. This is the model the Berlin fintech ran into. The product worked. The pricing did not.
There is a second class of agents — flat-priced ones, including ours — that charge a fixed monthly fee regardless of resolution volume. The unit economics are different because the actions an agent takes (a calendar API call, a CRM write) cost the vendor almost nothing. The marginal cost of an additional booking is fractions of a cent. Per-resolution pricing exists because vendors can charge for it, not because it reflects underlying cost.
Three questions that decide which one you need
I have walked through this with maybe 200 founders in the last year. There are three questions that decide it, and most evaluation processes get derailed because the founder answers the wrong one first.
First: when a customer messages you, is the goal to send information or to complete an action?
If it is information — "what is your refund policy", "where is your office", "how do I cancel" — you need a chatbot. The answers exist in your help center. A good chatbot retrieves and replies.
If it is an action — "book me an appointment", "process this refund", "qualify me as a lead and route me to sales" — you need an agent. No amount of information retrieval will book the appointment. Something has to call the calendar.
Most businesses have both. The mistake is treating them the same and buying the same tool for both.
Second: at your monthly volume, what does an agent actually cost?
This is the question Berlin missed. Per-resolution pricing looks cheap at small volumes and brutal at scale. Run the math at your actual monthly conversation count, not at the tier you are on today. A clinic doing 800 inquiries a month on Fin pays $792 in resolutions before the Intercom Suite seat fees. A Shopify store doing 5,000 conversations a month pays $4,950.
If those numbers are uncomfortable, the flat-priced agent vendors exist for a reason. The capability is the same. The pricing model is the difference.
Third: how many systems do you actually need the agent to act on?
The cleanest test of whether a vendor is selling you a chatbot or a real agent is to ask for a live demo of three actions across three systems. Book a meeting in Google Calendar. Write the lead into HubSpot with a score. Recover an abandoned cart in Shopify. If the demo is three iterations of "well, you could route to a human who would do that", you are looking at a chatbot with marketing copy that says "agent".
What the category is doing in 2026
The naming is converging. By late 2026 most vendors will call their products AI agents regardless of architecture. This is what happened to "AI-powered" in 2023 and "GPT-powered" in 2024. The label is downstream of marketing pressure, not engineering reality.
The pricing is bifurcating. The enterprise end (Sierra, Decagon, Fin at scale) is moving to outcome-based pricing — per resolution, per booked meeting, per recovered cart. The SMB end (Tidio, SimplyBoost, Chatbase) is staying flat-priced because the buyer needs predictability. The middle is uncomfortable. Intercom's per-resolution model is squeezed from both sides: enterprises can negotiate outcome contracts, SMBs flee to flat pricing.
The capability gap is closing. The flat-priced agents now do most of what the per-resolution agents do — calendar actions, CRM writes, Shopify actions, multi-step support flows. The exceptions are the heaviest enterprise workflows: cross-system reconciliation, complex policy reasoning, regulated decisions. If you are not running those, you do not need to pay enterprise prices.
A practical recommendation
If you are reading this trying to decide between a chatbot and an agent for your business:
If your inquiries are mostly informational and your support team is mostly answering the same 30 questions, buy a chatbot. Chatbase, Tidio's free tier, the entry plan on most platforms. Spend $30 to $99 a month, train it on your help center, get on with your life.
If your inquiries are commercial — lead qualification, meeting booking, product recommendations, sales handoff — you need an agent. The question is which pricing model survives at your volume. If you are doing more than 500 commercial conversations a month, flat-priced agents will be cheaper. If you are below that and want to test the category, per-resolution pricing on a low tier is fine.
If your inquiries are a mix, which is what most businesses actually have, you want an agent that handles informational queries without charging you per-resolution for them. SimplyBoost is built for that case — flat monthly pricing, the AI handles whatever comes in, no per-conversation surcharge.
Frequently asked questions
Is an AI agent the same as an AI chatbot?
No. A chatbot answers questions; an AI agent answers questions and also takes actions across your systems — booking meetings, writing CRM records, processing refunds, recommending products. Every AI agent contains a chatbot. Not every chatbot is an agent.
Why are AI agents so much more expensive than chatbots?
Two reasons, one legitimate and one not. The legitimate reason: agents make API calls to external systems, each of which has a small cost. The illegitimate reason: vendors discovered they can charge per resolution and price has decoupled from cost. The marginal cost of an additional resolution to the vendor is fractions of a cent. They charge $0.99 because they can.
When should I use a chatbot versus an AI agent?
Use a chatbot if your customers ask informational questions you already have answers to. Use an agent if they want something done — a meeting booked, a refund processed, a product recommended and added to cart. Most businesses need both, which is why flat-priced agents that handle informational and transactional traffic in one pricing tier are now common.
What is the cheapest AI agent in 2026?
Among flat-priced agents with real action capability, SimplyBoost starts at $39 a month with the calendar and CRM integrations included. Among per-resolution agents, the cheapest practical option is Intercom Fin at roughly $0.99 per resolution if your volume is genuinely low. Tidio's Lyro is in between — flat tier plus Lyro AI add-on for around $39 a month, capped at 50 conversations per month before the next tier.
How do I know if a vendor is selling me a real AI agent or a chatbot with new branding?
Ask for a live demo of three actions across three of your systems. Book a meeting in your calendar. Write a record into your CRM with custom fields. Take one workflow-specific action in your stack — a Shopify product recommendation, a Stripe refund, a Notion page update. If the demo cannot do this without "let me route to a human who will do that", you are looking at a chatbot.
---
A disclosure. I am the founder of SimplyBoost, a flat-priced AI agent for sales and support across web, WhatsApp and Instagram. We built it because we kept watching small businesses pay enterprise prices for things they could have done in a $39 monthly tier. If you want to see what your inbox looks like with an agent running on top of it, there is a no-credit-card trial at get.simplyboost.io.
SimplyBoost is registered in the Netherlands (KVK 87456346). Data hosted in Frankfurt, EU.