Most chatbot platforms still demand 2-8 weeks of setup, force you to pick between GPT-4 and Claude, then bill you twice. Here is why that is broken.
A founder I spoke with last month — runs a 40-person SaaS in Utrecht, sells project management software to mid-market construction firms — sent me a screenshot at 11 PM on a Tuesday. It was the Intercom Fin configuration screen. He had been on it for three hours. The specific question that broke him: should he route billing questions to GPT-4o or Claude Sonnet 4.5, and how would the choice affect his monthly token bill?
He had started the Intercom evaluation in early October. By the time he sent me that screenshot, it was the second week of November. Six weeks in. He had filled out a 47-field intent taxonomy, sat through two onboarding calls with a Customer Success Manager named Priya, written 31 sample conversations, configured escalation rules for nine support tiers, and connected exactly zero of his actual tools because his Stripe webhook required a custom OAuth flow that Intercom's solutions engineer was "happy to scope for an additional engagement."
His competitor — a smaller firm in Rotterdam — had launched their AI support agent over a long weekend in August using a different platform. It was answering 68% of their tickets by the time my founder friend was still picking which Large Language Model to assign to refund requests.
This is the chatbot setup tax. It is the gap between what AI support actually requires (a knowledge source, a way to talk to your tools, and a model good enough to reason about both) and what enterprise platforms charge you in time, decisions, and parallel invoices. Most platforms in 2026 still treat the customer like a systems integrator. They should not.
Why enterprise deploys still take 2-8 weeks
Intercom Fin's typical implementation runs 4-6 weeks for SMB and 8-12 weeks for mid-market, according to their own published case studies. Zendesk AI Agents — the platform formerly known as Ultimate.ai — quotes 6-8 weeks in their standard SOW. Ada quotes 6 weeks minimum. Sierra, the Bret Taylor venture that raised at a $4.5B valuation in October, does white-glove implementations measured in months, not weeks.
The reason is not that the technology is hard. The reason is that these platforms were architected before GPT-4 was good enough to reason without a flow tree. So the setup process is essentially a flow tree authoring tool wearing an LLM costume. You still build intents ("customer wants to cancel"), still write training utterances ("how do I cancel," "cancel my plan," "stop my subscription"), still configure handoff rules, still map the bot's persona variables, still wire up the integrations one webhook at a time.
A single Intercom Fin deployment I reviewed last quarter had 412 distinct configuration objects before launch. Four hundred and twelve. Each one a decision the customer had to make, document, and maintain.
The mid-tier platforms ship faster. Chatbase claims "5 minutes" but the honest number is closer to 3-4 hours once you actually want it to do something useful — escalate properly, handle multilingual customers, not hallucinate on pricing. Tidio's setup is genuinely fast for the FAQ layer, but the AI agent (Lyro) requires separate training and a different billing tier. ManyChat is fast because it is still mostly a flow builder; the AI bolt-on is a 2024-era retrofit.
The "AI-native" platforms tell a better story but mostly ship the same UX. Paste a URL. Configure a persona. Set escalation rules. Pick a model. Configure tools. Add brand voice guidelines. Set fallback behavior. By screen six you have forgotten which platform you are configuring.
The model selection trap
Here is the part of the industry that genuinely upsets me. Most chatbot platforms in 2026 ask the customer to choose between GPT-4o, Claude Sonnet 4.5, and Gemini 2.5 Pro during setup. Some let you toggle between them. A few let you route different conversation types to different models.
The customer almost never has the information to make that choice well.
The honest version of the trade-offs: GPT-4o is fast and cheap (around $2.50 per million input tokens, $10 per million output) and very good at structured outputs and tool calling. Claude Sonnet 4.5 (released September 2025, still the current Sonnet as of mid-2026) costs $3 per million input and $15 per million output, and is notably better at multi-step reasoning, long-context recall over a knowledge base, and refusing to make things up when uncertain. Gemini 2.5 Pro at $1.25 per million input is the cheapest of the three and ships with Google's grounding on web search, but its tool-calling reliability in production agent loops still lags Claude as of Q1 2026.
None of this should be the customer's problem. A founder selling project management software to construction firms should not have to know that Claude is better at refund disputes because it handles ambiguity more gracefully, while GPT-4o is better at the lookup-then-answer flow because its function calling round-trips faster. That is the platform's job.
Worse, the routing problem changes per conversation. The right model for "what is your refund policy" is the cheap one. The right model for "I was charged twice in March, but only for the seats I deactivated, and I think your renewal logic is wrong" is the expensive one that can actually reason about it. A platform that forces a single model choice at setup time is either overpaying on simple queries or underperforming on hard ones. Usually both.
The two-invoice problem
The billing structure is the second tax. Most platforms charge a platform fee and then pass through LLM costs separately. You sign up for Intercom Fin at $0.99 per resolution. You also sign up for a Zendesk AI Agents seat at $50-150/month. Then your OpenAI bill arrives, because the token spend lives in your own OpenAI account that you connected during onboarding.
The predictability problem is real. A founder I know hit an unexpected $2,400 OpenAI bill in a month where one of his enterprise customers ran an internal training exercise that involved every employee testing the chatbot for an hour. None of that traffic was anticipated in the original token-cost projection. The platform did not warn him. The platform did not care. The platform was on a flat fee.
The pass-through model exists because it protects the platform's margins. If usage spikes, the customer absorbs it. If a model gets more expensive, the customer absorbs it. If the platform picks a more expensive model to improve quality, the customer absorbs it. None of the platform's incentives are aligned with the customer's cost predictability.
The right structure is the opposite: the platform absorbs token costs, picks the model per conversation based on what the conversation actually needs, and bills the customer one flat number that does not move with usage. That requires the platform to be genuinely confident in its routing — to know which conversations are cheap and which are expensive, and to manage the blend. It is harder to build. It is correct.
What "ship in 5 minutes" actually requires
The honest five-minute setup looks like this. You drop a URL. The platform crawls your site and your help center, builds a vector index, identifies your products, your pricing, your refund policy, your contact information. You connect one tool — usually Shopify, Stripe, or your CRM — through a one-click OAuth, not a webhook configuration screen. You paste a snippet on your site or scan a QR code for WhatsApp. You are live.
What is NOT happening during that five minutes, but should be happening invisibly: the platform is deciding when to route to GPT-4o versus Claude based on conversation complexity. It is setting reasonable escalation defaults (confidence threshold, sentiment trigger, three-strike rule). It is enabling multilingual handling because the model already speaks 95 languages. It is configuring brand voice from the tone of the content it just crawled.
None of those are customer decisions. They are platform defaults the platform should own.
The enterprise platforms charge for setup time because their architecture assumes the customer is the integrator. The right architecture assumes the customer is a founder running a business and the platform is the integrator. That is a different product, not just a different price point.
Frequently asked questions
How long should AI chatbot setup actually take in 2026?
For a small to mid-market business with a website, a help center, and one core tool (Shopify, Stripe, Salesforce, HubSpot), end-to-end setup should be under 30 minutes. The crawl takes 5-10 minutes. The OAuth takes 30 seconds. Testing takes 15 minutes. Anything longer is either an enterprise-scale custom integration (justified) or a platform pushing its configuration burden onto you (not justified).
Should I pick GPT-4o, Claude Sonnet 4.5, or Gemini 2.5 Pro for my chatbot?
You should not have to. The platform should route per conversation: cheap fast models for simple lookups, more expensive reasoning models for ambiguous or multi-step problems. If your platform forces a single global choice, it is asking you to optimize for either cost or quality but not both. A good platform absorbs the model decision and the cost variance.
Why are Intercom Fin and Zendesk AI Agents implementations so long?
Both platforms were architected before LLMs were good enough to reason without explicit flow trees. Their setup process is still essentially flow authoring with an LLM front-end. The intent taxonomy, training utterances, and escalation matrix are inherited from pre-2023 chatbot UX. The models are 2026, the configuration UX is 2019.
What is a reasonable monthly chatbot bill for a small business?
For a business doing 500-5000 conversations per month, a flat all-in fee in the $39-$169 range is reasonable. If your bill has multiple variable components (per resolution, per message, per token) it will be unpredictable and will surprise you in a high-traffic month. Flat pricing protects you from your own success.
Can I switch chatbot platforms once I have set one up?
This is the underrated reason to avoid heavy-setup platforms. The 47 intents, 31 sample conversations, and 412 configuration objects you built in Intercom do not export. Switching costs are the moat. A platform that takes 30 minutes to set up also takes 30 minutes to leave. That is a feature.
---
Disclosure: I am one of the founders of [SimplyBoost](https://simplyboost.io), a flat-priced AI agent platform ($39-$169/month) for sales and support across web, WhatsApp, Instagram, and Facebook Messenger. We pick the model per conversation, absorb the token cost, and ship setup in under 30 minutes — which is why I am opinionated about this. SimplyBoost B.V., KVK 87456346, EU-hosted in Frankfurt.