The Question I Get Right Before "How Much Will This Cost?"
A restaurant group reached out last quarter wanting a chatbot to handle reservation questions, menu inquiries, and allergy information. Before we talked about price, I asked one question: "What happens right now when someone calls with a question your host can't answer?" The answer was that the host asks a manager, or the customer calls back. That gap — not the technology — is what determines whether an AI chatbot for business is worth building.
Most of the chatbot proposals I review skip that question entirely. They start with the model, the platform, or the demo, and never establish what specific interaction volume and failure cost justify the investment. I want to walk through the numbers I actually use with clients, including where this technology quietly falls apart.
The Fit Test Before the Budget Conversation
A chatbot earns its cost when three conditions are true at once:
- Volume is real. You are fielding at least a few hundred repetitive customer questions a month — order status, business hours, appointment availability, pricing tiers, return policy. Below that volume, the fixed cost of building and maintaining the thing outweighs the labor it saves.
- The knowledge is bounded and stable. The chatbot needs a defined universe of correct answers — your policies, your inventory, your calendar — not open-ended judgment calls. The more a question resembles "it depends," the worse a chatbot performs.
- A human safety net already exists. If the bot gets something wrong or hits its limit, there needs to be a fast, visible path to a person. Businesses that treat the chatbot as a replacement for support staff rather than a filter in front of them are the ones I see run into trouble.
If any of those three is missing, I recommend fixing that first. A chatbot layered on top of undocumented, inconsistent policies just answers questions wrong faster than a human would.
What This Actually Costs: Build vs. Buy
There are two real paths, and the cost structures are not close to each other.
Buying a platform (Intercom Fin, Ada, Freshchat, and similar tools) typically runs $50–$500 a month for SMB-scale volume, though many of these vendors now charge per resolution rather than a flat fee — commonly somewhere in the $0.50–$1.00 range per conversation the bot fully resolves. At 1,500 resolved conversations a month, that alone is $750–$1,500, on top of your base subscription. This model rewards you for low volume and punishes you as you scale, so run the math at your actual expected volume, not today's volume.
Building custom on an LLM API means paying for tokens directly — usually a few cents to under a dollar per conversation depending on the model and context length, plus $20,000–$50,000 in initial development for a properly scoped bot with retrieval over your actual business content, and $1,500–$5,000 a month in ongoing maintenance: content updates, monitoring, prompt tuning, and handling the edge cases your customers will inevitably find. I tell clients to budget maintenance at roughly 15–20% of the initial build cost per year, minimum, because a chatbot is not a "ship it and forget it" asset — your policies, prices, and inventory change, and the bot has to keep up.
In my experience, the total cost of ownership on both paths tends to land at two to three times whatever number was in the original pitch. Budget for that multiplier up front rather than discovering it in month four.
Where the ROI Actually Comes From
The return is not "we sound innovative." It is hours of human time removed from repetitive work, measured against the fully loaded cost of that time. If a support rep spends 15 hours a week answering "what are your hours" and "where's my order," and a chatbot can deflect 60–70% of that reliably, you have a real number to compare against the monthly cost above.
This is also where booking and scheduling use cases tend to pay off fastest, because the task is transactional and bounded — check availability, confirm a slot, send a reminder. I go deeper into that specific pattern in my article on automating business bookings, and it is worth reading before you scope a chatbot for anything appointment-related, because a dedicated booking flow often beats a general chatbot for that use case.
Realistic payback for a mid-cost implementation ($15,000–$30,000 all-in for year one) against genuine labor savings is usually 8–14 months, not the 60-day miracle some vendors imply. If your projected payback is longer than 18 months, I would slow down and question the volume assumptions before signing anything.
Where Chatbots Fail, and Why It's Rarely the Model's Fault
I've seen three failure patterns repeat across almost every project that goes sideways:
Hallucination on policy questions. When a chatbot isn't strictly grounded in your actual documents, it will confidently invent an answer that sounds plausible. This isn't hypothetical — a well-known 2024 case involved an airline held responsible by a tribunal after its support chatbot fabricated a bereavement fare policy that didn't exist, and the airline was ordered to honor it. Grounding the model tightly in your real content — and refusing to answer outside that scope — cuts this risk dramatically, but it doesn't eliminate it, so plan for occasional wrong answers rather than assuming zero.
No real handoff path. A chatbot that traps a frustrated customer in a loop does more brand damage than having no chatbot at all. The fix is a visible, low-friction escalation — not a hidden "contact us" link three menus deep — triggered automatically when the bot's confidence drops or a customer repeats a question.
Scope creep after launch. The bot that started as "answer FAQ questions" becomes "also handle complaints, also process returns, also upsell" within a few months, because it's easier to ask the bot to do one more thing than to build a new feature. Each addition multiplies the testing and monitoring burden. I recommend a hard-scoped v1, measured for 60–90 days, before any expansion.
This connects to a broader pattern I've written about in why AI projects fail once they hit production — chatbots are exposed directly to customers, which means the gap between a good demo and a reliable production system shows up faster and more publicly than with internal AI tools.
The Decision Framework, Simplified
Before approving a chatbot budget, I ask clients to answer three questions in writing:
- What is the actual monthly volume of repetitive questions this will handle, and how confident are we in that number?
- What does a wrong answer cost us — a refund, a bad review, a support ticket, or a compliance problem?
- Who owns this after launch, and what is the monthly budget for keeping it accurate?
If those three answers are solid, a chatbot is very likely worth building. If they're vague, the project needs another month of discovery before it needs a developer.
Chatbots are a genuinely useful tool for the right volume and the right scope — they are not a strategy, and they are not free after month one. If you want an honest read on whether the numbers work for your business specifically, let's talk through your situation.