AI Chatbot for Business: How to Choose and Build One
Three people at the same company say they want a chatbot, and they mean three different things. Sales wants a widget on the pricing page that qualifies a visitor and books a call. Support wants something to clear the "where is my order" queue. Operations wants an internal bot that answers "what's our PTO policy" so it stops landing in the founder's DMs. Same word, three separate builds, three different costs.
That overloaded term is why so much chatbot shopping goes sideways. A tool that is excellent at support deflection is the wrong pick for internal knowledge, and the demo that looked magical falls apart the moment it has to read your real order data instead of a canned FAQ.
This is a buyer's guide for a lean team that wants a chatbot to work, not just demo well. The route that fits you turns almost entirely on one variable most shoppers never name up front: how much of your real stack the bot has to touch. Sort that out and the rest of the decision gets simple.
The three jobs hiding inside "chatbot for business"
Before you compare a single tool, decide which of these you are actually building. They share a language model and almost nothing else.
- Customer support. Deflects repetitive tickets, looks up an order, changes a subscription, hands the hard ones to a person. Public-facing, tied to your helpdesk and commerce stack. This is the deepest category, and we cover its capabilities in detail in the customer service chatbot guide and the build itself in how to build a support chatbot.
- Lead capture and qualification. Sits on your marketing site, answers pre-sale questions, qualifies a visitor, and books a meeting or drops the lead into your CRM. Tied to your sales tools, not your helpdesk.
- Internal knowledge. Answers employee questions from your own docs, policies, and wikis. Never customer-facing, so the bar for tone is lower but the bar for not inventing a policy is just as high.
A tool sold for one of these can sometimes be bent to another, but the integrations, the guardrails, and the pricing model are built around one job. Pick the job first. Everything downstream depends on it.
The capabilities that decide whether it works
Across all three jobs, the same short list separates a chatbot that earns its cost from one that generates cleanup work. The 40-row feature chart on a vendor site is noise around these five.
Grounded answers from your own content
A bot answering from a model's general knowledge will confidently state a refund window or a policy you do not have. The capability that matters is retrieval: the bot reads your help center, your docs, or your past tickets, and answers only from that material. Whether that content is clean enough to answer from is its own question, one we walk through in is your data ready to automate. Without grounding, every confident wrong answer is a ticket you clean up twice.
Real actions, not just answers
Quoting your FAQ is trivia. Looking up order 48213 and telling the customer it ships Thursday is the job. The capability is tool use: the bot calls your order system, CRM, or billing provider through an API, reads the real record, and where you allow it, writes back. A bot that can only talk deflects the easy half of the easy questions. One that can act closes the whole category, which is where the volume lives.
A clean handoff to a human
Every bot hits questions it should not answer: an angry customer, a billing dispute, anything legal. A good handoff passes the full conversation and account context straight to a person inside your existing tools, with nothing repeated. A bad one dumps the customer back at the start of a queue to explain themselves again. This is where cheap bots quietly fail, and it is the part people screenshot.
Guardrails and scope control
A bot with no leash tries to answer everything, including what it should refuse. You want a confidence threshold below which it escalates instead of guessing, topics it is forbidden to touch, and a hard rule that it never states policy it cannot cite. For anything touching payments or personal data, a narrow reliable bot beats a broad confident one every time.
Integration with the stack you already run
A chatbot is only as useful as the systems it can reach. If it cannot see your Shopify orders, your Stripe subscriptions, or your HubSpot records, it is a nicer FAQ page. Real integration through supported APIs and webhooks is usually the hardest part of a deployment and the first thing a demo glosses over. It is also the part that decides everything, which is why we come back to it below.
The best ways to run an AI chatbot for your business
There are four honest routes, and they are not interchangeable. Here they are, ranked for a team in the 1-50 person range that wants the thing to actually work.
1. Work with bottta to design and build it
Buying a chatbot is the easy part. The work that decides whether it earns its cost is the part every vendor leaves to you: grounding it on clean content, wiring it into your real systems, writing the escalation rules, and watching it after launch. That build is what we do at bottta.
For a support or lead bot, that is our AI Automation work: retrieval over your own content so the bot answers from your material and not the model's imagination, plus the confidence thresholds and guardrails that keep it in its lane. It is our Integrations work to connect it to the systems that hold the answers, the order lookups, the CRM writes, the subscription changes through Stripe and your helpdesk by API and webhook. And it is a handoff into the tools your team already lives in, so a person never restarts from zero.
We work two ways. The $4K fixed-scope project fits when you want a defined bot built, grounded, integrated, and shipped, with 30 days of post-launch support to tune it against real conversations. The $3K/month retainer fits the reality that a bot is never done: knowledge bases drift, products change, and a bot nobody maintains slowly starts giving last quarter's answers. The retainer covers monitoring, content upkeep, and fixes across up to three active workflows, so the resolution rate you launched with is the one you still have in six months.
This is the right route when the bot has to touch real systems and real data, which is almost always the case once it does anything past reciting an FAQ. The platforms below are genuinely useful, but they hand you the model and leave the wiring, and the wiring is the job.
2. Off-the-shelf AI agent platforms
If your need is narrow and mostly support deflection, a platform gets you live fast. These tools bundle the bot, a help-center connector, and a helpdesk in one subscription, and most have moved to per-resolution pricing so the bill scales with volume instead of sitting flat.
| Platform | Base cost | AI pricing | Best for | |---|---|---|---| | Intercom (Fin) | Seats from $29/agent/month billed annually | $0.99 per resolution | Support teams already on or open to Intercom | | HubSpot (Customer Agent) | Bundled with Service Hub | $0.50 per resolved conversation (50 credits at $0.010 each) | Teams already running HubSpot CRM | | Tidio (Lyro) | Free tier, paid plans from $29/month | Lyro conversations from $0.50 each, sold in monthly packages | Small e-commerce and SMB sites |
Intercom lists Fin at $0.99 per resolution on its pricing page, billed on top of seats that start at $29 per agent per month annually. HubSpot moved its Customer Agent from $1.00 per conversation to $0.50 per resolved conversation in April 2026, paid as 50 credits at $0.010 each, per HubSpot's AI pricing page. Tidio's Lyro agent starts every account with 50 free non-renewable conversations, then sells renewable monthly packages, with a Lyro conversation starting from $0.50, per Tidio's pricing page.
The trap in per-resolution pricing is that a poorly grounded bot still bills you. If it "resolves" a chat by giving a wrong answer and the customer reopens the ticket, you paid for the bad resolution and still handle it by hand. Cheap-per-resolution and expensive-in-practice are the same platform when the knowledge base is thin. These platforms are a real fit for a narrow support job on a stack they already integrate with. They are a poor fit the moment the bot has to reach a system they do not natively support, which is exactly where a build earns its keep.
3. A custom build on the OpenAI or Anthropic API
If you have an engineer with spare cycles, the model itself is cheap and you can wire a bot directly to your stack with no platform in the middle. OpenAI's flagship GPT-5.5 runs $5 per million input tokens and $30 per million output tokens per its API pricing. Anthropic's Claude Sonnet 5 is $3 and $15, with the smaller Haiku 4.5 at $1 and $5, per Anthropic's pricing docs. A single chatbot conversation is a few thousand tokens, so the model call is almost never the expensive part.
The cost is not the tokens. It is the engineering around them: retrieval, guardrails, retries when the API times out, versioned prompts, monitoring for silent failures, and updating everything when a vendor changes an endpoint. The line between a chat window and a wired workflow is exactly this glue. A first version is a weekend. The version that survives real traffic and does not silently break at 2am is ongoing work, and it usually lands on the one person who least has time for it.
4. Hiring in-house
For a company that expects to run several bots and other automations for years, hiring an automation or support-ops engineer to own it is defensible. The trade-off is cost and lead time: you are paying a salary and waiting on a hire for a scope that a project ships in weeks. For most lean teams with one or two bots to stand up, it is more capacity than the problem needs.
How to choose
The decision comes down to how much the bot has to touch and how narrow the job is.
- One support job, on a stack a platform already supports. Start with an off-the-shelf platform. If Intercom, HubSpot, or Tidio already sits in your stack and you mostly need FAQ deflection, the fastest path is the built-in agent. Watch the per-resolution bill and keep the knowledge base clean.
- A bot that reaches real systems, or more than one job. This is a build. The moment the bot has to look up an order, write to the CRM, qualify a lead into your pipeline, or answer from internal docs the platform cannot see, the integration is the whole project, and that is what we do at bottta.
- Not sure it is worth automating at all. Run the task through the frequency, stakes, and input-stability test first. A low-volume question does not need a bot. A person answering the same 200 messages a week does.
One thing to skip for now: voice and phone AI agents are a separate, harder build. Solve chat first, where most repetitive volume lives, and come back to voice once the text version is earning its cost.
Frequently asked questions
How much does an AI chatbot for business cost?
It depends on the route. Off-the-shelf platforms bill per resolution, roughly $0.50 to $1.00 each on Intercom, HubSpot, and Tidio, on top of any seat cost. A custom build's model cost is tiny (a conversation is a few thousand tokens against $3 to $30 per million), so the real cost is the build and monitoring. A bottta project is a fixed $4K with integrations included, or a $3K/month retainer for ongoing work.
What is the difference between a chatbot and an AI agent?
A traditional chatbot follows scripted button-and-branch flows: click this, see that. An AI agent uses a language model to understand a free-text question, answer from your content, and take an action through your tools. In practice the terms have merged, but if a vendor still means rigid decision trees, it will struggle with anything a customer phrases in their own words.
Can an AI chatbot connect to our CRM and other tools?
That connection is the entire point, and it is where deployments succeed or fail. Platforms integrate natively with a fixed set of tools, so a bot works well inside a stack it supports and hits a wall outside it. A custom build reaches anything with an API. If the bot has to read and write across tools that do not natively talk, that integration work is exactly the layer we build.
Will a chatbot replace my support or sales team?
No, and building for that goal is how deployments fail. The realistic outcome is that the bot clears the repetitive, high-volume, low-stakes questions so your team spends its time on the conversations that need a human. Design the escalation path first, staff for the hard cases, and let the bot take the easy queue.
How do I stop a chatbot from making things up?
Ground it. Restrict it to answering from approved content, set a confidence threshold that escalates uncertain questions to a person, and forbid it from stating policy it cannot cite. A bot that says "let me get a teammate" when it is unsure is worth far more than one that guesses.
A chatbot for your business is not one purchase. It is a decision about which job you are solving and how much of your stack the bot has to reach, and the answer to that second half is almost always "more than a platform can see on its own." The gap between a bot that talks and a bot that resolves is grounding, integrations, and monitoring, and those are build decisions, not buying decisions. That build is the work bottta owns end to end: a bot that clears real volume instead of one people fight through to reach a person. Start a project or book a call with bottta.