Should You Build or Buy an AI Chatbot for Customer Service?

Buy, in almost every small business. A bought AI chatbot costs roughly $30 to $250 a month, or about $0.50 to $1 per resolved conversation, and goes live in days. Build only when the bot must act inside a system no product can connect to, or your volume is high enough that per-resolution fees exceed a developer's time.

The real choice has three options, not two. Between a ready-made product and a from-scratch build sits a middle route: assembling a bot on a platform that lets you add your own connections without writing an application. That middle route is where most "we need a custom bot" requirements end up, because it solves the one thing off-the-shelf products can't reach while leaving hosting, security patches and the chat widget to someone else.

Follow me on Instagram@sagnikteaches

Buy, assemble or build: what each one really means

Buy means a finished product: you paste in your website address and help documents, adjust the tone, add the widget to your site and switch it on. Intercom's Fin, HubSpot's Customer Agent, Tidio's Lyro, Chatbase and the free AI agent now built into Shopify Inbox all work this way. You control what the bot knows and when it hands over; you don't control how it works inside.

Connect on LinkedInSagnik Bhattacharya

Assemble means a bot-building platform such as Botpress or Voiceflow. You design the conversation, connect it to your own systems through their integrations or a web request, and the platform hosts it. You need someone comfortable with logic and settings, but not a software developer.

Subscribe on YouTube@codingliquids

Build means a developer writes an application that calls an AI model through its API (the programming connection vendors sell by usage), stores your knowledge, runs the chat widget and connects to your systems. You own every part, including every part that breaks.

A quick way to tell which you need: write down the five questions customers ask most, and next to each, what the bot must do to answer it. If every answer is "look it up in our policies", buy. If one answer is "check our job system", see whether a product or platform already connects to that system. Only if nothing does is building on the table.

What each route costs over 12 months

List prices as of September 2026. Treat the build figures as assumptions to test against real quotes; the product prices come from the vendors.

The surprise in that table is the AI model itself. A typical support chat of six exchanges might send around 18,000 tokens of instructions and knowledge to the model and get about 1,200 back. At Claude Haiku 4.5's API price of $1 per million input tokens and $5 per million output, that is roughly two and a half cents a chat. On OpenAI's gpt-6-luna, at $0.10 and $0.50 per million, it is a fraction of a cent. For a small business, building is never expensive because of the AI. It is expensive because of the people who write, host, test and fix the application around it.

The volume where per-resolution pricing overtakes a build

Per-resolution pricing looks alarming when you multiply it out, so it is worth doing the sum properly. Using the assumptions above for an illustrative build (a $12,000 project, $150 a month for hosting and monitoring, and four hours a month of developer upkeep at $75 an hour):

• Year-one build cost: $12,000 + $1,800 + $3,600 = $17,400. Later years: about $5,400.
• Fin at 300 resolutions a month: 300 × $0.99 × 12 = $3,564 a year.
• Fin at 2,000 resolutions a month: 2,000 × $0.99 × 12 = $23,760 a year.

Over three years, the build costs about $28,200 in total. Fin at 300 resolutions a month costs about $10,700 over the same period; at 2,000 a month it costs about $71,300. So building only starts to win on cost somewhere above roughly 800 resolutions a month, and only if the custom bot resolves as many chats as the product does. Few owner-run businesses reach that. The ones that do usually sell online at volume and have a developer already.

Seat fees and subscriptions shift the crossover, so rerun the sum with your own quotes. The pattern holds, though: below a few hundred chats a month, buying wins on cost almost regardless of which product you pick.

Four signs that building is the right call

1. The bot must take an action in a system nothing connects to. For example, checking live van availability in a job-management system that has an API but no chatbot integration, and no platform supports it.
2. Volume is in the thousands of resolutions a month, as in the sum above, and you've confirmed the per-resolution fee is the main cost.
3. You have someone to own it. A developer on staff or on a retainer who will fix it at 8am when it breaks, update it when the model is retired, and review conversations weekly.
4. Your data rules require it. A contract or regulator-facing obligation that no vendor's terms can meet, such as keeping every conversation on infrastructure you control. Check this with your adviser before assuming it; most vendors' business terms cover ordinary customer-service data.

If only one of the four is true, assemble on a platform. If none is true, buy. I have rarely seen all four in a business with fewer than 50 staff.

A removals firm works through the choice

Here is how the decision plays out for an illustrative six-person removals firm. In peak season its website chat gets about 400 conversations a month, falling to 150 in winter. The owner's list of the top questions:

1. Are you free on [date]?
2. Roughly how much for a three-bedroom house, 20 miles?
3. Do you supply boxes, and what do they cost?
4. Are my belongings insured during the move?
5. Do I need to arrange parking outside my house?

Questions 3 to 5 are policy lookups: any bought product can answer them from the firm's website and a two-page FAQ. Question 2 is a ballpark: a bought bot can quote a range from a published price guide ("most three-bedroom moves under 30 miles fall between X and Y; we confirm after a video survey"), which is what the firm wants anyway, since it never gives fixed prices without a survey. Question 1 was the reason the owner thought he needed a custom build: he wanted the bot to check the diary.

Three routes, costed:

• Buy Chatbase Standard at $150 a month. Peak season is about 400 chats × 6 messages = 2,400 message credits, inside the 4,000 included (depending on the model chosen, a message can use more than one credit, so check). The bot answers questions 2 to 5 and, for dates, says "I can't see the diary, but I'll pass your date to the team, who reply within two working hours", collecting name, date, addresses and phone.
• Assemble on Botpress Plus at $150 a month plus 150 extra peak conversations at $0.65, about $250 a month in peak. His job system has an API, so the bot could check free days. Build time: perhaps 20-30 hours for someone who knows the platform.
• Build: a developer quote of $9,500 plus hosting.

The deciding detail came from looking at the diary itself. A "free" day isn't really free until the owner has checked crew, van size and distance, so even a live lookup would have to say "provisionally available". Customers would still wait for a human confirmation. That made the diary connection worth very little, and the firm bought. Six months in, the date question is handled by the handover form, and replies within two hours convert just as well as the owner hoped the instant answer would.

What a custom build must include after launch day

If a developer quotes for a custom bot, check that these are priced, because they are where the ongoing cost lives:

• Knowledge updates. Where do prices and policies live, and who changes them? If they are pasted into the bot's instructions, every price rise needs a developer.
• Model retirements. AI vendors retire older models on a schedule. Someone has to move the bot to a newer model and retest it, typically at least once a year.
• Testing before changes. A fixed set of 30-50 test questions with expected answers, run after every change.
• Misuse protection. Customers will try to make the bot say odd things or promise discounts. The tutorial on protecting a customer chatbot from misuse covers the defences; a build has to include them.
• Handover to a human. Routing to email, phone or a help desk, with the conversation attached.
• Conversation logs and privacy. Where chats are stored, for how long, and how you delete a customer's data on request.
• Disclosure. If you sell to customers in the EU, the EU AI Act's transparency duty, which has applied since 2 August 2026, means telling people they are talking to an AI. Bought products usually have a setting for this; a build has to add it.

A realistic failure shows why the first item matters. Imagine an electrician's custom bot built with the price list written directly into its instructions. The firm raised its call-out charge in January, updated the website, and forgot the bot. For six weeks the bot quoted the old figure, and several customers held the firm to it. A bought product that reads the live website would have picked up the change on its next refresh; the build needed a paid developer hour and someone to remember.

Testing a bought bot on your real questions first

Before paying for any product, run it on a free trial against your own questions, including awkward ones. Twenty questions is enough to see how it behaves. For a cleaning company, a test set might include:

1. How much is a deep clean for a two-bed flat?
2. Do you bring your own products? I'm allergic to bleach.
3. Can you come this Saturday?
4. Your cleaner broke a vase last week. What happens now?
5. Do you clean ovens?
6. Can I get 50% off if I book weekly?
7. Are your cleaners insured and background-checked?
8. What if I'm not happy with the clean?

Illustrative answers from a bot trained only on the company's website, and what I'd fix:

Q2: "Yes, our cleaners bring all products and equipment. We use
eco-friendly products, but please let us know about allergies when
booking."
Q4: "I'm sorry to hear that. Accidents are covered by our insurance.
Please contact us and we'll arrange compensation."
Q6: "We offer discounts for regular bookings! Weekly clients save up
to 50%."

Q2 is fine. Q4 promises compensation, which the company's terms don't guarantee; the fix is a rule that any damage or complaint goes straight to a human with no promises made. Q6 is worse: the website mentions "regular client discounts" without a figure, and the bot echoed the customer's 50%. The fix is to publish the real discount (say, 10% for weekly) in the FAQ and add an instruction never to confirm a discount the knowledge doesn't state. If a product can't be tuned to fix both within the trial, try another. The fuller method is in the tutorial on stopping a chatbot giving wrong answers.

An assembled bot for a landscaper

The middle route earns its place when one connection matters. Take an illustrative landscaping firm with 120 maintenance contracts. Its most common chat question is "When is my next visit?", which lives in its scheduling system. A bought product can't see that. The scheduling system has an API, and an assembled bot on a platform can ask for the customer's house number and contract email, look up the next visit and reply.

That build is a few days' work for someone who knows the platform, and the platform handles hosting and the widget. The firm still keeps everything else (prices, services, seasonal advice) as knowledge the platform reads, so a price change is a document edit, not a developer job. If the landscaper had to build this from scratch, the same one connection would carry the cost of an entire application.

A filled-in scoring sheet

Score each statement 0 (not true), 1 (partly) or 2 (true). Here it is filled in for the removals firm above.

Reading it: high scores on the first two rows point to buying. A high score on rows 3 and 4 together points to assembling. Building needs high scores on rows 4, 5 and 6. The removals firm scored 2 and 1 on the buy rows, but 0 on row 4, so the API didn't matter and it bought. If you want the wider picture on whether you need a bot, a live chat tool or a help desk first, see chatbot vs live chat vs help desk, and for the costs of a bought website bot in more detail, what a website chatbot costs and whether it pays off.

Signs you chose wrong, and how to switch

Review the decision after three months with the bot's own reports. If you bought and the handover rate is above about 40% because customers keep asking something the bot can't reach, the missing connection is real, and assembling is the next step. If you built and your developer time is climbing each month, or the bot fell behind on prices, move to a product and keep your test questions and knowledge documents; they transfer to any platform. In either direction, the knowledge base you wrote is the asset you keep, so write it as ordinary documents rather than locking it inside one tool.

Further reads

• What to Ask an AI Chatbot Vendor Before You Sign Up — The questions to put to a vendor once you've decided to buy.
• How to Train an AI Chatbot on Your FAQs, Policies, and Prices — Loading your prices and policies so the bot answers correctly.
• How to Test a Customer Chatbot Before It Goes Live — A fuller pre-launch test plan than the one here.
• When Should an AI Chatbot Hand Over to a Human? — Setting the point where the bot passes a customer to a person.
• Can You Build an AI Chatbot for Your Business Without Coding? — How far no-code platforms go if you choose to assemble.
• How to Measure Whether Your AI Chatbot Is Actually Working — The numbers that show whether your choice was right.
• Should Coaches Build an AI Version of Their Method for Clients? — When an AI version of a coaching method helps clients and when it undermines you, what to build it on now custom GPTs are retiring, and the guardrails it needs.
• AI Chatbots for Nursery Enquiries and Visit Bookings — What a nursery chatbot should answer, what it must hand to the manager, how show-round booking works, and the two mistakes that cost parents' trust.
• AI Customer Service for Small Businesses: What to Automate First — Rank your customer messages by volume and risk, then automate in order: sorting, drafted replies, proactive messages, FAQ answers, and only then actions.
• Best AI Chatbots for Small Business Websites — Tidio, Crisp, tawk.to, Chatbase, Help Scout, Intercom Fin, Zendesk, Freshdesk, HubSpot and platform-native chat, priced and matched to real businesses.
• What Is Copilot Studio and Does a Small Business Need It? — What Microsoft's agent-building tool actually does, how its credit billing adds up, and the four situations where a small firm genuinely needs it.
• Should a Small Business Build a Custom AI Tool or Buy One? — A seven-question decision table, one community interest company's two-year sums for buy, configure and build, and what a build costs after launch.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: facts sheet for Intercom Fin and HubSpot Customer Agent pricing and API token prices; Tidio, Chatbase and Botpress pricing pages. Checked September 2026.