Custom AI Chatbot Development Services
A custom chatbot is a chat assistant built around one company's own records, systems and rules, not a template from a chatbot builder. We design, build and run them for US businesses: support and sales bots, assistants that answer from your documents, and copilots that draft replies for your team.
On this page
What our custom chatbot development services cover
The short answer: we design, build and run a chat assistant on your own records and systems. It answers from your documents, asks the questions your team would ask, and passes the conversation to a person when it should. It comes in six shapes, and one project can combine several.
- Support assistants: answer customers from your help center, policies and order records, and open a ticket for your team when they cannot.
- Sales and lead capture bots: ask the qualifying questions your sales team asks, collect contact details through a short form rather than in free text, and book the call.
- Assistants that answer from your documents: staff or customers ask in plain words and get an answer with the source named. The method is called retrieval augmented generation, or RAG.
- WhatsApp and website chat: one assistant behind the chat on your site and your WhatsApp Business number, writing to the same conversation record.
- Agentic workflows: assistants that take a step in another system, such as moving a booking or updating a CRM record, inside limits you set and with a person's approval where a mistake would cost money.
- Copilots for your team: draft replies and summaries for a person to check and send, as the copilot in our own CRM does.
Vendors often say a bot is trained on your data. In most builds the model is not retrained at all: it looks up the passages that fit the question and reads them at the moment it answers. That matters for you, because updating the bot then means updating a document, not paying for a new training run.
Not every request needs a model. If the questions are few and the answers fixed, a menu of buttons does the job for less, and our discovery sprint can end in exactly that advice. AI work that is not a conversation, such as pulling data out of documents or scoring leads, sits on our AI software development page.
The chat on this site, and the copilot that answers it
The Live Chat on this page is ours. We built the widget, the API behind it and the CRM inbox our team answers from, and an AI copilot inside that CRM writes the first replies. Open it and you are looking at the case study. We publish no results figure for it; what we can show is how it is built.
Six choices a chatbot of yours would inherit
- The opening line fits the page. The server holds 51 openers, one for each kind of page, and picks the line itself rather than accepting text from the browser, so nobody can plant words in our team's inbox. It also stores that line in the thread, so the copilot can read the question it opened with.
- Buttons answer the opener, and only a tap sends one. Three options reply to the greeting: I have a project, I need a quote, and Fix or improve my site. A button's message goes out only when the visitor picks it; the chat never posts a hello on anyone's behalf.
- It speaks only when spoken to. Opening the panel triggers no reply. A visitor's message wakes the copilot, and it sends one answer.
- A person can take over at any point. Once someone on our team takes a thread, the copilot stops replying in it. Its own greeting is stored as written by AI, so that line never counts as a person taking over.
- A form, not a fishing trip for an email. When the copilot needs contact details, it sends a marker that the widget turns into a short form. The request survives a reload or a new page, and once filled it never returns. Chatting needs no sign-up; the form becomes required after three messages.
- Files are checked before they are stored. Images, video, audio, PDFs, Office files, CSV and zip files up to 10 MB are accepted. Anything else is saved as a plain download, so an upload cannot run as script on our storage.
A visitor who comes back after five quiet hours starts a fresh conversation, and the old one stays on file in the CRM. The hand-offs to WhatsApp and email are written up on our CRM development page, and the API's other safeguards on our API and integration page.
Every chatbot build we take on has its architecture set and its code reviewed by our CTO, Divyansh Singh, and this one was no exception. On Clutch, our clients' rating stands at 5.0 from 66 verified reviews (checked 13 October 2026).
Where a chatbot earns its place, sector by sector
A chatbot pays off where the same questions arrive all day and the answers already sit in your records. Here is where that tends to hold in the sectors we build for, with the limit each one sets. Each industry page covers the wider software around the bot.
- Retail and online stores: order status, returns and product questions answered from your catalogue and order system. On Shopify, where we are a Shopify Select Partner, the store's own APIs supply the order data.
- Healthcare: booking, intake questions and after-hours answers from your own policies, never a diagnosis. The rules for health data are in the safety section.
- Real estate: questions about a listing, viewing requests and tenant maintenance reports, with the lead written straight into the agent's CRM.
- Logistics: shipment status looked up by reference number, and exceptions sorted and drafted for a dispatcher to approve.
- Education: enrollment, schedule and course questions, and a study assistant that answers only from the course's own material.
- Financial services: application status and account questions behind a sign-in. Identity checks and moving money stay in the systems built for them; the bot only reports on them.
- Professional services: client intake, document requests and appointment booking for firms whose staff spend the morning answering the same emails.
Buy a chatbot product, or build your own
Buy first if a product already does the job. Intercom prices its Fin AI agent at $0.99 per outcome on its pricing page (checked 17 October 2026). An outcome is counted once per conversation: when the customer confirms the issue is solved, asks for nothing more, or Fin completes a set workflow. On an existing helpdesk such as Salesforce it carries no seat or setup fees, above a monthly minimum.
The model under a custom bot costs far less per conversation. Anthropic's Claude Sonnet 5.5 and OpenAI's gpt-6.1-sol both list $2 per million input tokens and $10 per million output tokens (checked 17 October 2026). Both charge a twentieth of the input price for input they have already cached.
Here is our worked example. Say each visitor message makes the model read 3,000 tokens (its instructions, the passages it looked up and the chat so far) and write 300. Ten messages then come to 30,000 tokens in and 3,000 out: 6 cents plus 3 cents, so about 9 cents a conversation before any caching.
| Resolved chats a month | Fin at $0.99 each | Tokens for twice as many chats | Monthly gap | Months to repay $50,000 |
|---|---|---|---|---|
| 1,000 | $990 | $180 | $810 | About 62 |
| 10,000 | $9,900 | $1,800 | $8,100 | About 6 |
So fees alone justify a build only at volume. The stronger reasons are needs a packaged agent may not meet: the bot must act inside your own systems with each user's permissions, the data cannot leave your own cloud account, or the conversation is the product you sell. If none of those applies, buy, and our discovery will tell you so.
To run the comparison on your own numbers over five years, use our free build vs buy calculator.
What an AI chatbot costs to build and to run
GoodFirms surveyed more than 100 software companies for its 2026 cost report, updated 28 September 2026, and priced AI builds separately. The bands below are theirs, not ours. Most respondents also said AI integration alone can raise a build's cost by up to 30 percent, and the same report puts upkeep at 10 to 20 percent of the build each year.
| Tier | Usually holds | Benchmark | What moves it |
|---|---|---|---|
| Packaged agent | A product such as Fin answering from your help center | $0.99 per outcome at Intercom | How many chats it resolves |
| AI added to a system you have | One assistant inside an existing app, portal or CRM | Up to 30 percent on top of that build | How many systems it must read or write |
| First AI version (MVP) | A new assistant on your documents with one or two connections | $50,000 to $125,000 | Document clean-up, connections, the size of the test set |
| Medium AI project | Several workflows, user roles and channels | $125,000 to $250,000 | Actions in other systems, approvals, languages |
| Large or enterprise | Many teams, strict compliance, high volume | Above $250,000 | Security review, audit trail, uptime targets |
Running costs come on top: model tokens, as in the example above, plus hosting and, on WhatsApp, Meta's fees. Under Meta's WhatsApp Business Platform pricing, charges have been per template message since 1 July 2025, and free-form replies sent inside the 24-hour customer service window that opens when a customer writes cost nothing.
We do not print a chatbot price list. The first call takes 30 minutes, and if the work suits us a written scope and quote follow within 48 hours. The fixed price follows a paid two-week discovery sprint, and that sprint is allowed to conclude that a product like Fin would serve you better.
When the assistant is one part of a larger system, we price the whole as custom software: $40,000 to $150,000 for a focused first version, and $250,000 to $500,000 or more for complex enterprise platforms. The custom software cost guide explains what moves that figure, and our maintenance cost calculator estimates the yearly upkeep.
Models, search and connections: the four layers of a bot
A chatbot is four layers stacked on each other, and the choice made in each one decides what the bot costs to run and how hard it is to change later.
- The model. OpenAI's GPT models, Anthropic's Claude and Google's Gemini are all sold through APIs priced per million tokens. Pick per job: a small, cheap model to sort and route messages, a larger one to write answers. Keep the model behind one interface in the code, so a change of provider is a setting, not a rebuild.
- The search over your documents. When your records already sit in PostgreSQL, the open source pgvector extension keeps the search index beside them, under the same access rules. A separate vector database earns its place when the collection is very large.
- The connections. Your CRM, helpdesk, calendar, order system and WhatsApp Business account, reached through their published APIs. Our integration team builds and maintains these.
- The runtime. Our in-house stack: Python for AI features and data work, Node.js for APIs and background jobs, Cloudflare Workers at the network edge, and PostgreSQL on Supabase with row-level security. The chat on this site runs on the last two.
One more piece belongs in every build, though it is not code: a test set of real questions with agreed answers, rerun after every change to the instructions, the documents or the model. Without it, nobody can say whether last week's change made the bot better or worse.
Four steps from first call to a bot your customers use
- Discovery, weeks one and two. A sprint priced on its own that ends in a written scope and a fixed price, or with advice to buy. For a chatbot, ask for it to start from a sample of your real chats, emails or tickets: which questions repeat, what the bot may answer, what it may do, and when it must hand over. Ask for a draft test set out of discovery too.
- Build, from about day 15. Sprints of one to two weeks, each ending in a recorded demo of working code. A sensible order for a chatbot is the search over your documents first, then the conversation, the hand-over to a person and the connections to your systems.
- Test, before anyone outside sees it. Our QA team tests every build before you see a preview. A good chatbot build is also scored against the test set, with accuracy, response time and cost per conversation measured on real questions before launch.
- Launch, then run. Before the first release goes out, we prove the rollback works. After launch a named team keeps tuning it, or you take a handover pack (architecture notes, runbooks and a walkthrough) and run it yourselves.
On our builds, a first version (MVP) usually takes 8 to 12 weeks and a production platform 16 to 28. An assistant on your documents with one or two connections sits in the first group; one that acts across several systems is closer to the second. Our free requirements document generator helps you write the brief before the first call.
What the bot may say, and what it may do
In February 2024 a British Columbia tribunal held Air Canada responsible for what its website chatbot told a customer about bereavement fares. It called the airline's argument that the bot was a separate legal entity a "remarkable submission" and ordered it to pay 812.02 Canadian dollars. Treat every answer your bot gives as if it were printed on your website, because in that ruling it was.
The OWASP Top 10 for LLM applications, 2025 edition, ranks prompt injection first: text in a message or a document that tries to override the bot's instructions. Sixth is excessive agency, a bot allowed to do more than its job needs. So write these five rules into the scope at the start:
- It answers from your sources and names them. When the sources do not cover a question, it says so and offers a person.
- It reads only what the person in the chat may open, enforced in the database rather than requested in the prompt.
- It changes nothing important on its own. An action with money or a record behind it waits for a person to approve it.
- Keys stay on the server. The chat on this site keeps its database service key inside its own API; the browser never receives it.
- Usage has a ceiling. OWASP's tenth risk is unbounded consumption, a bot run up without limit; our own chat caps a visitor session at 300 messages.
Health data has its own rules. We sign a BAA whenever we handle protected health information, and our healthcare builds run on AWS, Azure or Google Cloud under the provider's BAA. A model provider that receives that information is a business associate as well, so it needs a signed BAA before the bot goes live.
How we contract the work, and what you own at the end
- Fixed bid: priced after discovery, against the written requirements. Anything beyond them is priced and dated in a written change request before anyone works on it.
- Dedicated team: a lead, engineers and QA billed monthly, normally for a quarter or more, for an assistant that keeps gaining jobs.
- Support after launch: a named team for tuning, new documents and incidents. A retainer starts at one senior engineer plus one designer; we do not staff anything smaller.
Under our MSA, the code, designs, copy and custom assets made specifically for your project are assigned to you on full payment; we keep the libraries and tools that existed before it or were built for reuse. From the first sprint, the code sits in repositories and cloud accounts you hold. Ask for the model provider account and its API keys to be yours as well, so the bot keeps running if you ever change teams.
Five questions to put to any chatbot vendor
- Which model will you use, and what will one conversation cost in tokens at our volume?
- Show us the test set: which real questions, and what score counts as ready to launch?
- What can the bot do in our systems, and which of those actions wait for a person?
- How does one of our staff take over a chat, and what do they see when they do?
- When the contract ends, do we receive the prompts, the test set and the search index along with the code?
Our own answers to the third and fourth are in the proof and safety sections above. Once you pick a vendor, put the agreed scope into our free SOW generator to get a statement of work you can both sign, and our custom software development company page sets out the rest of our terms.
Questions buyers ask before a chatbot build
How much does a custom chatbot cost?
GoodFirms' 2026 survey puts a first AI build at $50,000 to $125,000, a medium project at $125,000 to $250,000 and enterprise work above $250,000. A packaged agent costs less to start: Intercom's Fin is $0.99 per outcome, counted once per conversation. Running costs come on top, mostly model tokens, which our worked example puts at about 9 cents for a chat of ten messages. Our own pricing is in the cost section.
Can I build my own AI chatbot?
Yes. Chatbot builder tools and the model providers' own developer kits let one person put a bot on a handful of documents quickly. The hard parts arrive later: keeping answers inside your sources, limiting what the bot may do, handing a chat to a person, and testing every change. If those matter to you, budget for them whether you build in-house or hire a team.
Can I get paid to train an AI chatbot?
Yes. Model makers pay people to write example answers and rank model replies: for InstructGPT, OpenAI hired about 40 contractors through Upwork and Scale AI, according to its 2022 paper. On a custom business bot the model is usually not retrained at all; the paid work is reviewing transcripts and keeping the test set and source documents current.
How long does it take to build a custom chatbot?
With us, discovery fills the first two weeks and build sprints start around day 15. A first version usually takes 8 to 12 weeks in total; an assistant that acts across several systems is closer to our 16 to 28 week platform builds. Clean, current source documents shorten the work; scattered ones add a clean-up task before the first sprint.
Is a custom chatbot better than a SaaS chatbot tool?
Not by default. A packaged agent is cheaper to start and works well for answering from a help center. Custom wins when the bot must act inside your own systems with each user's permissions, when the data cannot leave your cloud account, or at volumes where per-outcome fees outgrow a build. The build or buy section shows where that line falls.
Can a chatbot connect to our CRM and helpdesk?
Yes, through their APIs. Reading is the easy half: order status, a contact's history, an open ticket. Writing needs rules about which fields the bot may change and which changes wait for a person. Our own chat writes every conversation into the CRM our team works from, on that visitor's record. If you are replacing the CRM itself, our custom CRM development company shortlist compares eight builders.
How do you stop a chatbot from making things up?
No model can be made incapable of it, so the build works around it. The bot answers from passages looked up in your documents and names them, it is told to say it does not know when they are silent, and a test set of real questions with agreed answers is rerun after every change. Questions with money or health behind them go to a person.
Do we have to tell people they are talking to a bot?
In California, saying so is what protects you. Business and Professions Code section 17941 makes it unlawful to use a bot that sets out to mislead people about its artificial identity in order to drive a sale, and says a clear disclosure removes that liability. Other states and countries set their own rules, so ask your counsel and design the disclosure into the chat window.
Who owns the chatbot once it is built?
With us, you do once the final invoice is paid: under our MSA, the code, designs, copy and custom assets made for your project are assigned to you, in repositories held in your name. With any vendor, ask for the prompts, the test set and the search index to be named in the handover, since they decide how the bot answers.
Last reviewed .
Bring the questions your team answers every day
List ten questions your customers or staff ask most, and where the answers live today. Bring that list to a 30-minute call; if the bot is worth building, a written scope and quote reach you within 48 hours, and discovery turns the quote into a fixed price. Book a call with a chatbot developer.
Why work with Digital Heroes
115 people across five studios in New York, Delhi, London, Sydney and Lucknow, shipping ecommerce, web, software and mobile work for founder-led brands. Senior engineers only, no account-manager relay, and the same team from kickoff to launch.
- Shopify Select Partner, as listed in the Shopify Partner Directory
- Tier 1 registered supplier on the United Nations Global Marketplace, which requires audited company documentation
- Top Rated Plus on Upwork, the bracket for the top 3% of talent by client outcomes
- 397 public five-star-weighted reviews across four Fiverr gigs, from clients in 36 countries
- DUNS verified business, No. 650878346, so procurement can check us before signing
- Listed by Clutch in Top 1000 Global B2B Companies and Top Web Designers 2024
- Verified profiles on GoodFirms, DesignRush, Digital.com, AppFutura and TopDevelopers
- 2,000+ brands built across 40+ countries, with $50M+ in client revenue scaled
More from Divyansh.
Chief Technology Officer · Lucknow · 50 posts on the site
Published .
