AI chatbots

AI chatbot development services that answer from your approved content

We build a website chat that quotes your own pages and policies, says when it is unsure, and passes the conversation to your staff.

Scope your chatbot
Trusted by the world's fastest-growing brands

What a chatbot build is

A website chat that answers from content you approved

AI chatbot development services build a chat for your site that answers from the pages, policies and price lists you approve, says when it does not know, and passes the conversation to a person. The test results come before launch, and you read them.

TAutomated assistantWhat if I pay late?A late fee is added after thedue date, per the fee schedulePayment and late feesCan you stop a lien sale?I’m not sure. Passing to staff.

Answers from your own material

The chat looks up the passage that matches the question and answers from it. You choose which documents it may read, and a passage you leave out is one it cannot quote.

Says when it does not know

A question with no matching passage gets a plain “I’m not sure” and an offer of a person. A confident guess is the failure we test for, because customers act on what a chat tells them.

Hands off to a person

Disputes, legal questions and anything the scope names go to your staff with the conversation attached. The customer does not have to repeat themselves.

Two kinds of chat

A scripted chat or an AI chat on your content

Both put a chat window on your site. They differ in how they find the answer, what they do with an odd question and what it takes to keep them current.

Scripted decision treeFollows a flowAI chat on approved contentAnswers from documents
How it finds an answerMatches buttons or keywords to a reply someone wrote in advanceReads the question, looks up matching passages in your approved documents and writes a reply from them
An odd questionFalls back to a menu or a contact form when no branch matchesCan answer a question nobody planned for, if the documents cover it, and hands off if they do not
Keeping it currentEdit the affected branches by hand, one flow at a timeUpdate the document and the chat reads the new version; the test set is re-run after a change
The riskA dead end the customer cannot get out of, and replies that go stale unnoticedA wrong or invented answer, which is why the build includes a test set and limits
SuitsA short, fixed set of tasks, such as booking a slot or tracking an orderSites with many policies, products or service details that customers ask about in their own words
Which one do you need?

Some sites need both: a flow for booking and an AI chat for questions. We look at a sample of the questions your staff answer, count how many are fixed tasks and how many are open questions, and recommend a mix in the written scope. If a scripted flow covers your volume, we say so.

What the sources say

Six facts to know before a chat goes live

A state disclosure law, two federal agencies, an accessibility standard and two security findings. The source sits under each point, and none of this is legal advice.

  1. State lawCalifornia

    California makes it unlawful to use a bot that hides its artificial identity so as to deceive people into a purchase or sale, unless it discloses that it is a bot.

    The law covers a bot used to communicate with a person in California online, with intent to mislead about its artificial identity so as to knowingly deceive them about the content of the communication and encourage a purchase or sale. A person who discloses is not liable under it. The disclosure must be clear, conspicuous and reasonably designed to inform. Our chat introduces itself as a bot.

  2. FederalCFPB

    The CFPB found that chatbots in consumer finance can give wrong answers, miss disputes and trap people in loops.

    Its June 2023 report says chatbots built on large language models sometimes get the answer wrong, and that only specific words may trigger the recognition of a dispute. It also describes “doom loops” where a customer cannot reach a person. It addresses banks, but the failures are the ones we test for: wrong answers, missed disputes and no way out to staff.

  3. FederalFTC

    The FTC said in its 2024 announcement that there is no AI exemption from the laws on the books.

    The September 25, 2024 announcement named five cases, among them one over a service marketed as an “AI lawyer” whose operator agreed to pay $193,000. The cases were about deceptive claims and schemes built around AI, including a tool that generated fake reviews. That is why the claims on your chat page, such as what it can answer, follow the test results in your own files.

  4. AccessibilityW3C

    WCAG 2.2 requires status messages to reach assistive technology without taking focus.

    Success criterion 4.1.3, at Level AA, says status messages can be programmatically determined through role or properties so assistive technologies can present them without receiving focus. In a chat, a new reply or a “typing” notice is that kind of message. We test the chat window with a screen reader as part of the build.

  5. SecurityOWASP

    OWASP ranks prompt injection first among the risks to language-model applications.

    A public chat box takes text from anyone, including people who try to talk it out of its instructions. OWASP’s 2025 list says retrieval does not fully remove the risk and names limited privileges and human approval for high-risk actions among the defenses. We keep the chat’s access narrow and test attempts to push it off its limits.

  6. Data exposureOWASP

    OWASP lists sensitive information disclosure as a top risk for language-model applications.

    A chat that can read customer records can repeat them to the wrong person. The 2025 entry recommends least-privilege access and limiting the data a model can reach. For a website chat, we begin with only the public and approved content and add account data only if the scope calls for it and a sign-in protects it.

The scope

What a chatbot build produces

Six parts of the work, each handing you a file to read and keep.

Content library33 approvedSourcePassagesCountUnit sizes and rates9Move-in checklist6Payment and late fees7Gate hours and access5Insurance plans6Liens and auctionsHeld back

An approved content library

We list the pages, policies, price sheets and PDFs the chat may read, mark each passage approved or held back, and fix the ones that contradict each other. You approve the library before the chat reads it. You receive: the content list with each passage’s status.

Limits sheetTo approveOpening line, shown firstv1 · Oct 8“I’m an automated assistantfor Tamarack Self Storage.”Unit sizes and ratesAnswersLate-fee disputesHands offLiens and auctionsHands off

Limits and the bot disclosure

We write what the chat may answer, what it must decline, and the wording that tells visitors it is an automated assistant. Your counsel can read the sheet before launch. You receive: a limits sheet and the disclosure text, for you to approve.

Staff inboxTranscript attachedDisputes a late feeTucson · 2:14 p.m.DisputeAsks about a lien noticeChandler · 1:52 p.m.LegalGate keypad will not openChandler · 11:08 a.m.AccessMessage left for the next day9:47 p.m.

Hand-off to your staff

We set the triggers that send a conversation to a person, who receives it, and what the customer sees outside your staff’s hours. The transcript travels with it. You receive: the hand-off rules and the after-hours message.

Test resultsAll 5 lines cleared98of 100, after fixesfirst run 89 · pass line 94Unit sizes and rates26 → 29 of 30Payment and late fees21 → 24 of 25Disputes and legal9 → 10 of 10

A test set and pass lines

We gather real questions from your inbox and phone notes, write the right answer for each with your staff, and set a pass line per kind of question before the chat is tested. You receive: the test set and the first-run results.

data-note.yamlFor your counsel1collects:2- question text3- name, email # at hand-off4sent_to_provider:5- question + passages6terms_read: Oct 8, 20267transcripts:8stored_in: your account9kept_for: set in the scope10card_numbers: not collected

Data and retention

We map what the chat collects, what is sent to the model provider, where transcripts are stored and for how long, and which provider data settings are on. You receive: a data note naming the provider terms we read, for your counsel.

Transcript logFirst 30 days1,200 chats started972 + 228972 answered228 to staffGaps from 120 chats readGate code reset after hoursNo passageHoliday gate hoursOutdatedCorrections added to the test set5

Transcripts and gaps

We set up a transcript log in your account, tag the questions the chat could not answer, and list the missing or outdated content behind each. You receive: access to the logs and a gaps list your team can act on.

How it runs

Approve the content, write the limits, test before launch

No work starts until you approve the scope, and no stage begins until you approve the one before. Each stage produces a document you keep.

content-library.csv

Approved content library

Sample · Sample Tamarack Self Storage (Tucson and Chandler, AZ)
SourcePassagesStatus
Unit sizes and rates9Approved
Move-in checklist6Approved
Payment and late fees7Approved
Gate hours and access5Approved
Insurance and protection plans6Approved
Lien and auction notices5Held back for staff
The chat reads 33 approved passages (9 + 6 + 7 + 5 + 6). The 5 lien and auction passages stay with staff, so the chat has nothing to quote on them.

Our stack

The tools a chat is built and watched with

The models, connections and monitoring behind a website chat.

OpenAI
Why OpenAI

OpenAI offers hosted language models through an API, with its own data controls and retention terms.

How we use it

We test its models on your questions, record the cost per conversation, and note which data settings are on.

ReasoningTool useRange
OpenAIBroad
Open modelsMid
Single vendorNarrow

Questions

AI chatbot development, answered

Start here

Put a chat on your site and read its test results first

Tell us where the chat would live and what it should answer from. You get the scope and its price in writing first, and you decide from there.