Useful answers need sources, uncertainty, and a human exit

Build an AI Chatbot That Knows Its Boundary

Describe the support job, approved knowledge, answer format, escalation route, and allowed actions. Playcode builds the interface, server, source retrieval, safety controls, tests, and hosting path around the model provider you choose.

No credit card required · No coding needed

Quick answer

What should an AI chatbot builder include?

An AI chatbot needs approved knowledge sources, grounded answers with citations, an explicit uncertain state, human escalation, prompt-injection defenses, rate limits, privacy and retention rules, and a safe boundary around external actions. Playcode can build those application controls around your chosen model provider, then test known answers, unknowns, abuse, failures, and recovery before launch.

1.1M+users
26M+projects
Since 2016

Used by people at

Counts registered accounts using corporate email domains from the companies above. Playcode is a self-serve product, not a contracted vendor to these brands.

From Approved Sources to a Chatbot People Can Verify

Prove answers and failure behavior before adding more channels or actions

01

Define Sources, Unknowns, and Escalation

Name the exact documents the chatbot may use, who maintains them, how answers cite them, what counts as insufficient evidence, and where a user goes when the bot is uncertain, blocked, or unavailable.

Start with a small fictional source set so retrieval, citations, privacy, and deletion can be tested without exposing customer data.

Preview
AI Chat
Publish
I'm a realtor in Miami. I need a website to show my listings and get buyer leads.
4 Questions
2 / 4
What's the main style for your real estate site?
AModern & minimal with clean lines
BLuxury feel with large property photos
CProfessional & corporate
DI'll describe my own style below
Or type your own answer...
Skip
Next
Describe what you want to change...
Economy
Build Progress
4/6
✓
Site layout & navigation
3s
✓
Hero section with CTA
5s
✓
Property listings grid
8s
Building contact form...
About page & bio
Footer & SEO meta
miami-homes.playcode.io
Miami Dream Homes
Your trusted partner in Miami real estate
Search by neighborhood...
Featured Listings
NEW
$1,250,000
Coral Gables Villa
4 bd|3 ba|2,400 sqft
OPEN
$890,000
Brickell Condo
2 bd|2 ba|1,200 sqft
Get in Touch
02

Build the Answer and Action Boundaries

Ask Playcode to build source retrieval, model-provider adapter, output validation, injection-resistant instruction hierarchy, rate limits, redacted audit events, and a separate confirmation-gated path for every tool action.

The model proposes; server code validates identity, permission, target, current state, idempotency, and confirmation before any side effect.

03

Test Abuse, Outages, and Human Handoff

Exercise known and unknown questions, conflicting sources, malicious instructions in user and retrieved text, PII, repeated requests, provider failure, dangerous actions, and escalation destination failure.

Publish only after the final HTTPS environment produces citations, uncertainty, safe retries, bounded logs, and a handoff a real person can receive.

Your website is live!

https://
miami-homes.playcode.io
Site is secure - SSL certificate active
Or connect your domainmiami-dream-homes.com
Miami Dream Homes
Your trusted partner
Miami Dream Homes
Your trusted partner
Replace fluent guessing

Make Every Answer Traceable or Explicitly Uncertain

A chatbot that sounds right

  • The prompt says be helpful but names no approved sources
  • Unknown questions produce plausible invented answers
  • Retrieved instructions can override the application boundary
  • The model can call external actions directly
  • Raw transcripts and personal data appear in routine logs

A chatbot that can be checked

  • Every grounded answer returns source IDs and links
  • Missing evidence becomes uncertainty plus human escalation
  • User and retrieved text are treated as untrusted data
  • Actions are typed proposals until server authorization and confirmation
  • Rate limits, retention, redaction, abuse, and outage states are testable
The application around the model

What Your AI Chatbot Can Include

01

Approved source registry

Versioned documents with owners, scope, effective dates, stable source IDs, access rules, update expectations, and a deletion path.

02

Grounded answer contract

A validated response shape containing answer, citations, confidence or evidence state, escalation choice, and no unsupported claim when the sources are silent.

03

Prompt-injection resistance

A fixed instruction hierarchy that treats user text and retrieved text as data, validates provider output, and never accepts source content as authority to reveal secrets or execute tools.

04

Safe action proposals

Typed previews for returns, bookings, tickets, or messages with exact targets and consequences, followed by server-side authorization, confirmation, and idempotent execution.

05

Privacy and abuse controls

PII minimization, redacted logs, retention and deletion rules, session and identity rate limits, content-abuse handling, and bounded audit events.

06

Failure and escalation paths

Explicit states for no grounding, conflicting sources, provider timeout, tool failure, handoff failure, rate limiting, and a real human owner who can continue the case.

A practical chatbot implementation guide

Build the Truth Boundary Before the Conversational Polish

The first useful release answers a small set well, refuses to guess, and hands the case to a person when evidence or automation ends.

Before you build

Prerequisites

  • A versioned approved-source set with an owner

    Grounding cannot be tested when the chatbot silently searches everything or no one owns outdated and conflicting policies.

    Ready when: Every source has an ID, title, location, access rule, effective date, content owner, and one known question it should answer.

  • An escalation destination with service expectations

    Uncertainty is only safe and useful when a real person or queue can receive enough context to continue without asking the user to start over.

    Ready when: A fictional unknown question creates a redacted handoff record in the intended queue and a person can identify the next action.

  • A privacy, retention, rate-limit, and abuse policy

    Chat messages can contain personal, regulated, hostile, or secret data, while provider and support costs can be abused by repeated traffic.

    Ready when: The brief names stored fields, purpose, retention, deletion owner, redaction, allowed users, per-boundary limits, and safety escalation rules.

Implementation sequence

Do, observe, verify
  1. 01
    Playcode project data model and source registry

    Create the source and answer contracts

    Define source ID, title, content, version, effective date, owner, access scope, and status. Define answer, citations, evidence state, escalation, and optional action proposal as a validated server response.

    Expected result

    Every answer can be traced to the exact approved source version or returns an explicit no-grounding state.

    Verify

    Run known, unknown, conflicting, stale, inaccessible, and deleted-source fixtures and inspect the complete response object.

  2. 02
    Server-side retrieval service

    Add retrieval behind authorization

    Resolve user and source scope on the server, retrieve a small bounded set, attach stable citations, and never let a browser-supplied role or source ID expand access. Keep secrets and private records out of prompt and logs unless strictly required.

    Expected result

    The same question returns only sources the current user may access, and cross-account requests reveal neither content nor source existence.

    Verify

    Repeat source search and answer requests signed out, in another account, after a role downgrade, and with a known private source ID.

  3. 03
    Server-only provider adapter

    Wrap the model provider with a strict adapter

    Label approved sources and the user message as untrusted data, require citations, validate provider output, set time and token budgets, and convert missing evidence, malformed output, refusal, timeout, or outage into explicit application states.

    Expected result

    Changing providers or models does not change the surrounding answer, uncertainty, citation, privacy, rate-limit, or escalation contract.

    Verify

    Run the same fixtures through a fake deterministic provider, the chosen provider, malformed output, timeout, and error responses.

  4. 04
    Tool proposal and confirmed server action

    Separate proposals from side effects

    Let the model propose only an allowlisted typed action. Show the user the exact target and consequence, obtain confirmation, then re-check identity, permission, current state, and idempotency in server code before executing.

    Expected result

    Prompt injection, duplicated replies, stale confirmations, and provider retries cannot create an unauthorized or repeated external action.

    Verify

    Test no confirmation, wrong user, changed order state, repeated idempotency key, provider success plus local failure, and an injected execute-now instruction.

  5. 05
    Published HTTPS chatbot and support queue

    Publish and prove the real handoff

    Deploy the exact tested version, configure bounded monitoring and alerts, then run a fictional end-to-end conversation for a cited answer, unknown question, injection, rate limit, provider outage, action proposal, and human escalation.

    Expected result

    The production path answers from approved evidence, fails closed, protects private data, and gives a person enough redacted context to continue.

    Verify

    Inspect browser, server, provider, tool, and queue records by correlation ID without raw secrets or unnecessary transcript content.

Decisions that change the build

Should the chatbot answer from a curated source set or broad search?

  • Curated approved sources
  • Broad web or workspace search

Choose: Start with curated approved sources for support and operational answers. Expand retrieval only when source ownership, authorization, freshness, and citation quality remain explicit.

Tradeoff: Curated sources cover fewer questions but are easier to verify and maintain. Broad search increases recall while raising injection, privacy, staleness, licensing, and contradiction risk.

Should the model execute support actions directly?

  • Proposal plus explicit confirmation
  • Direct action from the model response

Choose: Use a proposal plus server-confirmed action for any ticket, message, return, booking, refund, account, or data change. Keep the preview and execution tools separate.

Tradeoff: Confirmation adds one step, but it makes the target and consequence visible and gives server code a reliable point to authorize, deduplicate, and audit.

Before you share it

Test checklist

  • Happy path

    Ask a known question from an authorized account and then request a typed action proposal with a fictional target.

    Expected: The answer cites the approved source, the proposal names the exact target and consequence, and no action executes before confirmation.

  • Invalid input

    Ask an unknown question, use an inaccessible source ID, submit malformed provider output, and request a non-allowlisted action.

    Expected: The chatbot states uncertainty or denies access, returns no unsupported citation, validates the output, and offers a human handoff without a side effect.

  • Duplicate or retry

    Repeat one message through a provider timeout, retry the same confirmed action key, and resend the same escalation after an interrupted response.

    Expected: The user receives one stable result, the action executes at most once, and the support queue has one reviewable case.

  • Published smoke test

    On the published chatbot, run cited, unknown, injected, abusive, rate-limited, provider-outage, action-proposal, and human-handoff conversations.

    Expected: The final environment preserves citations, explicit uncertainty, safe tools, privacy, rate limits, failure states, and a working escalation destination.

If something goes wrong

Common failure cases

The chatbot gives a confident answer without a citation

Likely cause
The provider response bypassed output validation or the application treated general model knowledge as an approved source.
Check
Inspect retrieved source IDs, provider input, validated output, and the final response for the exact question.
Fix
Require citations to approved source IDs, convert missing evidence to uncertainty, and add the question as a regression test before re-enabling the answer.

A retrieved document changes the chatbot rules

Likely cause
Source text was concatenated as instruction authority or the model was allowed to produce an unvalidated action directly.
Check
Add a controlled malicious instruction to a fixture source and compare provider request boundaries, response, proposal, and server action logs.
Fix
Treat source and user text as untrusted data, validate output, allowlist action types, and keep server authorization and confirmation outside the model.

Users cannot reach a person after an unknown answer

Likely cause
The UI offered escalation but the queue destination, authentication, notification, ownership, or retry path was not exercised end to end.
Check
Trace one fictional escalation from the browser through the server into the queue and verify it is visible to the assigned owner.
Fix
Persist an idempotent handoff record before confirming success, alert on delivery failure, and provide a retry or alternate contact path.

Private text appears in provider or application logs

Likely cause
Raw prompts and retrieval payloads were logged by default or redaction occurred after the logging middleware.
Check
Search bounded logs for fixture email, phone, secret, and source markers across browser, server, provider, and queue layers.
Fix
Minimize before sending, redact before logging, disable raw body logs, shorten retention, restrict access, and document deletion and incident handling.
Example interface direction

Show the Answer, Citation, and Human Exit Together

A useful support interface makes the evidence and escalation route visible instead of hiding them behind fluent text.

Illustrative support chatbot concept with a cited returns answer, source panel, and human escalation control
Illustrative conceptIllustrative UI concept, not a product screenshot. The actual result depends on your brief, approved sources, model provider, and support workflow.
Provider-neutral working artifact

Inspect the Grounding, Safety, and Failure Tests

The support-triage package proves the application boundary without credentials or a hidden provider dependency.

9/9
grounding, privacy, abuse, tool, limit, and outage tests passing
2
versionable fictional approved source files
0
model credentials or external side effects required

Download the provider-neutral support chatbot ZIP

Choose the evidence boundary

Curated Support Bot or Broad General Assistant?

What you needBroad general assistantGrounded support chatbot
KnowledgeGeneral model knowledge plus broad retrievalVersioned approved sources with owners and stable citation IDs
Unknown answerMay produce a plausible continuationExplicit uncertainty and a visible human escalation route
ActionsTool calls may be coupled to conversation outputTyped proposal, confirmation, server authorization, idempotent execution
OperationsPrompt and provider metrics aloneSource freshness, citation coverage, handoff, abuse, limits, failures, privacy, and recovery
Bounded chatbot jobs

Start Where the Sources and Escalation Owner Are Clear

Customer Support

Cited policy answers

Answer shipping, returns, account, or product questions from owned policies and route missing evidence or special cases to the support queue.

Internal Help Desk

Role-scoped guidance

Retrieve only procedures the signed-in employee may access, cite the exact document, and send unresolved cases to the right internal owner.

Product Guide

Feature answers with sources

Explain supported product behavior from current docs and keep roadmap, pricing, or account-specific questions behind current authoritative sources.

Service Intake

Structured handoff

Collect the minimum facts, show consent and retention, propose the next action, and create a human-reviewed request only after confirmation.

The engine

A software team in one agent

Playcode AI runs the same frontier models that power ChatGPT and Claude - and orchestrates them like a team: it plans the work, delegates to sub-agents, runs long jobs in the background, and reviews its own changes.

Every frontier model

The latest models from every major lab, in one picker. Switch anytime.

ClaudeGPTGeminiGrok

Sub-agents

Big jobs split across specialists - one explores your code, one writes, one audits - working in parallel.

Background tasks

Long builds and migrations keep running while you keep talking. They report back when done.

migration - running

It sees your designs

Have a design in Figma? Paste a screenshot - or a page you like, or a bug - and it builds from what it sees.

Tuned by years of iteration to write production software - structured, typed, maintainable - not throwaway prototypes.

Production code

Real code you can open, read, and edit

Under every project is a codebase the agent keeps production-grade. And you are never locked out of it: open the file explorer and edit any file yourself - server included. No AI required.

Quality is the default

Complete states, secure boundaries, structured code - the bar is what a professional agency would ship.

Every file is yours to open

Browse the whole project - frontend, backend, configuration - and edit directly in the built-in editor.

Developers are welcome

Invite your developer with the right role, or export the project code and files. Nothing is trapped in Playcode.

For teams and organizations

Share it like you share a doc

Playcode is built for organizations, not just solo builders. Workspaces hold your projects, people, and billing; roles and teams decide exactly who sees what.

Workspaces with their own billing

Create a workspace per company, client, or department - each with its own members, projects, and subscription.

Three clear roles

Admins manage, editors build, viewers watch. Set roles on the whole workspace, on a team, or on a single project.

Teams that scope access

Group people into teams and give each team its own projects - or a whole folder of them. Private projects stay private, even inside a shared workspace.

Share outside the workspace

Send a project to any email - a client, a contractor - with exactly the access you choose. No extra seat needed.

Share "Inventory tracker"
Acme Ops workspace
Private
client@partner.co
Editor Invite
Operations team
8 people
Editor
MK
Maya Kowalski
maya@acme.co
ADMIN
JR
Jon Reyes
jon@acme.co
EDITOR
LS
Lena Sato
lena@acme.co
VIEWER
Restricted - only people invited can open it
Real-time

Multiplayer by default

Work on one project together. Write in the same AI chat together. Every message, edit, and setting is fully synchronized - live, for everyone, on every device.

acme-launch · Playcode
MKJR
Maya's laptop
Maya
Add a 'Book a call' button to the hero
Jon
And link it to our calendar, please
Playcode AIDone - button added to the hero and linked to your calendar.Preview updated
Describe the next change...
Jon's phone
Maya
Add a 'Book a call' button to the hero
Jon
And link it to our calendar, please
Playcode AIDone - button added to the hero and linked to your calendar.
Message...
Already there. No refresh.

Same project, same moment

Everyone works in the project at the same time - even in the code editor - without lock-outs or "who has the latest version".

Every device

Start on the laptop, check from your phone: the same live state follows you everywhere you sign in.

Nothing to refresh

Changes arrive over a live connection the instant they happen. Reloading the page is a habit you can drop.

Feels instant

It never makes you wait

Instant

Every click applies immediately on your device. Syncing happens behind you, not in front of you.

Offline

Connection dropped? Your changes queue locally and replay the moment you are back.

Reload-proof

The queue survives closing the tab. Nothing you did is lost while you are away.

Real People. Real Websites Built with AI.

"I built my entire portfolio site in 20 minutes. My clients think I hired a designer. Already got 3 new inquiries this month."
Marcus T. · Freelance Graphic Designer, Austin TX
"We switched from Wix to Playcode. The AI actually understands what we need instead of giving us cookie-cutter templates. Saved us thousands."
Sarah Chen · Owner, Bloom & Petal Floristry
"I update my property listings from my phone while showing apartments. Clients are impressed when I tell them I built the site myself."
David Morales · Real Estate Agent, Miami FL

Simple Pricing

Start small. Upgrade when you're ready.

Starter

$0to start

Try the AI website builder and see results

  • AI credits included to start
  • Publish your site instantly
  • Subdomain included
  • Website hosting included
  • No credit card required
Get Started

Pro

Most Popular
$21/month

Everything you need to build

  • 100 AI credits/month
  • All AI models (12+)
  • Visual editing
  • Custom domains
  • Website hosting included
  • Export your code anytime
  • Private projects
  • Unlimited collaborators
Start Building

Cancel anytime. No hidden fees.

How credits work

Credits are used when AI helps you. Simple edits cost less, complex features cost more.

  • "Change button color" ~0.3-0.5 credits
  • "Add a contact page" ~2-3 credits
  • "Build full landing page" ~5-10 credits

Most users never run out. 100 credits = lots of building.

AI Chatbot Builder Questions

This page describes the application built around the model provider you choose. Playcode can build the interface, server, approved-source retrieval, provider adapter, validation, tests, and hosting. The provider account, current API terms, credentials, usage cost, and model behavior remain separate boundaries.

Restrict answers to approved retrieved sources, require stable citations, validate the output, and make no grounding a first-class uncertain state with human escalation. No design eliminates every model error, so regression tests and production review remain necessary.

Prompt injection is untrusted user or retrieved text that tries to change instructions, reveal secrets, expand access, or execute actions. Treat all external text as data, keep authority in server code, validate output, minimize tools, and require confirmation for side effects.

Yes, when the target system has a supported API and you provide credentials and policy. Use separate proposal and execution steps: show the exact target and consequence, obtain confirmation, then authorize and execute idempotently on the server.

Store only what the business purpose requires. Decide fields, consent, access, retention, deletion, export, incident response, and provider handling before launch. Prefer redacted structured events over raw transcripts in routine logs, and never log credentials.

Fail closed: show that the answer service is unavailable, do not invent an answer, preserve any user work safely, offer a real human or retry path, and alert on repeated failures. Provider fallback must preserve the same grounding and privacy contract.

Use limits at the boundaries that match abuse and cost: session, authenticated user, IP or risk signal, provider budget, and external tool. Return a clear retry or human path, and avoid exposing exact thresholds when that would make abuse easier.

Pass known, unknown, conflicting-source, inaccessible-source, injection, malformed-output, PII, abuse, rate-limit, provider-outage, duplicate-action, and human-handoff tests. Then repeat a fictional end-to-end smoke test in the published HTTPS environment with bounded logs and a real escalation owner.

Still have questions? Contact us

Build the workflow around the conversation

AI app builder Build the chatbot interface, server, data model, and tests together.How to create an AI chatbot Follow the grounded retrieval, testing, safety, and operations guide.All Playcode builders Compare app, website, bot, and business-workflow paths.Client portal builder Put authenticated support conversations inside a private customer workspace.AI CRM builder Connect reviewed support or qualification handoffs to an owned workflow.

Define the sources. Make uncertainty visible.

Start with one support job, a small approved source set, stable citations, a real human destination, and no direct model side effects.

A support chatbot that cites our return and shipping policies and never guesses...Build My AI Chatbot

No credit card required. AI credits included.