Frequently Asked Questions (F.A.Q.)

JSFE is a deterministic workflow engine that takes over from the AI for any agentic action. Each “flow” is a self-contained execution unit that provides the user-facing prompts, collects user input, and uses conventional programming (conditionals, loops, tool calls) to control execution, exactly like a traditional application would. The AI generates the natural-language wording of prompts inside a flow but never decides what to call, when, or with what parameters. Across 200,000+ flows including 100,000+ phone calls in production, 100% of business flows (payment, account inquiries, product inventory queries, dispute tickets creation) completed without any error, exactly as scripted.

Why deterministic flows instead of letting the AI choose tools at runtime? Do AI Assistants Really Need Tools? walks through the engineering case in detail.

By removing the AI from action execution. Traditional “agentic” AI systems let the LLM choose what tools to call and what parameters to pass, which is where hallucination becomes dangerous (wrong charges, corrupted state, invented data). instantAIguru routes every action through JSFE's deterministic flows, where conventional code controls the call, the parameters, and the user-facing prompts. Hallucination is eliminated by construction: there is no AI in the path that commits the action. (Conversational accuracy is a separate guarantee, covered in the next question.)

We argued the broader case for this architectural pattern in Agentic AI Can Be Hallucination-Proof.

Three layers, validated independently:

• Hybrid RAG combines dense vector retrieval (semantic match) with sparse keyword retrieval (exact-term match) on AWS OpenSearch. This catches both what the customer means AND the exact SKU or policy clause they referenced.
• Multi-vendor model orchestration routes each role in the answer pipeline (intent classification, retrieval reranking, draft generation, validation) to the model best suited for it. Frontier labs in production: OpenAI, Anthropic, Google, Meta, Groq, DeepSeek.
• Automatic hallucination detection plus cross-vendor escalation runs every draft through a validation loop (fact-checking against retrieved sources, confidence scoring, citation matching). When validation fails, the system escalates to an intervention model from a competing vendor and retries. The cross-vendor switch is deliberate: re-prompting the same model often produces the same wrong answer (a “cache trap” of cached reasoning patterns); switching vendors breaks that pattern.

The 97%+ number was first established in May 2024 from a 30,000-question review by Omie's tier-1 support engineers. Curacao Department Stores re-substantiated it across 100,000+ monthly interactions, validated by the Customer Service team led by SVP Joseph Jiron.

Why FAQ-based answer systems hit a ceiling well below this accuracy: Your FAQ Page is Costing You Customers.

99%+ measured at Curacao in April 2026. instantAIguru runs intent detection through two top models from different vendors in parallel as a robustness measure: detection has a strict response-time budget, and routing each request to a second vendor in parallel means a transient provider slowdown never causes a missed intent, driving failed detections to near zero. Once an intent is identified and a JSFE flow is entered, every step prompts the user explicitly. A real example from production:

“To send you a payment link I need to locate your account. To use your caller ID please say YES. Otherwise please enter or say the account number, or either the account's phone or email.”

The user retains control until the moment of commit. In the rare 1% where intent is not confidently detected, no scripted action runs at all. Instead, the conversational AI answers (for example, “Here's how to make a payment” rather than executing one). The architectural guarantee: a wrong action is never committed silently.

Phone calls flow through Twilio Conversation Relay, a WebSocket bridge that streams speech-to-text into the same Hybrid RAG plus JSFE pipeline that handles web, email, SMS, and WhatsApp, then streams text-to-speech back to the caller. Sub-300ms response latency in normal conditions. At Curacao in April 2026, 85%+ of inbound phone calls were resolved without transfer to a live agent.

Each channel connects from the admin portal after your Guru is live on your website. For WhatsApp, voice, and SMS you own the underlying Meta and Twilio accounts directly (BYOC), so instantAIguru never sits between you and the carrier and you pay no usage premium to us. Email is even simpler: instantAIguru provisions a dedicated inbox address for you instantly, with no external account required.

WhatsApp

instantAIguru is a certified Meta WhatsApp Tech Partner with an approved app. The connection has two phases. First, on Meta directly: create a WhatsApp Business Account (WABA) and verify a mobile phone number. Second, in the instantAIguru admin portal: launch Meta's Embedded Login (Meta's standardized authorization flow) and authorize our certified app to use your verified WABA number. Your Guru is live on WhatsApp immediately. There are no per-message charges, including no Meta fees: the Guru only responds to user-initiated questions, and Meta allows free responses within the 24-hour conversation window. You pay only your flat SaaS fee per connected channel to instantAIguru.

Voice (phone)

Create a Twilio account, provision a phone number, and enter your Twilio credentials (Account SID and Auth Token) in the admin panel. Voice connects instantly. You pay Twilio directly for the number and per-minute calls; instantAIguru charges only the flat SaaS fee per connected channel.

SMS

Same Twilio flow as voice, plus one US-carrier requirement: you must submit and pass an SMS Campaign application before your number can send messages. Once approved, enter the Twilio credentials in the admin panel and you're live. Same model: you pay Twilio for messages, instantAIguru charges only the flat SaaS fee per connected channel.

Email

Click Add Email Channel in the admin portal. instantAIguru auto-provisions a dedicated inbound address for you, with no external account needed. Most customers don't change their public-facing email; they configure their existing email provider (Google Workspace, Microsoft 365, etc.) to forward all support@ mail to the provisioned address. Replies thread properly via RFC headers so they land in the original conversation in your customer's inbox, and an AI-disclosure footer is appended for compliance with California's BOT law and the EU AI Act. You can add multiple email channels per Guru (e.g., one each for support, sales, and billing) by setting different Department values.

For the strategic case for owning your own carrier numbers, including what happens to your customer relationships when a platform owns them instead, see Bring Your Own Numbers.

Two reasons: cost predictability and risk isolation.

Cost predictability

Platforms that resell WhatsApp, voice, or SMS as a bundled service typically markup carrier costs and add per-message or per-minute fees. With instantAIguru's Bring Your Own Channels (BYOC) model, you pay carrier rates directly (Twilio for voice and SMS, no Meta fees on user-initiated WhatsApp conversations within the 24-hour window) plus our flat SaaS fee per connected channel. No usage-based markup, ever.

Risk isolation

When a platform owns your WABA or your phone number, leaving the platform means losing the number, the conversation history, the customer associations, and the sender reputation tied to it. With BYOC, the WABA and Twilio number are yours; instantAIguru is just the application authorized to use them. If you ever disconnect, your number, your customers, and your reputation come with you.

For a deeper take on the risks of platform-owned numbers, see Bring Your Own Numbers.

Yes, every channel, not just web chat. Web chat works the moment you paste the snippet on your site. During the trial you can also connect your own Twilio account (voice and SMS) and your WhatsApp Business account from the admin portal and test them live against your own content, and email runs through a dedicated inbox we provision instantly that you forward your support mail to. You can even configure live-agent handoff and Live Agent service hours during the trial. The trial is 7 days with no credit card (extending to 14 if you add a Stripe payment method), so you evaluate the full omnichannel experience before paying anything.

The Guru is a responsive assistant: it answers messages your customers start, across web chat, email, phone, SMS, and WhatsApp. The only message it can initiate is an optional CSAT survey invite. It does not run proactive outbound or marketing campaigns, which also keeps you on the right side of messaging rules around unsolicited contact.

Each channel keeps its own conversation history. When a conversation is explicitly moved between channels the history carries with it: a phone call can offer to continue by text, or a text can invite the customer to continue on WhatsApp, and the Guru picks up where it left off. Phone calls start fresh each time by design. Your site can also pass a signed-in customer's context, and a JSFE integration can pull records from your CRM. Channels link when the move is explicit, not by guessing identities automatically.

Yes, directly and live. You connect your store through a custom app authorized with your Shopify admin credentials, and the Guru works with your real store data in real time: product search and availability, order status and tracking, and your store policies. Order-specific details are gated behind verification, the customer confirms identity with a one-time code sent to their phone or email first, so account information only ever reaches the right person. A shopper asking where their order is gets a true, current answer from Shopify rather than a guess.

After you authorize a custom app with your Shopify admin credentials, the Guru has read-only access to order tracking and status, product details and availability, inventory (per location where applicable), store policies, and store locations and directions. It can also recommend products from your catalog, drawn from what it can read. Order-specific details are released only after the customer verifies identity with a one-time code.

Connections to your other systems are built as deterministic flows in our integrated Flow Designer, and we expose a REST API, webhooks, and an MCP endpoint to wire them up. This is production-proven: a Microsoft Dynamics 365 CRM integration runs live for Curacao. We do not ship a brittle pre-labeled connector for every tool; instead we scope the specific action you need, a status lookup, a record write, a custom workflow, during onboarding, so the integration fits your stack rather than forcing your stack to fit a template. Because these run on the JavaScript Flow Engine, the action executes as conventional scripted code with the AI removed from the commit.

Yes, through JSFE deterministic flows rather than the conversational model, so a transaction runs as scripted code with zero hallucination on the commit. In production at Curacao, the Guru collected 105 payments in a single weekend and sent hundreds of outbound payment links by SMS and email. Any action that touches money or records, a payment, an account change, a structured return, is built and scoped deliberately as a flow, never left to the model's discretion.

On phone calls and web chats, a conversation can transfer to a live agent, and you configure the live-support contact details for a direct handoff. The Guru is built to recognize the edge of its competence: it escalates when it cannot source an answer, on detected customer frustration, or on the topics you mark as human-only, and it can switch a conversation to text to continue by SMS. You set the rules.

Live Agent coverage is configurable: you set service hours and agent scheduling per phone line, per day of the week, so out-of-hours calls are handled the way you choose rather than ringing an empty desk.

Yes. The admin console supports role-based access with per-user scopes, so you decide who on your team can see and change what. Access to customer data is restricted to authorized users and is logged.

Yes, through the open-source CSAT survey and analytics platform we provide. You launch it on your own infrastructure, then enable the integration from the Guru admin portal: paste your CSAT site URL and token secret, and choose a delivery method (SMS through your own Twilio number, email, or both). After a call, the Guru sends the customer a one-tap survey link by text or email, and the responses, CSAT, NPS, and verbatim comments live in your own self-hosted CSAT dashboard, so the satisfaction data stays in your hands.

Every conversation is logged in the Chat Dashboard, a complete view of what your Guru has discussed with visitors, with search, filtering, and tagging, a per-conversation quality rating, and the response time for each answer. It loads new logs incrementally and auto-refreshes while you watch a live window, and a separate Analytics Dashboard summarizes the patterns, what customers are asking and where your content has gaps, so you can both spot-check individual answers and act on the trends.

Yes, from the Chat Dashboard, in three formats, each respecting your current filters and visible columns: CSV, a pretty-printed JSON array (handy for programmatic analysis), and a styled Excel workbook with a frozen header row and auto-sized columns (best for sharing and review). Your conversation data is yours to take with you.

Yes. The web chat widget lets a customer rate each answer, turn by turn, and add a note, and they can undo or change a rating at any point, so the signal stays honest rather than locked in on a misclick. Feedback is captured at the level of the individual response rather than only as a session-end survey, which is the most precise kind there is: it tells you exactly which reply landed and which one missed, often in the customer's own words.

Every rating and note flows into the Chat Dashboard as a per-answer quality mark (correct, incorrect, or unmarked) and into the Analytics Dashboard's accuracy view. You can filter straight to the answers customers flagged and, for each one, inspect the sources the Guru retrieved and the model it used, then fix the underlying content or tuning. That turns real customer feedback into a measurable improvement loop rather than guesswork, and it is the same kind of answer-level review that underpins our independently validated accuracy.

Yes. The chat widget gives the visitor a download button that saves their full conversation as a JSON file they keep, and a delete button that clears it. The person you are helping stays in control of their own transcript, which is both a courtesy and a clean answer for a privacy-conscious customer.

Yes, the Guru answers 24/7, every day of the year. Separately, you set live-agent service hours, by phone line and by day, that decide when the Guru offers or routes a conversation to a human. Outside those hours the Guru keeps answering on its own, so customers are never left waiting.

Yes. On your website or chat widget, a Live Connect icon appears in the widget header whenever an agent is available, and the visitor taps it to reach a live agent over audio, video, or chat. Agents choose when to be available or pause, from any web-enabled device including mobile, and Live Connect runs on our managed infrastructure by default or fully self-hosted (it is open source). On the phone channel, when a line has a Live Agent phone number configured, a caller who asks for a person is handed to that agent during the line's defined live-agent service hours. The AI Guru handles the volume, and a human steps in when one is available.

Yes. Choose from eight preset color themes or build your own with custom color pickers, primary, background, gradient, in-color text, and hover, so the widget's header, send button, and accents match your site. A live preview panel reflects every change as you make it, so you tune the look before it goes live.

Yes. Upload a custom image to replace the default chat bubble on the widget launcher; it previews immediately and you save to apply, and you can clear it to revert to the default at any time. If you set an icon directly in the embed code, that takes priority over the one configured in the panel.

Yes. You can set up to four chat starters, suggested opening questions that help a visitor begin, and add teaser text bubbles that appear near the widget. An optional visual effect, such as a pulsing glow on the launcher with a delay and duration you set, draws the eye without being intrusive. As with theming, you preview it all live before publishing.

Yes. You write the voice in a single free-text field: describe the tone you want, give examples of a good reply, and add any directions that matter, including language-specific guidance. There is no rigid form, you describe the character in your own words and the Guru speaks in it across every channel. This is on top of visual branding: your colors, logo, launcher icon, and chat starters.

Over 100 languages on text channels (web, email, SMS, WhatsApp). Voice supports the major commercial languages (English, Spanish, French, Portuguese, German, Italian, Arabic, Mandarin, Hindi, Japanese, and others) with continuous expansion. The same Hybrid RAG knowledge base serves every language without per-language re-indexing.

The Guru detects the language of each incoming message and replies in it, and it can switch mid-conversation if the customer does. Language handling is AI-driven and best-effort, following the model's reading of each message rather than a fixed rule. One knowledge base serves every language, so you never maintain translated copies.

Guard rails. When the retrieval pipeline finds no relevant references for a question, the response is required to start with a transparent disclosure, for example:

“I was not able to find any document that addresses your question. Here is the answer I generated without reference documents.”

Customers can also configure live-support contact details for a direct handoff, and on phone calls and web chats the conversation can transfer to a live agent. The architectural rule: never improvise medical, legal, or financial claims, never invent product attributes that are not in the indexed source, and always disclose when reference documents were not used.

About five minutes for a typical small site (under 50 pages). Larger sites take roughly one hour per 1,000 crawled pages, depending on source-content complexity. The self-serve flow: paste your URL, and the AI customer service agent goes live as soon as the crawl completes. WhatsApp, voice, SMS, and email channels connect later from the admin portal (see the channel-connection question above). Custom JSFE workflows for proprietary CRM, niche commerce platforms, or industry-specific compliance are scoped during onboarding.

Yes. Your website crawl is the starting point, not the limit. From the Knowledge Files tab you can upload individual files or an entire folder of documents, internal policies, spec sheets, manuals, anything you want the Guru to know, and they become part of the same knowledge base it answers from (supported file formats depend on your plan, and files upload to secure cloud storage). You can also enable Crawl External Pages to let the crawler follow links into other domains. After you add or remove files you re-index, and the Guru incorporates the changes. So private or unpublished material can inform its answers without ever appearing on your public website.

The Guru learns from three sources that work together. First, your website: paste your URL and it crawls your site (running JavaScript so dynamically rendered content is captured) and indexes the text. Second, Knowledge Files you upload, available on every plan, in HTML, PDF, DOC/DOCX, XLS/XLSX, Markdown, CSV, and TXT. Third, the Instructions field: a powerful free-text source where you state authoritative facts, answers, policies, and directions in your own words, which strongly shape how the Guru responds. When you change your site content or files you re-index, and the next answer reflects it.

Yes, at a few levels. In the admin you can turn on Crawl External Pages, which also indexes the external pages your site links out to, not just your own. For finer control, additional options are available on request through support: a hard limit on the number of indexed pages, path exclusions, additional seed domains, and fenced paths. And scheduled re-crawls, which refresh your knowledge base automatically, are available through the API.

Three architectural facts shape it:

• No model training on your data. All AI vendors we use (OpenAI API, Anthropic via AWS Bedrock, Google Gemini API, Groq, Cloudflare Workers AI) operate under their respective API “no-training” terms. Your customer conversations are never used to train any AI model, by us or by any vendor.
• Encrypted storage on AWS DynamoDB. Conversation history is the only customer-related data we persist. It is stored in AWS DynamoDB with encryption at rest (AWS-managed keys, AES-256) and encryption in transit (TLS 1.2+). All other data flows are stateless: requests to AI vendors complete and exit, with no intermediate storage.
• Least-privilege access. Access to customer data is restricted to limited certified personnel under documented access controls. Production access is authenticated and logged.

You retain control of your conversation history. From the admin portal you can configure retention duration and request erasure of any subset (a specific user, a date range, or all history) at any time.

We are committed to supporting customers' obligations under both.

• GDPR compliant. We follow GDPR data-handling principles: no model training on customer data, encryption at rest and in transit, user-controlled retention and erasure, named sub-processor disclosure, and data-minimization. Data is hosted on enterprise-grade AWS infrastructure in the us-east-1 region (Northern Virginia, USA).
• HIPAA compliant. Our infrastructure follows HIPAA controls (encryption, access logging, minimum-necessary access). For workloads involving Protected Health Information, we work with customers to design appropriate data flows, including BAA-covered vendor routing through AWS Bedrock or OpenAI Enterprise as needed.

Regular internal security audits are performed.

ISO/IEC 42001 is the international management-system standard for AI governance. We are tracking 42001 controls (AI lifecycle documentation, risk assessment, vendor oversight, incident response, bias and quality monitoring), and certification is on our 2026 compliance roadmap.

For procurement reviews evaluating 42001 posture today: existing controls map onto the standard's core requirements.

• AI lifecycle. Versioned model and pipeline configuration, change-controlled deploys, named sub-processor disclosure.
• Risk and quality. Continuous accuracy monitoring (97%+ measured at Omie 2024 and Curacao 2025), zero-hallucination architecture for agentic actions via the JavaScript Flow Engine, named-validator substantiation.
• Data governance. No model training on customer data, encryption at rest and in transit, user-controlled retention and erasure (see GDPR question above).
• Vendor oversight. Multi-vendor frontier-AI orchestration (OpenAI, Anthropic, Google, Meta, Groq) under stateless inference terms.

Detailed mappings for 42001 questionnaires are available under NDA on request to [email protected].

Sub-processors and their roles:

• AWS primary cloud infrastructure (DynamoDB for conversation storage, Bedrock for Anthropic model access, EC2 for application hosting, OpenSearch for retrieval indexing).
• OpenAI model API (intent classification, generation, validation).
• Google Gemini model API (model orchestration, retrieval reranking).
• Groq model API (low-latency inference for selected pipeline stages).
• Cloudflare CDN, edge caching, Workers AI for selected inference, security and WAF.
• Stripe payment processing for our SaaS subscription billing.

For customers using BYOC channels (WhatsApp, voice, SMS), Meta and Twilio are not our sub-processors: those accounts are owned by the customer directly, and we are the application authorized to use them on the customer's behalf.

All AI-vendor APIs operate under their respective “no-training” terms. AWS, Cloudflare, and Stripe are SOC 2 Type II certified.

Conversation history is stored in AWS DynamoDB in the us-east-1 region (Northern Virginia, USA). Configuration data (your Guru's settings, channel connections, knowledge base index pointers) is co-located. No data is replicated to other regions, so you always know exactly where your data lives.

Access to customer data is restricted to limited certified personnel under documented least-privilege controls. Production access requires explicit authentication and is logged. AI vendor APIs operate stateless: requests are sent for inference and exit immediately, with no persistent vendor-side storage of your data.

Yes. From the admin portal you can configure conversation-history retention duration and request erasure of any subset (a specific user, a date range, or all history) at any time. For GDPR right-to-erasure or other regulatory requests directed at a specific data subject, contact [email protected] and the request is honored within standard regulatory timelines.

The 97%+ figure measures conversational accuracy: the percentage of customer questions for which our pipeline returned a factually correct answer drawn from the customer's indexed knowledge base.

How it was established

• First validation (Omie, May 2024). Omie's tier-1 support engineering team reviewed 30,000+ Guru-generated answers and confirmed accuracy above 97%. This validation was the basis for production go-live on 2024-05-22.
• Re-substantiation (Curacao, October 2025). Subsequent review of 100,000+ monthly customer interactions at Curacao Department Stores confirmed the same threshold. Validated by the Customer Service team led by SVP Joseph Jiron.

What it does not measure

• It is not a measure of the deterministic action layer (commits, payments, account changes). Those are protected separately by the JavaScript Flow Engine, which removes the AI from the action path entirely. See the zero-hallucination question above for the architectural guarantee.
• It is not a benchmark across all possible questions, only those answerable from the customer's indexed source. Out-of-scope questions are handled by guard rails (see the “what happens when instantAIguru doesn't know the answer” question above).
• Independent industry baselines for comparable single-vendor AI customer service platforms are typically around 90%.

A separate operational metric, resolution rate, was measured at Curacao in April 2026: 85%+ of inbound phone calls handled by the AI were resolved without transfer to a live agent.

Unconditional. If at any time within the first 90 days of paid service you request a refund, you receive a full refund of all SaaS payments made to instantAIguru during that period. No questions asked, no proof of dissatisfaction required, no proration.

The guarantee covers our SaaS fees only. Carrier or platform pass-through costs you incur directly with third parties (Twilio voice and SMS minutes, Meta WhatsApp template message fees if applicable, Stripe processing fees) are paid to those providers and are not refundable through us.

To request a refund, contact [email protected] from the email associated with your subscription.