Haptik vs Moveo.AI (2026)

The integration catalog is documented page by page rather than asserted as a round number: Shopify, Magento, WooCommerce and Fynd for commerce; Stripe, PayPal, PayU, Razorpay and SETU for payments; Zendesk, Freshdesk and Salesforce for service; HubSpot, LeadSquared and Microsoft Dynamics 365 for CRM; Twilio, Exotel and Helo for communications; CleverTap for engagement; and CIBIL for credit data. That is breadth across classes. The Indian market integrations are the distinguishing set: SETU, Razorpay, Exotel, CIBIL and Fynd are the local payment, telephony and credit rails a platform selling to Indian banks and retailers needs. Integrations are write paths, not only lookups: agents create tickets, check inventory, issue payment links and schedule appointments, and Code Step Integration covers custom logic where no connector exists.

AI Agents call tools in the form of webhooks and external MCP servers, which the customer adds and tests, and Dialogs connect to external services via webhooks and run task automation such as sending emails and updating databases. The plans page adds custom webhooks connecting to the customer's own APIs, maintained by Moveo, alongside native Zendesk, Intercom and messaging channels, with CRM systems and internal data sources supplying real-time customer context. Documented custom tool support with authenticated action in outside systems is the threshold on this axis, which is a capability test rather than a catalog-class test, and it is met.

Haptik's agents plan and execute tasks by calling APIs and databases, and the vendor explicitly distinguishes this from language models that respond to prompts rather than executing tasks. The documented flow surface is a step graph: Connections represent the path a conversation takes, each mapped to the user inputs, entity presence or entity values that route to the next step, with Static, Code and Output step types and Code Step and Static Step integrations for custom logic. Execution ends in real actions across systems: ticket creation, inventory checks, payment links and appointment scheduling. No multi agent coordination is documented; this is multi step task execution with conditional routing in a conversational frame. The eighty percent automation figure is the vendor's own reported outcome.

Moveo's Dialogs provide the control-flow primitives this axis looks for, and the vendor documents all four: multi-step workflows guiding users through complex processes, conditional logic creating branching paths based on user input, API integrations connecting to external services via webhooks, and task automation executing specific actions, with the vendor's own examples being sending emails and updating databases. Updating a database is back-office action taking. The architecture is worth carrying because it is a considered design rather than a feature list. Knowledge and guidelines handle most conversational needs, and Dialogs are documented as optional, complementing the knowledge-based approach when deterministic structured workflows are needed. So the platform is probabilistic by default and deterministic where it matters, which is the right split for collections and payment conversations where a wrong step has consequences. On top sits the proactive dimension: agents initiate conversations, manage objections and drive toward an outcome such as completing a payment or activating a card, which is goal-directed multi-step execution rather than turn-taking.

Haptik documents Proactive Messaging that starts conversations without a user message: campaigns sent to a selected audience immediately or on a scheduled custom date, and event based triggers, with SMS for US users and WhatsApp HSM templates for outbound use. The scheduled wake is an outbound message campaign rather than a scheduled agent task. Channel coverage is wide: Web SDK, WhatsApp, Facebook, Instagram, Sunshine Conversation, LINE, Google Business Messages, Telegram, Microsoft Teams, Bot as an API, and iOS and Android SDKs, plus voice with inbound and outbound telephony through an Exotel integration. More than a hundred languages, including twenty two Indian languages, is why the platform works in tier two and tier three markets, where customers use WhatsApp in a regional language rather than a web widget in English.

The standout cell, on channels, proactive initiation and schedules. Channels are broad: chat, voice through telephony, email, SMS, WhatsApp, Facebook Messenger and Viber, plus Zendesk and Intercom as embedded surfaces, across more than twenty languages. Viber in particular signals a vendor that has done the work for its actual markets, Greece and Brazil, rather than shipping a US channel list. The distinguishing property is proactive outbound. Agents initiate conversations rather than waiting, running outbound campaigns the vendor describes as reaching down to a segment of one, to chase a payment, prompt a card's first use or intervene before a customer churns. Outbound initiation is an unusual event class for a conversational platform, and here it is the commercial point of the product: a collections agent that cannot start the conversation is useless. Campaigns can also be scheduled for a future launch with date, time and timezone, retry with a fallback when delivery fails, and follow up when a contact stays silent, which is an autonomous wake on a schedule.

Haptik grounds agents on a maintained knowledge base built from the customer's own material: PDFs, DOCX, PPTX, CSVs and the customer's website, with a product catalog as a separate structured source for recommendations. The vendor states it improves accuracy by combining multiple RAG systems, a considered retrieval architecture rather than a single vector lookup, and that answers stay based solely on provided information. Knowledge Base Integration connects agents to third party knowledge management systems, with documented integrations for Zendesk Guide, Freshdesk KMS, Salesforce KMS and USU Unymira, so the corpus tracks the customer's existing system of record rather than requiring a parallel copy. Per answer source attribution is not documented, a gap regulated buyers may notice.

Persistence is the bar here, not citation-grounded retrieval, and Moveo has a maintained knowledge structure documented first party. The vendor's documentation states AI Agents operate on a knowledge-based system using retrieval augmented generation over a knowledge base holding curated sets of information from various sources, combined with guidelines that shape behavior. That is an explicit RAG architecture rather than context assembled per prompt. The part that settles persistence is the ingestion behavior of Collections, Moveo's knowledge layer: data sources include a website that is re-crawled every 24 hours, a help center, and uploaded CSV, PDF or HTML documents. A corpus that re-crawls its sources on a schedule is a standing structure that tracks what it was built from, which is what separates Full from Partial on this axis. Past chat logs can also be uploaded as grounding material. Worth carrying: grounding here is paired with real-time customer data from CRM and internal systems, so an agent answers from the corpus and personalizes from the customer record. For collections and retention conversations, which is what this product is for, both halves are needed.

Haptik's agents retain context for personalized experiences and handle multi turn conversations, which is within conversation state, and the Archives section preserves transcripts for review. Neither carries what an agent learned in one conversation into the next. The vendor's argument that language models respond to prompts rather than actively executing tasks or learning from interactions implies its own agents learn from interactions, but that is marketing prose about something else, not documentation. The knowledge base persists and counts under knowledge grounding. The documentation is extensive, and a user state or context persistence page could exist.

Session state, with stronger claims in the marketing than in the documentation. The documentation shows conversation context held as variables and user information through a session, visible per turn in the logs, and simulation criteria can assert on the final session context. The vendor's marketing says memory travels with the customer rather than the ticket, and describes agents as self-learning, evolving from every interaction and driving higher automation rates over time. Both are claims without a documented mechanism: nothing read says what is retained across sessions, for how long, whether a customer profile is a queryable object, or how a customer reviews or deletes it. What is corroborated is cross-channel continuity: live agents receive handoffs with full cross-channel context, which requires state to travel with the customer, so the claim is at least partly load-bearing in the product. A memory or customer-profile page in the documentation would settle it.

What is documented is a handoff rather than a gate. Agents escalate high touch scenarios to human counterparts through live chat, with a centralized agent inbox so no interaction is missed and people can jump into a conversation; the Agent Inbox, Live Traffic and Teams sections confirm the surfaces exist. That suits customer experience work, where the failure mode is a frustrated customer stuck with a bot rather than an agent taking a destructive action. But escalation transfers the conversation once the agent cannot proceed; it does not pause the agent for approval before it acts, and nothing gates ticket creation, a payment link or a booking on a person's sign off. Guardrails are named and SOP compliance is tracked, but neither is described as a mechanism. Grounding does real work: an agent stated to answer solely from provided information is constrained at the output layer.

What is documented is substantial and it is the right set for the use case. Live agents pick up complex cases through assignment modes with full cross-channel context, so escalation is a designed path rather than a fallback. Compliance officers review and modify conversation flows, which is oversight by a second role over what the agent is permitted to say, and that matters more than usual here because these are debt collection conversations where what is said is regulated. Agents are constrained to negotiate within defined parameters, payment-recovery campaigns run under a required compliance profile, and changes to an agent go to a draft before they are published. What is absent is a gate in the execution path. Nothing documented pauses a conversation for approval before an agent commits to a repayment plan, sends a payment link or writes to a system. Constraining what an agent may offer and reviewing flows beforehand are both design-time controls. So this is oversight before and around, not during, which is Partial. A documentation section on approvals or human-in-the-loop configuration would settle it.

The grade rests on a published trust center rather than a claim. Attestations are deep: ISO 27001 and ISO 27701, GDPR, CCPA, CPRA, HIPAA, FIPS 140-2, NIST SP 800-53 and CyberGRX, with independent penetration testing by a CREST certified third party and audits by BDO India. ISO 27701 is the privacy information management extension, so privacy governance is certified separately rather than asserted, and FIPS 140-2 with NIST SP 800-53 is United States federal cryptographic and control baseline territory, a strong signal for a company centered on Indian enterprise. The control surface is thinner than the certifications: single sign-on, role based access and retention configuration are not documented.

Both halves first party. The documentation states SOC 2 Type II covering security, availability and confidentiality, ISO 27001, and HIPAA compliance with a BAA available, with a Trust Center carrying certifications, reports, security practices and penetration test summaries. Single Sign-On and role-based access control through Permissions are documented, and only Admins and Owners can create API keys. The control that matters most for a bank running collections conversations is private model hosting: with Moveo's own models hosted privately and private or on-premises deployment available, conversation data need not pass to a third-party model provider. Compliance officers can also review and adjust conversation flows. Sovereign delivery properties count on deployment and are not counted twice here.

Agent Analytics, integrated with an Insights Agent, provides custom KPI dashboards, sentiment measured from the start to the end of a conversation, and tracking of agents' compliance with standard operating procedures, and the Analytics, Live Traffic, Archives and MyChats sections confirm the surfaces. Conversation level visibility is strong, and SOP compliance tracking is a substantive quality signal. It holds at Partial because it observes conversations, not agent execution: no trace shows which integrations an agent called in what order, with what arguments, or why it chose one path. When an agent creates a ticket, checks inventory and issues a payment link, the analytics show the outcome and the customer's sentiment, not the sequence.

A per-turn record of the agent's own decisions. Logs replay production conversations in the administrator view, showing for each turn the predicted intents with their confidence scores and the conversation context, including variables and user information, with an AI Agent log view for debugging individual agents and filters by channel, coverage, containment, tags and ratings. The Moveo MCP server can read session transcripts, and the API reference lists session and log-content endpoints. Analytics dashboards and a live-chat agent desk with a thread view and customer information panel sit around it. Whether webhook or tool calls appear in the production log is not stated.

Multi region data residency at the Enterprise tier is reported only in third party pricing material, described there as critical for finance, healthcare and government and as a negotiated enterprise term rather than a documented product feature. Haptik never states it on its own pages, so it supports Partial and no more. The circumstantial case is strong: as a Reliance Jio subsidiary serving Indian banks, insurers and telecoms under RBI and IRDAI rules, where local data residency is frequently required, Haptik would struggle to sell to most of its named customers without Indian data centers. But circumstance is not evidence, and the trust center is where a first party statement would sit.

Private and on-premises deployment inside the customer's own infrastructure is documented, alongside EU data residency and three named hosting regions, Europe, the United States and Brazil, and the vendor's proprietary models are hosted privately, which is what makes on-premises meaningful here. That last point deserves emphasis. An on-premises platform that still calls out to a third-party model API lets the sensitive content, the conversation, leave the perimeter even when the application does not. Moveo can run its own models, so a private deployment can be self-contained. For a bank negotiating repayment terms with a customer, that is the difference between an on-premises checkbox and an on-premises fact. EU residency plus European engineering is the second half of the story and is the reason the named customers, Allianz, Alpha Bank, National Bank of Greece, are reachable at all. No detail is documented on what an on-premises installation requires. A vendor selling to banks would normally publish that, and its absence is worth noting.

Haptik ships three named prebuilt agents: an AI Support Agent resolving queries with multi turn context, an AI Sales Agent giving catalog aware recommendations, and an AI Booking Agent for reservations and scheduling. Those are ready made agents a customer adopts and configures, not components to assemble, and Agent Assist adds prebuilt integrations into Zendesk, Freshdesk and Salesforce agent desktops. Three agent types read partly as a product taxonomy rather than a library, and no browsable catalog exists. The product is sold sales led, so how a customer adopts these agents without a full build is not documented.

The bar asks whether the customer receives packaged assets ready to adopt, and a set of prebuilt agents qualifies; a browsable marketplace is evidence of the bar, not its definition. The vendor's documentation states "AI agents come in different types, each trained and optimized for a specific use case", and that some carry predefined guidelines a customer completes to get the best out of the agent. That is a named set of ready-made agents arriving pre-configured for a purpose. The campaigns documentation also ships two campaign types, Outreach and Payment recovery. The use cases these are built around are specific to the product's market, onboarding, collections, card activation, churn prevention and support, and the Auto Builder converts a plain-text description into a ready conversational flow for anything unlisted. The agent-type list itself was not read, so how many exist and how much they arrive pre-configured with rests on the documentation's description.

Haptik states that teams choose between top models including GPT, Llama and Claude based on their needs for performance, security and scale, and invites them to experiment across those models to find the best fit: the customer chooses the model, which is what this axis measures. Haptik names security as a selection criterion alongside performance and scale. For an Indian enterprise weighing which provider may process customer conversations under RBI or IRDAI rules, which model runs where is a compliance decision, and letting a bank choose Llama over a US hosted API solves a real problem. Third party pricing material indicates premium model access is charged as pass through, so choice has a direct cost.

Customer choice across several makers. Customers connect large language models from external providers, OpenAI, Azure OpenAI, Anthropic, Gemini and DeepSeek, by adding a provider under Deploy, Developer Tools, LLMs with their own API key and choosing the model from a dropdown. Connected models are listed with provider, model, creator and last use, and the model strategy guide configures which models power an AI Agent on each channel. Moveo's own proprietary models, fine-tuned for business conversations and hosted privately, remain one option among the connected providers, and they are what make a self-contained on-premises deployment possible when the conversation should not leave the perimeter.

Haptik's documentation publishes three SDKs, a Web SDK, an iOS SDK and an Android SDK, alongside a channel documented as Bot as an API, so the platform is callable from outside through documented interfaces. The SDK set says something about the product: customers embed Haptik agents inside their own applications, a more demanding form of extensibility than an admin calling a REST endpoint, and Bot as an API makes an agent addressable as a service. Code Step Integration and Static Step Integration let custom logic be inserted into an agent's flow, and the integration catalog is documented page by page rather than asserted as a number. No MCP server is documented.

Moveo publishes its own MCP server with regional endpoints for Europe, the United States and Brazil, signed in through the browser or with a Manage-type API key sent as documented headers, and about 100 tools covering the platform's own objects: creating, cloning, training, publishing, comparing and rolling back AI Agents, changing guidelines and workflows, managing knowledge bases, environments and routing rules, and running simulations. The documentation site also carries a REST and analytics API reference, with endpoints for brains, dialogs, collections, campaigns and API keys. Both let an outside caller drive the platform. Webhooks from Dialogs run the other way, Moveo calling other systems, and count on integrations.

Haptik documents a Bot QA Tool for regression testing: a Generate API turns mock conversations covering the test scenarios into a test case CSV based on step traversal, and a Run API executes those cases and returns a result CSV with success or failure for every case, run each time a bot goes live. A functional testing guide covers test case preparation across flows, APIs, error handling and multilingual behavior, and the Test bot shows per message debug logs. That is a harness that scores the bot's behavior against recorded cases, documented for step based bots through staging API endpoints.

Simulations are regression tests for an AI Agent. The customer writes a scenario and success criteria, a simulated customer plays it through the agent's real production pipeline over chat, email or voice, and an evaluator grades the finished conversation pass or fail with written reasoning. Criteria can assert on tool calls, webhooks fired, handover, fallback responses and final context variables. Runs can be launched singly, in bulk or on a weekday, weekly or monthly schedule, and the Runs tab reports success rate, failures and duration per run. For a product negotiating repayment terms under financial-services rules, testing behavior against scenarios before deployment is exactly what a buyer needs, and it is documented. Simulations require a paid plan.

Everything a Haptik agent does runs through a documented programmatic interface. Channels are messaging and voice surfaces the agent converses on: WhatsApp, Instagram, LINE, Telegram, Microsoft Teams, web and mobile SDKs. Actions run through named integrations: creating a Zendesk ticket, checking inventory, issuing a Razorpay payment link, scheduling through a calendar. Code Step Integration executes custom logic. Real time web data integration, mentioned in third party descriptions of the platform, is retrieval over fetched content, not computer use. The absence is structural: agents live inside a customer's messaging channels, with no surface on which browser control would operate.

Every documented action path is programmatic. Agents answer from a knowledge base using retrieval, and where deterministic behavior is needed Dialogs connect to external services via webhooks and execute task automation such as sending emails and updating databases. Webhooks and database writes are programmatic interfaces by definition. Channels are conversational surfaces the agent talks on, not software it operates. There is no near miss on this record: Moveo never claims interface-level operation anywhere. The absence is structural. A conversational platform whose agents live inside WhatsApp, Viber and telephony has no surface on which browser control would operate, and the integration model, webhooks into systems of record, is precisely the alternative to driving an interface.