Spoki vs Wassist (2026)

Stands at F, re-based, with two facts removed that belonged to other axes and one that ran the wrong direction. WHAT CAME OUT. Website and document import is grounding, credited on Know. The TypeScript SDK and REST API are the outbound developer surface, credited on Ext. WhatsMCP was cited on both this cell and Ext, and under the standing one-line test it belongs here alone: Wassist is the MCP CLIENT calling remote servers, so the vendor's product is consuming third-party tools. That is tool calling, not extensibility. WHAT CARRIES THE GRADE IS TWO THINGS OF DIFFERENT KINDS. The documentation states plainly, CONNECT YOUR AGENT TO ANY EXTERNAL API. BOOK APPOINTMENTS, CHECK INVENTORY, PROCESS PAYMENTS, so tool calling is unbounded rather than catalogue-limited. And WHATSMCP connects any remote MCP server, which turns an open standard into an open tool catalogue without Wassist building a connector for anything. THE NAMED INTEGRATIONS ARE THE HONEST LIMIT AND WORTH STATING. Shopify, Klaviyo, Yotpo and Recharge are all the Shopify commerce stack, which is one ecosystem class, and on its own that would sit at Partial under the convention that a deep catalogue inside a single class is not breadth. Breadth here comes from the arbitrary API and MCP routes rather than from the connector list. WHAT MAKES THE COMMERCE INTEGRATION SUBSTANTIVE rather than a logo is that it reaches transaction: agents surface product cards, carts and payments natively inside the thread, and orders and shipping updates return to the same conversation. Reading a catalogue is easy; completing a purchase through it is the harder half. One open question recorded: whether WhatsMCP is a component of the agent product or a separate consumer offering was not established on the pages reached.

Stands at P. The July basis was accurate and the finding this pass is that the shape is deliberate rather than unfinished. WHAT IS DOCUMENTED IS CONVERSATIONAL TOOL USE, WELL DONE. Agents hold multi-turn conversations, call external APIs to book appointments, check inventory or take payment, and escalate to the right human team. Sequencing a purchase inside one thread, browse to cart to payment to receipt to shipping update, is genuine multi-step work with external side effects at several points. WHAT IS ABSENT IS ANY ORCHESTRATION LAYER ABOVE THE CONVERSATION. No sub-agents, no agent-to-agent handoff, no branching, looping or conditional construct, no visual flow surface, and no documented process object that outlives a single thread. The unit of execution is the conversation, and when it ends the work ends. THAT IS COHERENT FOR THE PRODUCT and worth stating so the grade is not read as immaturity. A front-of-house commerce agent answering a shopper does not need a directed graph; it needs to answer well and call the right API. The vendor is explicit that it targets simplicity, a working agent in under a minute with no coding and no API keys, and orchestration machinery would work against that. WHERE IT COSTS THE BUYER is anything spanning conversations: a returns process that waits three days for a parcel, or a campaign whose follow-up depends on an earlier reply. Cart recovery and order updates are the vendor's answer to that and they are event-triggered rather than orchestrated, which is a narrower mechanism. Confidence is medium; the documentation's concepts page describing how agents, tools and conversations work together was not reached and would settle whether any durable process object exists.

Stands at P, re-based with richer evidence, and the single-channel limit is what holds it below Full. THE EVENT CLASSES ARE GENUINELY PLURAL AND COMMERCIALLY SPECIFIC. Inbound customer messages; abandoned cart recovery, which fires on a state change in the store rather than on anything the customer does; order and delivery updates, which fire on fulfilment events; and event subscriptions through the API for anything else. Cart abandonment in particular is a real store-side trigger and is the product's headline use case, with the vendor claiming it four times more effective than email. ENTRY POINTS ARE DOCUMENTED AS A DISTINCT SURFACE: footer links, buttons and QR codes that open a WhatsApp thread. That is the acquisition path for a conversational agent and most vendors in this lane do not document it at all. OUTBOUND IS ALSO COVERED, through WhatsApp's native broadcast for campaigns, which became commercially viable when Meta removed outbound messaging costs in November 2024. WHAT HOLDS IT AT PARTIAL IS THAT ALL OF IT ARRIVES ON ONE CHANNEL. Every route terminates in a WhatsApp thread. There is no email, web chat, Slack, voice or embedded surface, and no scheduled or cron capability is documented. Against kalcend earlier today, five simultaneous channels plus API, this is one channel with several event types, which is the Partial shape. THE STRATEGIC POINT WORTH RECORDING rather than treating as a deficiency: single-channel is the product thesis, not an omission. Wassist is a bet that WhatsApp is the channel, citing 98 percent open rates against 20 percent for email. A buyer choosing it is choosing that bet, and the grade should read as narrowness rather than immaturity.

Stands at F on the persistence line, with the limit named more precisely than the July basis managed. THE GROUNDING STRUCTURE IS BUILT ONCE AND PERSISTS, which is what separates Full from per-request assembly. An agent is created from a website import, a connected Shopify store, or uploaded documents, and the platform generates a first version grounded in that specific corpus in the brand's voice. The Shopify connection is the strongest form, because a live store connection tracks the catalogue rather than freezing a snapshot: prices, stock and product copy stay current without anyone re-uploading. WHAT THE GROUNDING IS FOR IS UNUSUALLY CONCRETE. Agents answer on returns, order status, sizing and product questions and make recommendations from the catalogue. Those are answerable only from live commercial data, so grounding here is transactional rather than documentary, and getting it wrong is immediately visible to a paying customer. ONE FACT MOVED OFF THIS CELL: brand voice. A configured tone is prompt configuration, not knowledge, and the July basis leaned on it as though the two were the same. It belongs in the description. CONFIDENCE IS MEDIUM AND THE GAP IS SPECIFIC. No retrieval mechanism is documented anywhere reached: no vector store, embedding, chunking, re-index cadence or citation behaviour. For the Shopify path a live connection is implied but the sync model is not stated, so whether the catalogue is queried at request time or periodically indexed is unknown. That distinction decides how stale an answer about stock can be, which for this buyer is the whole question. The documentation's concepts page covering how agents, tools and conversations work together was not reached and is the natural place to settle it.

Stands at P under the state limb of the 31 August ruling, with one fact removed that was not memory at all. WHAT CAME OUT: a persistent configuration in the brand's voice. That is the agent's definition, authored once by the operator, not state the agent writes and later reads. The July basis treated durable configuration as durable memory, which would credit every platform in the index with memory for the fact that its settings save. WHAT IS DOCUMENTED IS SESSION STATE. Agents follow context across a multi-turn conversation, which is exactly the ruled definition of Partial: state that lives only within a session, or a conversation buffer with no documented store. ONE CHANNEL PROPERTY CUTS IN THE VENDOR'S FAVOUR AND IS WORTH RECORDING. WhatsApp threads are persistent and identity-bound by construction: the same phone number returns to the same thread months later, and the history is visibly there for both parties. So a returning shopper's prior conversation exists whether or not Wassist maintains a memory layer, and an agent reading recent thread context would appear to remember. That is the platform's property rather than the vendor's, and under the standing question of whose mechanism it is, it cannot carry the grade. WHAT WOULD MOVE THIS is documentation of a memory store: retained customer attributes across conversations, purchase history the agent recalls unprompted, or preferences accumulated over time. For a commerce agent whose whole pitch is personalised recommendation and lifetime value, per-customer memory is the obvious capability, and its absence from the documentation is more likely a gap in what is written than in what exists. The concepts page on how agents, tools and conversations work together was not reached across three passes and is where this should be settled at lane close.

Stands at P. Human handoff is real and documented, but it is the weaker of the two things this axis measures. WHAT IS DOCUMENTED: agents ANSWER FAQS, HANDLE INQUIRIES, AND ESCALATE TO HUMANS WHEN NEEDED, handing off to the right team for complex cases. Escalation is a genuine mechanism and for a customer-facing commerce agent it is the one that matters most in practice, because the failure mode is a shopper stuck with a bot rather than an agent taking a catastrophic action. WHY IT DOES NOT REACH FULL, and the distinction is worth keeping sharp. Escalation is an EXIT: the agent gives up and a person takes over. An approval gate is a CHECKPOINT: the agent proposes an action and waits for a person to permit it. Nothing documented pauses an agent before it acts, and this agent does act, taking payments, applying discounts, triggering refunds or returns through connected APIs. Nothing requires a human to confirm any of those. WHAT IS ALSO ABSENT is the guardrail half: no content policy, topic boundary, spend limit, action allowlist or confidence threshold is documented on any page reached. So the cell rests on escalation alone, which is one mechanism out of the several this axis contemplates. THE RISK CONCENTRATION IS WORTH RECORDING because it is unusual in this lane. Wassist agents transact with the general public at scale on a channel with 98 percent open rates, using a brand's own voice and number. An agent that can complete purchases and issue refunds with no documented approval step, guardrail or spend limit is carrying more consequence than most Partial cells in this lane imply, and a merchant should ask about it. The escalation destination and whether a human can take over a live thread were not established on the pages reached.

Stands at P, with the testing fact removed because it was being counted on two axes. WHAT CAME OUT: the ability to test an agent before publishing. That is a build-time quality gate and it belongs on Eval, where it now sits alone. The July basis used it here and there, which is the same one-fact-two-axes pattern found on four records in this batch. WHAT REMAINS IS ANALYTICS AND CONVERSATION REVIEW. The documentation lists analytics alongside templates and business profiles as part of what the platform provides, and conversations are reviewable. For a conversational commerce agent that is a meaningful surface: an operator can read what the agent actually said to a customer, which for a text-native product carries a good deal of the reasoning. WHY IT HOLDS AT PARTIAL RATHER THAN RISING. This axis distinguishes reporting from auditing, seeing what happened from reconstructing why. Nothing documented records the agent's actions as distinct events: which external API was called, with what parameters, and what came back. That gap is sharper here than on most Partial cells because these agents take payments and call inventory and fulfilment APIs. An operator can read the conversation in which a refund was discussed and cannot see the refund call itself. No audit log, retention period, export path or per-action record appears on any page reached across three passes. CONFIDENCE IS MEDIUM. The vendor publishes analytics as a named platform feature but I reached the word rather than the surface, so what it measures, engagement volumes or agent behaviour, is not established. Given event subscriptions exist through the API, an operator could build their own action record; that is the customer's work rather than the platform's and is not credited.

P>N. The July basis correctly conceded that no on-premises or residency options are documented, then graded Partial anyway on facts that are not deployment facts. WHAT IT CITED AND WHY IT DOES NOT COUNT. Publishing agents natively to WhatsApp is a distribution channel, credited on Trig. Using the customer's own WhatsApp Business account or a Wassist-provided number is a telephony and account arrangement, not a statement about where software runs or data sits; a customer bringing their own WABA still has Wassist processing every message on Wassist's infrastructure. WHAT THE PLATFORM ACTUALLY IS, on its own documentation: a hosted service with three usage paths, a no-code dashboard, programmatic control through the REST API and TypeScript SDK, and connecting an existing agent to WhatsApp USING OUR INFRASTRUCTURE. All three are the vendor's cloud. No region selection, virtual private cloud, self-hosting or data localisation commitment appears on any page reached across three passes. UNDER SECTION 7 SINGLE-REGION SAAS WITH NO SELECTION IS NONE, and this is that. WORTH RECORDING BECAUSE IT IS STRUCTURAL RATHER THAN AN OVERSIGHT: this product is entirely dependent on Meta's WhatsApp Business Platform, so conversation data traverses Meta's infrastructure by construction regardless of what Wassist does. A residency option would be constrained by that dependency, which makes the absence coherent for the architecture rather than a gap the vendor could close cheaply. The third usage path, connecting a customer's existing agent over Wassist's infrastructure, is a programmatic route credited on Ext, and it is the closest thing here to customer control over where the agent itself runs.

Stands at F under the 31 August bar, which asks whether the customer receives packaged assets ready to adopt and treats a browsable catalogue as evidence rather than definition. TWO ROUTES QUALIFY AND THEY ARE DIFFERENT IN KIND. Templates are vendor-supplied starting points. A COMMUNITY GALLERY OF AGENTS CRAFTED BY OTHER USERS is the browsable catalogue form, and a peer-built one is a stronger signal than a vendor-curated shelf: it only exists if people are actually building things worth cloning, and it grows without the vendor's effort. THE THIRD ROUTE IS THE ONE MOST BUYERS WILL USE and it is the reason this reaches Full rather than sitting at Partial like the generation-only records reviewed earlier today. Entering a store URL produces a working agent in under a minute, pre-configured for the ecommerce job with cart recovery, order updates and product recommendations BUILT IN. Unlike a blank generator, what arrives is a specific, complete agent for a known business shape. That is a packaged asset by any reasonable reading, delivered by generation rather than by catalogue. THE CONTRAST WITH KALCEND EARLIER TODAY IS INSTRUCTIVE and worth carrying: kalcend generates from a free-text description with no catalogue and dropped to None; Wassist generates against a known vertical AND publishes a gallery of adoptable peer agents. Same technique, different outcome, because here the customer also has finished things to take. CONFIDENCE IS MEDIUM because I reached descriptions of the gallery and templates rather than the gallery itself, so how many agents it holds and whether they are genuinely one-click adoptable is the vendor's characterisation. If a lane-close pass finds the gallery thin, the pre-configured ecommerce build would still support Partial on its own.

Stands at N, and the absence is now observed on first-party pages rather than merely unfound, though two leads are recorded that could move it. WHAT IS DOCUMENTED IS THE OPPOSITE OF CHOICE, AND IT IS THE PRODUCT'S SELLING POINT. Wassist configures the AI automatically: a merchant enters a store URL and receives a working agent in under a minute with NO CODING, NO API KEYS. No model provider is named anywhere reached, no picker exists, and no bring-your-own-key or endpoint configuration is documented. Not requiring an API key is advertised as a feature, which is a clear statement that model procurement is the vendor's job. THIS IS THE WEAKER FORM OF ABSENCE and the note should say so. Per the ground rules a vendor naming a single provider is the best-evidenced absence, because the buyer at least learns what powers the product. Here nothing is named, so a merchant cannot tell which model reads their customer conversations, which is a fair question for a product handling payments and personal data. TWO LEADS RECORDED, NEITHER GRADED. First, the pricing page states IMAGE GENERATION CREDITS VARY BY MODEL, which is the only first-party acknowledgement that more than one model exists and implies selection somewhere, though for image generation rather than the conversational agent. Second, the documentation's third usage path lets a customer CONNECT YOUR EXISTING AI AGENT TO WHATSAPP USING OUR INFRASTRUCTURE, where the customer controls the model completely. I have not credited that here because on that path Wassist is the transport rather than the agent platform, and it is graded on Ext where it belongs; crediting it would let one fact do work on two axes. CONFIDENCE MEDIUM: I reached the docs landing and pricing pages but not a model or configuration reference, and the image-generation line suggests one may exist.

Stands at F, re-based, with WhatsMCP moved off this cell. THE CORRECTION THAT MATTERS: the July basis cited WhatsMCP here as extensibility. It is an MCP client, meaning Wassist calls out to remote servers, which is the opposite direction from what this axis measures. Under the standing one-line test, the customer's assistant calling the vendor is Ext and the vendor calling out is Int. WhatsMCP now sits on Int alone. THE GRADE DOES NOT DEPEND ON IT AND CLEARS MIKE'S 30 AUGUST BAR TWICE OVER. The documentation names a FULLY TYPED TYPESCRIPT SDK WITH COMPREHENSIVE REST API DOCUMENTATION, described as the same interfaces Wassist uses internally, plus event subscriptions. Fully typed matters: types are a maintenance commitment a vendor does not make for a surface it treats as an afterthought. THE THIRD USAGE PATH IS THE MOST INTERESTING AND WAS NOT CREDITED BEFORE. The documentation offers three entry points, a no-code dashboard, full programmatic control, and CONNECT YOUR EXISTING AI AGENT TO WHATSAPP USING OUR INFRASTRUCTURE, MAXIMUM FLEXIBILITY. That last one is the platform operating as pure infrastructure for somebody else's agent, which is an unusually complete form of extensibility: the customer keeps their agent, their model and their logic, and takes only the WhatsApp transport. FOR THE LANE that path is worth carrying, because it makes Wassist usable as a channel layer beneath another builder in this index rather than only as a competitor to one. What is not documented is a first-party MCP server exposing Wassist agents to outside assistants, which under the ruling does not withhold the grade.

Stands at P under the 31 August Eval bar, and it now holds the pre-publish testing fact exclusively rather than sharing it with Obs. WHAT IS DOCUMENTED IS A REAL PRE-PUBLISH LOOP, and it is structural rather than optional. The platform's stated purpose is to CREATE, TEST AND PUBLISH agents, with test in the middle, and a creator previews how an agent responds and refines it before it goes live. The vendor also guides a first agent through a tutorial in five minutes, so the create-test-refine cycle is the documented path rather than an advanced feature. FOR THIS PRODUCT THE TESTING STAKE IS HIGH and the loop is well placed. A published agent talks to real customers on the brand's own WhatsApp number in the brand's voice, so a bad answer is a public one. Requiring a preview before publication is the right gate. WHY IT DOES NOT REACH FULL. The ruled bar asks for a result the customer can read AND COMPARE about the agent's behaviour on their own work. Previewing shows what the agent says; nothing scores it, retains the scenarios as a test set, records expected answers, or lets a merchant establish whether a change to the knowledge base or prompt made behaviour better rather than merely different. Preview-and-refine is a quality gate with no readable result, which is the ruled definition of Partial. THE GAP IS SHARPENED BY THE PRODUCT'S OWN MECHANICS, and this is the part worth carrying. The agent is grounded on a live Shopify catalogue, so its inputs change continuously without anyone editing the agent. A platform whose knowledge shifts underneath it and which offers no regression check is one where quality can drift silently between publishes, and the merchant would learn about it from a customer.

Stands at N, and confidence rises to high because the architecture settles it rather than a search returning nothing. Everything an agent does runs through documented programmatic paths: WhatsApp message exchange over Meta's Business Platform, calls to any external API, native Shopify and commerce-stack integrations, and remote MCP servers through WhatsMCP. Comp is non-zero only where an agent operates software through its interface BECAUSE no programmatic interface exists, and this product is constructed entirely from interfaces that do. ONE NEAR-MISS WORTH NAMING AND REFUSING. Agents complete purchases, and the vendor states the shopper adds to cart with PAYMENT HAPPENING THROUGH THE WEBSITE CHECKOUT, ALL SHOWN IN-APP. An agent that drives a checkout sounds like the positive case for this axis, and a grader working from a summary could read it that way. It is not: the checkout is surfaced to the human shopper inside the WhatsApp thread and the shopper completes it. The agent presents a payment surface rather than operating one, and Meta's native payment message types are the mechanism. THAT DISTINCTION IS WORTH CARRYING to the ecommerce agent records in Enterprise operations and Customer support, where agent-completes-checkout claims will recur and the same question applies: does the agent drive the checkout, or hand it to the person. Website import is a one-time content ingestion for grounding, credited on Know, and is retrieval rather than navigation. No browser control, form filling, screen interaction, visual grounding or desktop automation appears on any page reached across three passes, and the vendor makes no claim in that territory.