FLOWX.AI vs Obin AI (2026)

Stands at F, and it is the least contestable cell on this record because the integration targets are named individually rather than counted. WHAT MAKES THIS STRONG IS THE CLASS OF SYSTEM, NOT THE NUMBER. Jack Henry, FIS, Finastra and Temenos are the four core banking platforms, and COBOL mainframes and transportation management systems sit alongside them. A vendor claiming a thousand connectors is claiming breadth; a vendor naming the specific core banking systems it reaches is claiming something a prospective bank can immediately verify against its own stack, and getting it wrong would be commercially fatal. THE ARCHITECTURAL POINT IS THAT NOTHING IS REPLACED. The vendor's stated position is that agents plug into systems the customer already runs, core banking, mainframes and document stores, through GOVERNED CONNECTORS, with the legacy estate untouched. For this buyer the alternative to integration is a multi-year core replacement, so integration depth is the entire product rather than a feature of it. Breadth across classes is met: core banking platforms, mainframes, document stores, databases, APIs and transport systems span genuinely different integration classes rather than a deep catalogue inside one ecosystem, which is the distinction that holds a vendor at Partial. WHAT I DID NOT REACH is the connector documentation itself, so the mechanism by which a connector is built or configured, and whether tool calling is exposed to agents as a first-class construct, rest on product-page description. Confidence stays high because the named systems are specific and checkable, but the integration docs are the natural place to verify at lane close.

Stands at F, re-based onto the architecture documentation, and with one claim in the July basis handled more carefully than it was. THE PLATFORM LAYER IS DOCUMENTED. FlowX is a process and workflow engine before it is an agent platform: the documentation's first two entry paths are creating a process with a UI in the visual designer and building a workflow, and the architecture is a layered microservices platform with process execution as its own named concern, communicating through Apache Kafka. Multi-step orchestration is the platform's original function rather than a capability added for agents. THE AGENT LAYER SITS ON TOP AND IS MULTI-AGENT BY CONSTRUCTION. The architecture names BUSINESS AGENTS THAT POWER END-USER EXPERIENCES AT RUNTIME, and agent stacks combine numbers of cooperating agents against a single process, nine for mortgage underwriting being the vendor's own example. THE EVOLVING ARCHITECTURE CLAIM IS REAL BUT CARRIED CAREFULLY. Agents proposing system changes that builder agents implement after approval is a genuine orchestration pattern and an unusual one, agents modifying the workflow rather than only executing it. It comes from the FlowX.AI 5 launch announcement, which is a first-party vendor announcement and clears the floor, but I did not reach documentation of the mechanism, so it corroborates rather than carries the grade. The grade rests on the process engine and the multi-agent stacks. A NOTE FOR THE READER OF THE ARCHITECTURE PAGE: Kafka is inter-service communication within the platform, not a customer-facing event surface. It is recorded here as execution architecture and is deliberately not counted on triggers, where it would be the wrong fact. Control-flow vocabulary was not reached; the visual process designer implies it but no branching or looping construct was named on any page retrieved.

Stands at F, resolved from unresolved, and now the best-evidenced Trig cell in the lane. The 31 August pass failed three times on term collision; the route that worked was the /llms.txt index the architecture page advertises, followed to the 217-page v5.9 documentation index. EVERY CLASS IS A NAMED DOCUMENTATION PAGE, not an inference: SCHEDULE. Scheduled processes automate process instance creation on specific dates, times or intervals using START TIMER EVENT NODES and predefined schedules, and timer expressions are configured with CRON SYNTAX for scheduling or ISO 8601 for durations and timeouts. Timer start, intermediate and boundary events are separately documented. INBOUND WEBHOOK. Incoming Webhooks start a process from an external HTTP POST or resume a running instance and inject the payload, SECURED WITH API KEYS, EVENT-DRIVEN, NO POLLING. EMAIL. Email Trigger connects to IMAP mail servers and monitors specific mailboxes to initiate workflows, and Message Start Event starts an instance on a message, INCOMING EMAIL, OR INCOMING WEBHOOK. INBOUND API. Four documented REST endpoints start processes, including by workspace and process name, described as the simplest public entry point for triggering from a backend or scheduled job. EVENT. Manage Triggers is a centralised interface to view, activate and deactivate EVENT-BASED TRIGGERS that automatically start process instances, and message catch boundary events interrupt or run alongside a user task on an incoming message. CHANNEL. A Chat UI component enables interactive AI agent conversations with end users, and client SDKs cover web, iOS and Android. THE MANAGE TRIGGERS PAGE IS THE DETAIL THAT CONFIRMS OPERATION RATHER THAN CAPABILITY: a vendor builds an interface for activating and deactivating triggers across builds only when customers are running enough of them to need governing. Recorded for method: Kafka remains inter-service communication and is still deliberately not counted here.

Stands at F, with the marketing term that carried the July basis removed. ZERO HALLUCINATION TECHNOLOGY IS NOT GRADED AND SHOULD NOT HAVE BEEN. It is a proprietary marketing name for a claim no system can make absolutely, and grading grounding from it credits a slogan. The vendor's current and more careful phrasing is HALLUCINATION CONTROLS MEASURED IN PRODUCTION, which is a different and more defensible statement; it is recorded but it belongs to governance rather than to grounding. WHAT CARRIES THE GRADE IS THE ONTOLOGY LAYER, and it is the right kind of evidence for the persistence line this axis turns on. The vendor names it as one of three platform components on which every catalogue agent was composed, alongside Agent Builder and Observatory. An ontology is a maintained structure describing entities and their relationships, which is the strongest form of maintained knowledge structure and the opposite of per-request context assembly. THE SECOND HALF IS THE UNIFIED SOURCE OF TRUTH. The platform's stated function is integrating thousands of disjointed legacy systems into one, and agents are trained on the institution's own data under its own rules and compliance. Grounding on operational systems of record rather than on an uploaded document corpus is the shape that matters for the work sold here: an underwriting agent needs the current state of the core banking system, not a snapshot. Unified case files aggregating and correlating evidence across source systems are real grounding output, though the case file as an artefact is the AML product's deliverable and is deliberately not also counted on observability, where the July record used it twice. CONFIDENCE IS MEDIUM AND THE GAP IS NAMED: no retrieval, indexing, embedding or ontology configuration documentation was reached, so the mechanism rests on product-page description of a named component.

P>F, resolved from unresolved. The July basis said no explicit per-agent memory construct is documented, which was true of the marketing pages and is not true of the documentation. THE DECISIVE PAGE IS TITLED FOR THIS QUESTION: PASSING DATA BETWEEN NODES, described as how outputs from subprocesses, workflows, AI AGENTS, AND BUSINESS RULES LAND IN PROCESS VARIABLES, AND HOW DOWNSTREAM NODES READ THEM. Agent output written to durable process variables and read by later steps is agent-addressable state, named for agents specifically rather than inferred from a general data model. THE STATE OUTLIVES THE RUN IN TWO WAYS. Within a process, variables persist for the life of the instance, and BPMN instances in this domain run for days or weeks: the platform documents updating process variables on ACTIVE instances to resolve production issues without restarts, and migrating running instances between builds. Across processes, FlowX Database stores STRUCTURED DATA ACROSS PROCESSES AND APPS, which is durable shared state readable by any later agent. THAT IS THE SAME ARCHITECTURE CREDITED AT FULL ON JOGET AND OUTSYSTEMS this lane, and grading it differently would be inconsistent. joget reached Full on store-to-form and store-to-process-variable enhancers writing agent output into the application data layer; outsystems on StoreMemory plus process variables. FlowX has both halves: process variables carrying agent output, and a platform database spanning processes. A THIRD SURFACE SUPPORTS IT. Knowledge Bases hold static documents and DYNAMIC DATA FEEDS with documented create, append, replace, delete and edit-metadata operations, so an agent's accumulated context is a maintained object rather than a transient one. CONFIDENCE IS MEDIUM AND THE LIMIT IS PRECISE: this is process state that agents read and write, not a named per-agent memory primitive with retention or per-user scoping. Nothing documents an agent recalling a prior conversation with a particular person. For a BPMN platform that is the coherent shape, and it clears the ruled state limb.

Stands at F on a documented mechanism, with its scope named because the July basis overstated it as oversight in general. THE MECHANISM IS SPECIFIC AND IT IS THE VENDOR'S OWN. In the FlowX.AI 5 evolving architecture, agents PROPOSE system and workflow changes for human review, and only after designated operators approve do builder agents implement them. The worked example is an Auditor Agent that interprets a regulatory update, proposes the corresponding workflow change, and waits. That is propose, approve, implement, with a person holding the gate, and it is FlowX's own surface rather than one the customer already owned. IT IS AN UNUSUAL AND STRONG SHAPE FOR THIS AXIS. Most approval mechanisms in this lane gate an agent's business action. This one gates the agent's ability to change the system itself, which is the higher-consequence class: a wrongly executed transaction is one bad outcome, a wrongly altered underwriting workflow is every subsequent outcome. For a regulated buyer that is the right place to put the gate. THE SCOPE LIMIT, WHICH THE JULY BASIS DID NOT STATE. This gates self-modification. Nothing reached documents a runtime approval step before an agent takes a business action, no checkpoint before a case is decided, a payment released or a submission filed. Given the vendor sells AML investigation and lending underwriting, whether a human approves the agent's substantive output is a separate and unanswered question. POLICY-GATED OUTPUTS, stated on the product page alongside audit trails, is the guardrail half and is a real constraint rather than an essay. CONFIDENCE IS MEDIUM because the mechanism rests on the FlowX.AI 5 launch announcement, which is a first-party vendor announcement and clears the evidence floor, rather than on product documentation. Three passes failed to reach the AI Platform documentation where it would be described.

F>P, resolved from unresolved. The July basis rested on BANKING GRADE SECURITY and BUILT FOR REGULATORS, which are adjectives, plus perimeter and residency facts that belong on Dep. Stripped of those, no attestation remained, and this pass looked properly for one. NO CERTIFICATION IS CLAIMED ANYWHERE. The 217-page v5.9 documentation index contains no security, compliance, trust or certification page. The marketing site carries no trust centre. No SOC 2, ISO 27001, ISO 27017, ISO 27018, DORA or PCI DSS claim appears on any first-party surface reached. THE VENDOR'S OWN BUYER-GUIDANCE CONTENT IS THE SHARPEST EVIDENCE, and it cuts against them. FlowX publishes a hub article on evaluating banking automation vendors which tells banks the baseline attestation set is SOC 2 TYPE II, ISO/IEC 27001, ISO/IEC 27017 AND 27018, and instructs them to REQUEST THE SOC 2 TYPE II REPORT AND ISO 27001 STATEMENT OF APPLICABILITY UNDER NDA, NOT A SUMMARY DECK. Where the same article mentions FlowX, it claims only banking-grade safety with zero hallucinations and deterministic outputs, adding that this is A CLAIM BANKS SHOULD VALIDATE UNDER THEIR OWN MODEL-RISK GOVERNANCE. A vendor writing the procurement checklist for its own category, and not placing itself on it, is conspicuous. This is the same shape refused on EverWorker in this lane, where blog posts advising buyers to require SOC 2 were mistakable for the vendor's own posture. Here it is worth grading because of what is absent from it. THE CONTROL HALF IS GENUINELY MET, which is why this is Partial rather than lower. The documentation names Keycloak-based authorisation and access roles, swimlanes grouping nodes by participant, BUSINESS FILTERS restricting access to process instances by a business value such as branch, permission-based expressions controlling UI visibility by role and process data, an anonymous runtime role for unauthenticated flows, and a dedicated Audit service providing a centralised location for all audit events. Under the conjunction bar, controls documented and attestation absent is Partial. A published certification would move it immediately.

Stands at F, but one fact has been removed from the basis because it was doing work on two axes, and that removal matters more than the grade. WHAT CAME OUT: human-readable event timelines and unified case files. Those are the AML and fraud product's DELIVERABLE, the artefact an investigator receives, built by aggregating evidence across source systems. They evidence grounding, which is where they now sit, and crediting them here counted the product's output as though it were visibility into the agent that produced it. That is the same error as counting SnapLogic's trillions of records as observability. WHAT CARRIES THE GRADE IS OBSERVATORY, a named platform governance component that the vendor states every catalogue agent was composed on, alongside Agent Builder and the ontology layer. A named component whose stated purpose is governance is a different class of evidence from an adjective. AROUND IT SIT THREE STATED PROPERTIES: audit trails, POLICY-GATED OUTPUTS, and HALLUCINATION CONTROLS MEASURED IN PRODUCTION. The third is the one worth carrying. Measured in production means the vendor claims a running quality signal rather than a pre-release check, and for a regulated buyer a documented measurement of model reliability in live operation is closer to what they actually need than a trace viewer. CONFIDENCE IS MEDIUM AND THE GAP IS SPECIFIC. I did not reach the Observatory documentation, so what it exposes, whether a customer sees per-agent run traces or only aggregate governance reporting, and what retention or export exists, are all unestablished. Three passes failed to reach the AI Platform documentation section. Given this vendor sells to regulators as its primary pitch, the capability is more likely real than not, but the grade currently rests on a component name plus three product-page properties. This is the first cell to verify at lane close, together with Sec.

Stands at F, re-based off marketing prose and onto the installation documentation, which is what this axis should rest on. THE JULY BASIS WAS THE RIGHT ANSWER FROM THE WRONG SOURCE. It cited data staying inside the customer's perimeter, agents running next to their systems under their policies, and no vendor lock-in. Those are homepage sentences. Under the outsystems Dep correction earlier in this batch, a deployment claim on a sales page is not a documented deployment option, and I would have had to drop this cell if nothing else existed. WHAT REPLACES IT IS DOCUMENTED AND SPECIFIC. The documentation landing page carries a named path, I NEED TO DEPLOY FLOWX, INSTALL AND CONFIGURE FLOWX IN YOUR INFRASTRUCTURE. That is an installation route offered to customers rather than a bespoke deployment the vendor performs, which is the distinction that decided ai-library's Dep cell today. THE ARCHITECTURE MAKES IT CREDIBLE RATHER THAN ASSERTED. All backend services are containerised and deployed on Kubernetes, with more than twenty-three services, Kafka for inter-service communication, and a published data architecture page covering database-to-service mapping, storage types and sizing guidance. Sizing guidance in particular is something a vendor writes only for customers who run the software themselves. FOR THE BUYER THIS SEGMENT SERVES, this is the load-bearing property. National Bank of Canada, OTP and Banca Transilvania are not going to run mission-critical lending on a multi-tenant SaaS, and the whole premise of leaving legacy core banking untouched requires the platform to sit beside it. One limit recorded: no region selection or managed-cloud residency options were documented on the pages reached, so the grade rests on customer-infrastructure installation rather than on a residency menu.

Stands at F and clears the 31 August Pack bar in its strongest form, because the vendor uses the word this axis looks for: EVERY AGENT IN THE CATALOG. A catalogue is the evidence the bar names, and this is one of the few records in the lane where the vendor calls it that itself. THE COUNT IS SPECIFIC AND STRUCTURED. More than 220 production-ready, tested agents across 20 categories, bundled into banking agent stacks with named compositions: nine agents for mortgage underwriting, seven for commercial onboarding, four for invoicing. Naming how many agents constitute a stack is a level of specificity that marketing prose does not usually reach, and it describes a finished assembly for a business process rather than a template someone starts from. THE STACK CONCEPT IS THE PART WORTH CARRYING TO COMPARISON PAGES. Most vendors at Full on this axis ship components a customer assembles. A mortgage underwriting stack of nine cooperating agents is the finished job, which is why the vendor can talk about going live in weeks rather than quarters. ONE THING THE CATALOGUE CLAIM ALSO TELLS US, and it is why this cell is credible rather than promotional: the vendor states every catalogue agent was composed on its own platform using Agent Builder, Observatory governance and the ontology layer. The packs are built with the same tooling sold to customers rather than hand-crafted separately, which is the internally consistent version of this claim. What I did not reach is the catalogue itself as a browsable surface, so the count and the category structure rest on the vendor's own summary rather than on an enumerated list. That is the check for lane close.

Stands at P, resolved from unresolved, and the basis is replaced because the July version reasoned from marketing rather than from evidence. WHAT CAME OUT. The July basis read NO VENDOR LOCK IN and WORKING WITH THE CUSTOMER'S EXISTING STACK as SUGGESTING FLEXIBILITY. That is inference from positioning, the same error corrected on mobagel, and it cannot support a grade in either direction. WHAT THE DOCUMENTATION SHOWS. The 217-page v5.9 index contains no model configuration, provider selection or routing page. Agent-related pages cover the Custom Agent node creating agents with MCP tools, MCP integration and data sources, Knowledge Base integration, Context Retrieval nodes, document classifiers and a Chat component. Model choice is not among them. THE ONE FIRST-PARTY STATEMENT ABOUT MODELS is in the 5.1 documentation, where AI Core is described as A SERIES OF AI AGENTS CONNECTED TO A CENTRALIZED REPOSITORY OF LLMS. A centralised repository is a vendor-managed pool: the platform holds the models and agents draw from them. That is vendor-side provisioning without a documented customer selection surface, which is the Partial position under the standing convention, and it is a founded Partial rather than the inferred one it replaces. ONE ADJACENT FACT RECORDED AND NOT CREDITED, because it belongs to a different axis. The vendor's own banking hub article names LLM-AGNOSTIC DESIGN, NO LOCK-IN TO A SINGLE MODEL PROVIDER, WITH LLM ISOLATION SO THERE IS NO THIRD-PARTY SAAS DATA PATH FOR SENSITIVE PROMPTS as something banks should look for. It is written as buyer guidance rather than as a FlowX claim, the same construction that left the Sec attestation absent, and I am not reading it as a self-description. WHAT WOULD MOVE THIS: a model configuration page, or documentation of the AI Core repository showing which models a customer may select per agent. The 5.9 AI section is thinner in the index than 5.1's AI Core section was, which may mean the pages moved rather than the capability.

Stands at F, and the basis is replaced entirely because the July version graded the wrong things. It cited the Agent Builder Platform, low-code and natural-language authoring, and INTEGRATES VIA APIS WITH SYSTEM INTEGRATOR PARTNERS. The builder is the product itself, not an extensibility surface, and a system integrator partnership is a go-to-market arrangement that belongs nowhere in the grid. WHAT ACTUALLY CLEARS MIKE'S 30 AUGUST BAR, and it clears it comfortably. The documentation landing page offers a named path, I WANT TO INTEGRATE APIS, USE THE SDKS (ANGULAR, REACT, IOS, ANDROID) OR REST APIS. Four named client SDKs across web and both mobile platforms, plus REST, is a developer surface documented as a first-class route rather than inferred. THE ARCHITECTURE PAGE CORROBORATES IT STRUCTURALLY. The platform is 23-plus backend microservices with client applications connecting through an API GATEWAY THAT HANDLES OAUTH2 TOKEN VALIDATION AND REQUEST ROUTING. A published gateway with a named auth standard is what makes the API usable by something other than the vendor's own front end. ONE DETAIL WORTH CARRYING for a platform selling into banks: the documentation publishes an /llms.txt index inviting retrieval of the complete documentation index. A vendor that machine-publishes its own docs index is unusually open for this segment, and it is also the artefact that would have saved the July pass. No MCP server was found on any page reached, which under the ruling does not withhold the grade. WHAT THE JULY BASIS GOT RIGHT BY ACCIDENT is that the grade was correct. It is now correct for reasons a reader can check.

Stands at P, resolved from unresolved, and the shape is the familiar one: a strong testing culture aimed at everything except the agent. WHAT IS DOCUMENTED, AND IT IS MORE THAN THE JULY BASIS FOUND. Testing a knowledge base runs create, append, replace, delete, edit-metadata and query operations against a deployed build BEFORE RUNNING THEM IN PRODUCTION. Testing data sources from Runtime runs database operations and SQL queries against a deployed build WITHOUT TRIGGERING A WORKFLOW. Mock integrations let a process be exercised before integrations are complete. Failed process start exceptions are viewable with detailed error information and stack traces. Observability with OpenTelemetry provides distributed tracing, metrics and logging. THAT IS A GENUINE PRE-PRODUCTION VALIDATION LAYER and it is better than most Partials in this lane. Being able to test a knowledge base and a data source against a deployed build, independently of the workflow that uses them, is careful engineering aimed at the failure modes that actually break regulated deployments. WHY IT STILL DOES NOT CLEAR THE BAR. Every one of those tests exercises a COMPONENT: a knowledge base, a data source, an integration. None scores the agent. The ruled bar asks for a readable comparable result about the agent's behaviour on the customer's own work, and nothing in a 217-page index provides a test set, expected outputs, a judge, a scored run or a comparison between agent versions. THE CLAIM THAT LOOKS LIKE EVALUATION AND IS NOT: agents are described as producing deterministic, zero-hallucination, fully auditable outputs, and customers measure impact before scaling. Determinism is an architectural assertion, and business impact is a commercial outcome. Neither is a customer-runnable evaluation, and the same reasoning held ai-library and integrail at Partial and None respectively this lane. THE ABSENCE IS SHARPEST FOR THIS BUYER. The vendor's own banking article tells banks to confirm that each new agent does not trigger a fresh model-risk review by VALIDATING THE UNDERLYING DETERMINISM GUARANTEES ONCE, THEN REUSING THEM. That argument only holds if determinism can be demonstrated, and no surface for demonstrating it is documented.

Stands at N, and confidence rises to high because the architecture settles it rather than a search failing to find a claim. Everything the platform does runs through documented programmatic paths: governed connectors into core banking systems, mainframes and document stores, REST APIs, client SDKs, and Kafka-mediated services. Comp is non-zero only where an agent operates software through its interface BECAUSE no programmatic interface exists, and FlowX's entire commercial premise is the opposite, that it can reach systems programmatically which others cannot. THE NEAR-MISS IS THE COBOL MAINFRAME CLAIM AND IT NEEDS REFUSING EXPLICITLY, because it is the most plausible false positive I have seen on this axis. A vendor that reaches forty-year-old mainframe systems sounds exactly like a computer-use vendor: those systems predate modern APIs, and green-screen terminal emulation is a real and common integration technique that WOULD count here, since a 3270 screen is an interface with no programmatic alternative. BUT THE VENDOR DESCRIBES CONNECTORS, NOT EMULATION. The stated mechanism is governed connectors integrating systems into a single source of truth while leaving them untouched, and nothing on any page reached mentions screen scraping, terminal emulation, robotic process automation or interface driving. If a later pass finds green-screen automation documented, this cell should move, and that is a real possibility worth recording rather than a hypothetical. That distinction, integration versus emulation on legacy systems, is worth carrying to the Enterprise operations lane, where mainframe modernisation vendors will recur and the same near-miss will present each time.