Convey vs Spinnable (2026)

Stands at F, and the breadth argument here is different from every other record in this lane, which is why it holds. Most vendors earn this cell with a connector catalogue. Convey has no catalogue and does not need one, because a teammate reaches a system either programmatically where an interface exists or BY OPERATING THE COMPUTER DIRECTLY where one does not. That removes the ceiling entirely: the set of reachable systems is the set a human employee could use, which is strictly larger than any connector list. The legacy case is the point. Named workloads run across shipping portals, spreadsheets, ERP and advertising systems, and the vendor is explicit that it connects to LEGACY TOOLS WITHOUT MODERN APIS. Those are precisely the systems that appear on no vendor's integration page and that block most automation projects. Action is real rather than read-only: teammates process invoices, reconcile financial data, ingest orders and build reports, all of which write. One honest caveat for a comparison page: reach and reliability are not the same thing. An API integration breaks loudly when the contract changes; a taught screen procedure breaks silently when a button moves. Deterministic replay makes that trade explicit, and the vendor's own emphasis on verified steps suggests they know it.

Stands at F. Teammates own multistep operational processes end to end across several systems, and the named workloads are the right kind of evidence because each is inherently multi-step: invoice processing across shipping portals, financial reconciliation, recurring reporting, order intake and advertising asset ingestion. Reconciling finances means pulling from one system, comparing against another and writing a result somewhere third. The vendor's framing draws the distinction this axis cares about explicitly: unlike assistants and chatbots that help one person work faster, teammates TAKE OWNERSHIP OF RESPONSIBILITIES previously managed manually. Ownership of a process rather than assistance with a task is the difference between orchestration and completion, and the outcome figures cited, tens of thousands of hours reclaimed annually at named customers, are only achievable if work completes unattended. Execution is deterministic: once taught, a teammate follows verified steps every time rather than re-planning each run. That is a design choice with a real trade, and worth carrying to a comparison page. It buys reliability across hundreds of thousands of repetitions, which is what finance operations need, at the cost of adaptability when a process changes. What is not documented is multi-agent composition: nothing describes teammates coordinating with one another or handing work between them.

F>P on a consistency call. The July basis credited autonomous background operation on recurring processes running unattended, and asserted scheduled and event-driven operation. The first part is documented; the second is inferred from it. What the vendor actually says is that teammates work autonomously in the background, take ownership of responsibilities, and handle recurring processes including reporting and order intake. Recurring work running unattended implies a schedule exists, but no scheduling capability, cron, inbound webhook or event subscription is named anywhere. SEMA4-AI SITS AT PARTIAL ON THE IDENTICAL CLAIM in this same batch, that agents work around the clock finding and completing work autonomously, and grading Convey higher on the same evidence would be arbitrary. The channel class is absent entirely, which is coherent: nobody talks to these teammates, they do back-office work unattended, and the operator's surface is management rather than conversation. One property worth carrying that does not affect the grade: because teammates operate a computer directly, work can reach them through interfaces that have no programmatic trigger at all, which is a genuinely different reach from webhook-driven platforms even though it is not a documented trigger class.

Stands at P. What is documented is context gathering rather than a maintained knowledge structure: teammates LEARN TASKS AND GATHER CONTEXT, acquiring what they need by observation and demonstration as an operator describes a process or shares their screen. That is a real and unusual form of grounding, and worth stating precisely because it does not fit the usual shape of this axis. Most vendors ground an agent by pointing it at documents; Convey grounds one by showing it the work. What the teammate ends up knowing is a procedure and the state of the systems it touches, not a corpus it retrieves from. Under the persistence line that is Partial. The taught procedure persists, and teammates read live data from connected systems and screens at run time, but no knowledge base, document index, ingestion pipeline or retrieval configuration exists as a maintained object the customer curates. Grounding is assembled per run from whatever the systems show. Worth noting for completeness: for the workloads described, invoices, reconciliations, order intake, that may be the right architecture. The ground truth for an invoice is the invoice in the portal, not a document in an index, and copying it into a vector store would add staleness rather than remove it. Confidence is medium because no documentation site exists to confirm what, if anything, is held between runs.

F>P. The July basis read learning by demonstration as memory: a teammate learns a task once and retains it, running reliably across hundreds of thousands of repetitions. What that describes is a PROCEDURE CAPTURED AT BUILD TIME AND REPLAYED, which is configuration rather than accumulated state. The teaching step is genuinely impressive and unusual, an operator shares a screen and the teammate encodes what it saw, but it is the equivalent of authoring a workflow, just authored by demonstration instead of by drawing a graph. Once taught, execution is described as DETERMINISTIC, following verified steps every time. Determinism is the opposite of accumulation: a teammate that behaves identically on run one hundred thousand and run one is by definition not learning from the intervening runs. So persistence is real and clears N: the taught procedure survives, teammates hold their own identity, and they own a responsibility over time. But nothing documented describes a teammate carrying context between runs, learning from a correction, or updating its own procedure from experience. That is the line applied to imbue, tektonic and maisa in this session. One third-party analyst makes the same observation from the other direction, noting that vendors without persistent state risk being read as sophisticated macro recorders; the demonstration-teaching is what separates Convey from that, not memory.

F>P. The July basis credited role-based access, permissions configured in partnership with IT, enterprise governance, and operators overseeing teammates. All of that is real and none of it is an approval gate. What is documented is ACCESS CONTROL AND MANAGEMENT: IT controls identity management, permissions and governance, each teammate is provisioned with its own identity, and operations teams configure and manage teammates within those established boundaries. That constrains what a teammate may reach and who may change it. It does not pause a run for a human to approve what the teammate is about to do. The distinction matters more here than almost anywhere in this index. These teammates operate a computer directly and own invoice processing and financial reconciliation unattended across hundreds of thousands of repetitions. A misconfigured procedure replayed at that volume against a finance system is precisely the case where a checkpoint before a consequential write would matter, and none is documented. The architecture supplies a different kind of safety, and it is worth crediting even though it is graded elsewhere: teach-then-replay deterministic execution means a teammate follows verified steps rather than improvising, so the failure mode is a wrong procedure faithfully repeated rather than an agent inventing an action. That is a real answer to the reliability problem and not an answer to this axis.

Stands at F. Both halves of the section 7 conjunction are met and the attestation half is stated with the type specified: SOC 2 TYPE II CERTIFIED AND HIPAA COMPLIANT, with regular penetration testing. The control surface is the interesting part and it is well matched to the risk. Each digital teammate is provisioned with ITS OWN DISTINCT IDENTITY and role-based permissions, configured in partnership with IT, so a teammate is a governed principal in the customer's directory rather than a script running under a shared human account. That is the right answer to the question this product raises: if software is going to log into systems and click through them, whose credentials is it using and what can it reach. Treating each teammate as a named identity makes the work attributable and revocable. IT retains control of identity management, permissions and governance while operations teams configure teammates within those boundaries, which is the separation of duties an enterprise security review looks for. Data handling is documented separately: customer data is kept out of model training and zero-retention agreements are maintained with model providers. For a company a year past founding this is a stronger posture than most, and the certifications are presumably what made NBCUniversal and Samsara possible.

Stands at P. Operators manage and oversee teammates, and reporting on completed work exists, with the vendor and customers citing measured outcomes such as hours reclaimed and processes automated. Each teammate holds a distinct identity, which means its actions are attributable in principle to a named principal rather than lost in a shared account. What holds it below Full is that nothing documented reconstructs a run. There is no execution trace, no record of which applications a teammate opened, which fields it read or wrote, or why it took one path rather than another, and no audit log a compliance function could query. That gap is sharper here than the grade conveys, and it is worth flagging rather than filing. These teammates work by CLICKING THROUGH APPLICATIONS, which is the hardest kind of activity to reconstruct after the fact: an API call leaves a request log somewhere, while a screen interaction leaves nothing unless the platform records it deliberately. When a finance reconciliation running unattended across hundreds of thousands of repetitions produces a wrong result, the question is which step went wrong, and the published material does not say how a customer would answer it. Deterministic execution mitigates this by making behaviour predictable, and per-teammate identity gives the systems being operated their own audit trail, but neither is the same as observing the teammate.

P>N. The July basis credited SaaS delivery with customer data ownership and zero-retention model terms, then noted self-host and residency are undocumented. Those commitments are real and they are Sec properties, not Dep ones: zero retention with model providers governs whether a third party keeps a copy, not where the platform runs or who controls it. On what this axis actually measures, nothing is documented. Convey is delivered as SaaS with no self-hosting, no on-premises option, no deployment into the customer's own cloud, no virtual private cloud, no customer-managed keys and no selectable region. Under the standing rule, nothing documented equals N, and the same reading moved zapier and teneo this session. The residency question is sharper here than the grade suggests, and it is worth flagging for a buyer rather than leaving implicit. A teammate that operates a computer directly holds credentials to the customer's systems and sees whatever appears on screen while working invoices, financial reconciliation and advertising assets. That is a concentrated exposure, and where the platform runs and how those sessions are stored is a question the published material does not answer. The vendor is a year old with a June 2026 Series A, so a hosted-only posture is expected rather than surprising, and this may change as enterprise deals demand it.

P>N, and the July basis argued for N in its own words while grading P: teammates are built per customer by demonstration RATHER THAN shipped as a prebuilt agent library, and a template catalogue IS NOT THE MODEL. That is correct and it is the whole answer. Under the 31 August wider bar the question is whether the customer receives packaged assets ready to adopt by any route, and here they receive a method, not an asset. Every teammate is created by an operator demonstrating their own process, which is the opposite of adopting something prebuilt. The near-miss is the white-glove motion, where Convey specialists help stand up the workforce and then hand back control, with the vendor citing about three hours to onboard an end-to-end teammate. That is a service, and services are not packaged assets; the same refusal held moterra at Partial for professional-services delivery and applies more strongly here because Convey does not even claim a component library. The absence is coherent rather than a gap. A platform whose premise is that every enterprise's invoice process is idiosyncratic enough to need teaching by demonstration has no use for a generic invoice-processing template, and shipping one would undercut the pitch. Worth revisiting after a year of enterprise deployment, when common patterns across NBCUniversal, Faire and Samsara might justify a starter library.

Stands at N. No customer-facing model choice, model picker, bring-your-own-key arrangement or routing configuration is documented anywhere. The platform manages the model layer itself, with the only model-related commitment being zero-retention agreements with providers, which governs what those providers may keep rather than which of them the customer picks. That is a coherent position for this product and the vendor is candid about it. The framing is a managed, outcome-focused platform, with the vendor's own material acknowledging that teams wanting an open model choice will find something else. Selling outcomes rather than infrastructure means owning the stack that produces them, and letting each customer swap the model underneath a taught procedure would undermine the deterministic reliability the whole pitch rests on. Worth carrying to a comparison page as a trade rather than a deficiency: a buyer with a model-approval policy has no lever here, and a buyer who wants a teammate to keep working identically across hundreds of thousands of repetitions is better served by the vendor controlling it. Those buyers are frequently the same organisation, which is where the friction will appear. Confidence is medium because no documentation site exists and an enterprise arrangement could exist unpublished, though the vendor's positioning points the other way.

P>N, correcting a direction error. The July basis credited connecting programmatically to company systems and integrating with IT-managed identity. Both run INBOUND: Convey calling the customer's systems, and the customer's identity provider governing Convey. Neither makes Convey callable from outside, which is what this axis measures. That is the same correction applied to maisa earlier in this batch, where hundreds of API integrations were credited on Ext when they belonged on Int. Nothing documented provides a public API, an SDK, a webhook endpoint, an MCP server or any route by which an external system invokes a teammate or manages the platform. Under the standing rule, nothing documented equals N. Convey publishes no developer documentation site that I could reach, which is consistent with a sales-led product a year past founding. The positioning explains it and is worth stating plainly for a buyer: this is a managed outcome platform for non-technical operators, not a developer surface. The vendor's own framing draws the same line, that teams wanting a self-serve developer framework will find a managed platform instead. A buyer who needs teammates triggered from their own orchestration layer, or wants to embed one in an internal tool, has no documented route today.

Stands at N under the 31 August Eval bar, which needs a readable comparable result for Full and a quality gate for Partial. Neither is documented. The near-miss is DETERMINISTIC EXECUTION, described as teammates following VERIFIED STEPS every time after being taught. Verified is doing tempting work in that sentence and it should be refused: it describes steps a human confirmed at teaching time, not a system that tests output at run time. A procedure validated once and replayed faithfully is reliability engineering, and the same refusal was applied to taskade's durable execution and cassidy's retry handling in this batch. What is absent is anything a customer could point at their own workload: no test suite, no scoring, no accuracy measurement, no comparison of one version of a taught procedure against another, and no gate that checks a teammate's output before it is committed. The absence is more consequential here than elsewhere, and it belongs on a comparison page. Teammates run finance work unattended across hundreds of thousands of repetitions, and the vendor's own reliability argument is that they do the same thing every time. That is exactly the situation where a customer needs to measure whether the same thing is still the right thing after a portal changes, and nothing documented lets them.

Stands at F and is one of the clearest positive computer-use cases in the index, which matters because this axis has been refused repeatedly across this session for sandboxes, terminals, screenshots, scraping and browser testing. The capability is stated as the product's reason for existing rather than as a feature: teammates OPERATE A COMPUTER DIRECTLY, CLICKING THROUGH APPLICATIONS THE WAY A PERSON WOULD, specifically so they can reach LEGACY TOOLS WITHOUT MODERN APIS. That is the axis stated almost verbatim, which asks for operation of software the vendor does not control because no programmatic interface exists. The commercial evidence corroborates it. Faire is cited automating hundreds of hours of invoice processing ACROSS SHIPPING PORTALS, which is exactly the class of system that has a login and no API, and the teaching mechanism only makes sense for interface-level work: an operator shares their screen and the teammate learns what to click. With orby-ai watchlisted, this becomes a useful reference point for the axis alongside sema4-ai's Robocorp libraries and gumloop's Chrome extension, and it is the strongest of the three on positioning: the other two ship computer use among other things, whereas Convey's proposition depends on it. Confidence is high because the claim is central, specific and repeated across the vendor's own materials and independent coverage.