FurtherAI

The July 2026 Connectors launch documents a connector library through which every FurtherAI workflow reads from and writes to the customer's existing systems directly, with enrichment sources layering third party risk and compliance data on top of internal records, each connector's scope (data reached, actions it can take, authentication) reviewed before it ships, and admins controlling which connectors are active. Agents taking authenticated action in outside systems through a named catalogue is the Full line.

SourceFurtherAI, furtherai.com/blog/furtherai-launches-connectors-one-workspace-every-insurance-systemread 2026-09-07

The August 2026 Multi-Agent System launch documents specialised agents coordinated end to end across an insurance workflow: a dedicated agent preps a submission with the scope an underwriting assistant would have, hands off to the next agent in sequence or to the underwriter whose judgment the next decision requires, and the context travels with the work. Multi step, multi agent execution with human participants in the sequence is the axis at Full.

SourceFurtherAI, furtherai.com/blog/furtherai-launches-multi-agent-system-running-the-full-insurance-relayread 2026-09-07

Agents clear submissions against the customer's underwriting guidelines and appetite, with Guideline Checks a named workflow; the guidelines corpus is loaded once and checked against on every submission, a maintained structure over the customer's knowledge that persists and stays queryable across runs, and a customer describes the platform as data centralisation with an intelligence layer putting information at underwriters' fingertips.

Persistence is the discriminator and it is met by the standing guidelines corpus. Run time enrichment with third party and web data is spent on Integrations, not here.

SourceFurtherAI, furtherai.com and furtherai.com/blog/furtherai-launches-multi-agent-system-running-the-full-insurance-relayread 2026-09-07

The workspace pauses for human review inside workflows, and the Multi-Agent System post documents the mechanism: when a decision needs a human, the agent brings it to the person who owns it with the context they need, presented where they need it, with the underwriter's judgment gating the sequence; triage recommendations are drafted with reasoning laid out for review rather than executed. This is the vendor's own review and approve surface in the workflow, not a gate the customer already owns.

SourceFurtherAI, furtherai.com/blog/furtherai-launches-multi-agent-system-running-the-full-insurance-relay and furtherai.com/blog/building-the-ai-workspace-for-insuranceread 2026-09-07

Both halves are on the vendor's own security page. Compliance posture: SOC 2 Type II described as independently audited controls, ISO 27001, GDPR and HIPAA, with a Vanta security portal linked. Access surface: authentication with SSO, role based permissions, audit logs, customer managed encryption keys, and customer controlled retention and deletion. The earlier Partial rested on single tenant isolation alone and predates this page. Audit logging is recorded here as supporting evidence and spent on Observability.

SourceFurtherAI, furtherai.com/securityread 2026-09-07

The security page names audit logs in its access control feature alongside SSO and role based permissions, and each triage recommendation is drafted with the reasoning laid out for the reviewer, giving per decision traceability. What is not documented is a runtime trail of agent actions a customer can open to reconstruct why an agent did what it did across a workflow run: the audit logs are presented in an access control context, and the a16z post's auditability claim names no surface. Partial with the gap named.

SourceFurtherAI, furtherai.com/securityread 2026-09-07

Near line, refused: the Multi-Agent System post documents context travelling with the work between agents in a sequence, which is orchestration state, spent on Workflow orchestration; deleting it would delete the case in flight. No memory layer with a stated scope and lifetime, per user, per team or per workflow, and no session state a buyer can locate and delete is documented on any page read.

SourceFurtherAI, furtherai.com/blog/furtherai-launches-multi-agent-system-running-the-full-insurance-relayread 2026-09-07

The security page states the customer retains full control over data, deciding what gets uploaded, how long it is kept and where it lives, with customer managed encryption keys; the OpenAI collaboration post names single tenant deployment.

Under the 6 September 2026 ruling on Q41 the mechanism decides, not the claim: a where-it-lives sentence with no named region list, no named customer environment option and no documented selection surface is a claim that a mechanism exists, which is Partial with the gap named, and confidence is high because the grade is about scope rather than doubt. Single tenancy is isolation, not residency.

SourceFurtherAI, furtherai.com/securityread 2026-09-07

The homepage presents a named set of prebuilt workflows a customer selects among and adopts: Submission Intake, Submission Processing, Policy Check and Compare, Underwriting Audit, Claims Intake, FNOL, SOV Mapping, Contract Analysis, Authority Check, Tower Analysis, Bind Order Verification, Cyber Analysis and Guideline Checks, and the vendor states most partners start with one workflow and expand from there, with case studies showing single workflow adoptions.

Under the removal test each workflow does its own job when selected and the others remain whole, which is customer adoptable packaged assets, the Full bar by the named set route. The forward deployed engineer configuration is the deployment method, not an absence of packaging; the earlier Partial withheld Full on the lack of a self serve template library, which is not the bar.

SourceFurtherAI, furtherai.com and furtherai.com/blog/building-the-ai-workspace-for-insuranceread 2026-09-07

Work reaches the agents as submissions, claims and FNOL files arrive, forwarded into intake and processed around the clock, and the connectors let workflows read from the customer's systems; no schedule, webhook or broader event trigger a customer configures is documented on the pages read. Inbound document and case arrival across a few channels is Partial with the gap named.

SourceFurtherAI, furtherai.com and furtherai.com/blog/furtherai-launches-connectors-one-workspace-every-insurance-systemread 2026-09-07

The vendor documents routing across OpenAI, Anthropic and other providers, using whichever delivers the best outcome for a given workflow step, and states the moat is in the workflow engine rather than the underlying model, with the OpenAI collaboration post confirming multi provider work. That is disclosed vendor internal multi provider routing, which is the published Partial on this axis: the axis measures who chooses, and no customer or admin control of the model is documented on any page read. The earlier Full credited the routing itself, which is the how-many error.

SourceFurtherAI, furtherai.com/blog/building-the-ai-workspace-for-insurance and furtherai.com/blog/furtherai-x-openai-improving-ai-agents-for-insuranceread 2026-09-07

No API, SDK or MCP server for calling FurtherAI's own platform from outside is documented on any page read; the connectors run the other direction, FurtherAI's workflows reading from and writing to the customer's systems, which is Integrations, and the Microsoft Marketplace listing is a purchase channel, not an interface. Nothing documented for driving this platform is None.

SourceFurtherAI, furtherai.com/blog/furtherai-launches-connectors-one-workspace-every-insurance-systemread 2026-09-07

The documented evaluation is the vendor's own engineering practice: stress testing extraction against internal benchmarks of dense insurance documents, including the OpenAI collaboration on 900 page loss runs, whose improvements roll out to customers with no action required.

Under the ironclad precedent and the 7 September 2026 ruling on Q43, a vendor's own release practice is None on this axis: the artifact under test is the vendor's platform, the customer neither runs the layer nor pays for it, and no harness, scored test cases, quality gate or optimisation loop over the customer's deployment is documented. The published 95 percent accuracy figure is a marketing benchmark, not an eval surface. The earlier Partial credited the internal practice.

SourceFurtherAI, furtherai.com/blog/furtherai-x-openai-improving-ai-agents-for-insuranceread 2026-09-07

The connectors post states every workflow reaches the customer's systems through the connector library, reading and writing directly, a positive statement that reach runs through integrations, which is None under the platform action rule. Near line named: enriching with third party and web data appears in feature copy without a mechanism; the third party portion is data providers, spent on Integrations, and no run time browser, fetch tool or computer control is documented.

SourceFurtherAI, furtherai.com/blog/furtherai-launches-connectors-one-workspace-every-insurance-systemread 2026-09-07