TruVelocity™ for AI TEVV fitment: Intelligent Core Copilot (L&AH)
17 Nov 2026 · Assessment of the customer's deck Intelligent Core Copilot · Immersive AI Kit for Innominds (11 slides) against what TruVelocity™ does today.
Jump to: 1 Their product2 Fit, concept by concept3 Their bar, our test types4 Pilot5 What we need6 Gaps7 Questions8 Next steps
1Their product, in one minute
"L&AH Copilot": one conversational layer across the life, accident and health insurance lifecycle, from RFP and underwriting through enrollment, servicing, billing and claims.
- Inquire: every answer must trace back to the record, document or transaction it came from.
- Act: bounded operations on a named record, inside the user's rights, with an audit trail (8 claims, 17 policy and member, 3 operations actions).
- Orchestrate: agents that interpret a request, email, document or event, gather evidence, invoke actions, validate and route exceptions to a person.
- Boundaries: the Copilot must never price or approve underwriting, adjudicate or pay a claim, generate bills or refunds, act outside a user's authority, or submit what the user has not confirmed.
- Their own quality bar (slide 11): 80 prompt variants across four endorsements, and six things "understood the request" must mean.
"Recognizing the sentence is the easy part. Resolving the subject, validating against the group's configuration, warning about consequence, and refusing the rest, is the product." (customer deck, slide 11)
That sentence is TruVelocity™'s premise: judge what the agent did, not only what it said.
2Fit, concept by concept
| Their deck | What TruVelocity™ does with it |
|---|---|
| Three layers: Inquire, Act, Orchestrate fits | TruVelocity™'s three test profiles: conversational (chat inquiry), task (one bounded action) and workflow (multi-step and multi-agent), each already proven on a reference agent of that kind. |
| "Every answer traces back to its record" fits | Answers are checked against independent evidence: the facts the reply must state, the tools actually called and grounding. SE-01GF-01GF-05 |
| 28 governed actions: permissioned, human-confirmed, audited fits | The core of TruVelocity™: it reads the affected records before and after each turn (read-only) and catches unrequested writes, "done" with nothing written, actions without approval and actions outside a role's rights. SE-02SE-03SE-04SS-03AT-05 |
| Three operating modes fits | One catalog, checked per mode: answer only expects no writes at all; propose for approval expects an approval prompt, and nothing written when the user declines; execute under permission expects exactly the right records changed, and nothing else. |
| "What Copilot is not": never adjudicate, approve, pay or exceed authority fits | Negative tests at critical severity. TruVelocity™ has caught this class of defect in our reference claims agent, which approved a payout on its own. |
| Proof story: a bad effective date refused, the rule explained fits | Precondition enforcement, plus the database check that proves the refusal really left the record untouched. TT-06SE-04 |
| 80 prompt variants per endorsement fits | TruVelocity™'s wording model: every request asked many ways, each case scored PASS / FLAKY / FAIL with a 95% confidence range, and reliability over repeated tries (pass^k). IU-06RC-01RC-02 |
| Agents triggered by inbound email or scanned PDF fits, needs data | The task profile with queue events (Service Bus, Kafka) and attachments; extracted values are checked against expected ones, which needs reference documents with known correct values. |
| "Route exceptions to a person with the context attached" fits | Escalation and handoff, agent-to-agent handoffs (recorded by TruVelocity™'s agent-to-agent tap) and errors passing down a chain. AT-04AT-08SS-09 |
| Audit trail and conversation logs fits | Auditability CP-06; their audit log also becomes one more independent evidence source. |
| Multi-turn context ("set member context, then…") partly | Multi-turn conversations are supported today; long-conversation stability CQ-05 and memory correctness AT-07 are not validated yet (see section 6). |
3Their quality bar is our test catalogue
Slide 11 lists six things "understood the request" must mean. Each maps to test types TruVelocity™ already measures:
| Their bar (slide 11) | Example | TruVelocity™ test types |
|---|---|---|
| 1 Resolve the subject: never guess | "Change class for John Doe." | IU-02 disambiguation ES-02 slot elicitation |
| 2 Map words to values | "Promoted to management", "next quarter" | ES-01 slot extraction against expected values |
| 3 Validate the configuration | "Move him to Class B." | TT-06 precondition enforcement |
| 4 Split the request | "His ID and his name both need fixing." | IU-03 multi-intent decomposition |
| 5 Warn before submit | "To a class with no life cover." | CP-03 disclosure SE-02 approval vs state |
| 6 Refuse or challenge | "Move everyone in Warehouse." | IU-04 out of scope AT-05 scope boundary SS-03 privilege abuse |
TruVelocity™ measures every run against all 71 test types of the Intent-Based Testing Catalogue and reports gaps instead of hiding them.
4Proposed pilot, built on their own material
- Servicing endorsements (class, employee ID, hours worked, location). They already wrote 80 prompt variants, and proof story 3 gives the expected behaviour. Turn them into a catalog and run it in all three modes: the fastest route to a result they will recognise.
- Claims inquiry: "Why is this claim pended and what is missing?", grounded against the claim record.
- One agent end to end: the Member Profile Changes agent, including approvals and routing exceptions to a person.
- Negative tests across the board: never adjudicate, never pay, never exceed authority limits, no bulk changes.
| Operating mode | What TruVelocity™ expects | What would be a defect |
|---|---|---|
| 1 Answer only | Correct, grounded answer; no record changes | Any write; an answer the record does not support |
| 2 Propose for approval | The prepared change and its evidence, then an approval prompt | A write before approval; a write after the user declines; approving the wrong change |
| 3 Execute under permission | Exactly the requested records changed, with reason code, effective date and audit entry | Extra or missing changes; "done" with nothing written; action beyond authority limits |
The output is the same as for our own pilots: an execution report per run, a release gate, trends across builds and the evidence portfolio.
5What we need from the customer
- An interface TruVelocity™ can drive. A chat API, a WebSocket or an OpenAI-compatible endpoint all work today. If the Copilot is reachable only through its web UI, that is a gap: TruVelocity™ lists UI-only cases for manual testing and has no automated UI driver yet.
- A non-production environment with seeded test data that can be reset, and test users per permission and authority profile.
- Independent, read-only evidence:
- Tool calls: if their tools and actions are exposed through MCP, TruVelocity™'s proxy records them directly; otherwise OpenTelemetry traces, or their execution history and audit API.
- Data: read-only access (SQL or REST) to policy, member and claims records, to read them before and after each turn, ideally through a path separate from the Copilot's own tools.
- Agent handoffs: their orchestration logs, or TruVelocity™'s agent-to-agent tap.
- Each environment's settings: permissions, required fields, authority limits and approval behaviour. Their deck says these differ per environment, and the expected results depend on them.
6Gaps and risks, stated openly
- Long, context-carrying conversations. Multi-turn is supported; long-conversation stability (CQ-05) and memory correctness (AT-07) are among the 19 test types TruVelocity™ has not validated yet.
- Document extraction (True Bill Verification, RFP from a PDF) needs reference documents with known correct values, supplied or approved by the customer.
- Size of the surface. 52 tools, 28 actions and 25 agents is a large catalog to write; a phased rollout starting from slide 11 keeps the first result fast.
- Bias and fairness (CP-04) is not covered yet. The Copilot does not decide underwriting, so this matters less, but a regulated customer may still ask.
- UI-only access would limit automation (section 5).
7Questions to put to the customer
- How is the Copilot accessed: API, WebSocket, or only the web UI?
- Are the inquiry tools and governed actions exposed through MCP, REST or internal SDK calls? Do they emit OpenTelemetry traces?
- Is there a non-production environment with resettable data, and can we have a read-only data path into it?
- Can they share the 80 prompt variants and the expected outcome of each?
- Which operating mode is each environment in, and what are its authority limits and approval rules?
8Next steps
- Share this assessment and the questions in section 7 with the customer.
- On answers, agree the pilot scope (section 4) and environment access (section 5).
- Draft the starter catalog for the four endorsements on slide 11 from their prompt variants; run it once access is in place.
- Present the first execution report, release gate and coverage against the 71 test types.
Internal to Innominds. This page summarises a customer's deck; share outside Innominds only as agreed with the customer.