Before giving an AI agent more responsibility, verify where its data comes from, who checks it and who approves decisions with real consequences. Keep the evidence needed to reconstruct the outcome. Spej's recommendation is to build that data foundation for agentic AI in insurance around one narrowly scoped workflow, with an accountable owner and measures agreed before implementation.
By Sean Blair, Co-Founder & Chief AI Officer, Spej | November 11, 2026
Chris Newman, CEO of ACORD Solutions Group, described data trust as a condition for greater automation in an interview with Elizabeth Blosfield at InsurTech Connect. His questions concern data origins, verified sources and a carrier's ability to account for its decisions. Those are Newman's assessments, reported by Insurance Journal, and a useful starting point for a carrier evaluating agentic AI.
Why the data layer comes first
An agent's proposed action needs a traceable basis. For a claims workflow, a reviewer should be able to identify which policy terms and claim records informed a recommendation. If documents conflict, the workflow needs a defined way to stop and ask for review. Otherwise, a convincing summary can hide an unresolved difference in the underlying records. Insurers need to be able to show how an AI-assisted decision was reached, which depends on traceable data in the trusted data layer.
Newman recommends verified data and human oversight as guardrails before greater automation. He also suggests insurers should anticipate requests for evidence and accountability as scrutiny increases. That is his assessment of what could happen; the interview establishes no specific regulatory mandate. Insurance Journal
Spej's view is that agent features are becoming commodities. A carrier's practical advantage comes from choosing useful work, keeping a person responsible and checking whether the workflow improves. Data preparation belongs inside that decision because the cost of resolving uncertain inputs affects whether automation is worth doing.
Define a trusted data layer in operational terms
Spej recommends recording the source and owner of each critical data element, alongside verification status, freshness and permitted use. A field that appears in several systems needs an agreed rule for which record governs the workflow. Someone also needs authority to resolve a disagreement when that rule doesn't produce a usable answer.
This can be a set of controls across existing systems. The design should answer questions a reviewer can actually investigate: Which record supplied the effective date? Was it extracted from a document or read from a structured field? Has someone verified the extraction? What happens if an endorsement changes the relevant terms?
Approved data use needs the same attention. A tool's access should match the work it's allowed to perform. Spej's guidance on shadow AI and approved data use can help frame that discussion without assuming every available record belongs in every AI workflow.
Trace a claims review workflow from intake to action
Consider a hypothetical design exercise for claims document review. This describes a possible workflow, separate from Spej's published engagements. Start with document intake and identify the source record for each file. Preserve enough information to connect extracted content to the original document and distinguish a later revision from the version previously reviewed.
Next, define how the system finds the relevant policy and policy version. A matching failure should create a review task. An ambiguous match should reach a person with the competing records visible. The agent's job at this stage could be to assemble evidence and flag inconsistencies for review.
The reviewer then needs a usable work item: the issue, supporting passages and the records that disagree. Keep the proposed recommendation separate from the final disposition. Define what happens when the reviewer requests more information, rejects the suggestion or sends the matter to someone with different authority.
For underwriting, apply the same questions to submission review. Identify which submitted information has been checked, which remains uncertain and who can approve the next action. Keep the scope narrow enough that the team can inspect how the workflow behaves when the inputs are incomplete.
Put human approval before action with real consequences
Spej advises naming the approver before implementation. Specify the decisions that person can make and the evidence they need. Give the reviewer enough time to check the recommendation and a clear route to reverse, modify or escalate it. A queue that merely asks someone to click approve provides little help when the underlying records disagree.
Record the inputs, relevant policy version, AI output and human disposition. Preserve the reasons for an override or escalation where those reasons are needed to understand the result. Strategy and governance work should establish these decision rights; a named accountable owner should keep them current.
The NAIC's AI Risk Evaluation Supplement v7.0 is marked DRAFT. Its optional exhibits address AI-system inventories and data sources, and its discussion of human oversight includes reviewer authority and documented overrides. It says the supplement does not independently create substantive legal duties. These are supporting reference points for the design questions above. NAIC draft supplement
The working group's agenda schedules consideration of adoption for November 16, 2026. That outcome is unknown on this article's publication date. The agenda doesn't establish adoption or a new mandate. NAIC working group agenda
Work with the systems the carrier already has
Newman observes that carriers have invested heavily in their systems and cannot change them overnight. He describes his company's approach as connecting systems and converting data into usable, trusted formats while accommodating different levels of digital maturity. This is his description of ACORD Solutions Group's approach. Insurance Journal
For a carrier considering implementation, Spej's advice is to check the selected workflow's access paths and data mappings before choosing an architecture. Confirm which records can be read, how updates become available and whether the workflow can identify the version it used. Test what happens when a connection fails or a required field is absent.
Existing systems may support a narrowly scoped workflow with limited changes. They may also expose constraints that change the scope or make a proposed workflow unsuitable. AI implementation should include that assessment, with uncertainty reflected in the delivery plan.
Measure before expanding
Agree a baseline for handling time and rework before the workflow changes. Include time spent checking recommendations, correcting extracted information and investigating exceptions. Track operating costs alongside the business measure so the team can see what the improvement requires to maintain.
Spej's published claims-review case describes document checks against policy terms and previous records, with people checking flagged issues. Its separate operating-model case describes prioritization and governance work. They illustrate distinct scopes; neither establishes an outcome or delivery timetable for another carrier.
Set conditions to continue, change or stop the workflow. Review recurring exceptions and whether staff are using the output. If review effort outweighs the benefit, revisit the workflow and its inputs. AI Office support can provide continuing review within an agreed scope, while the carrier retains business accountability.
Where AI falls short
Verified sources can still be incomplete or stale. Traceable inputs help a reviewer inspect a result, although they cannot guarantee a correct interpretation. Human review also needs time, authority and access to evidence. Assigning a reviewer without providing those conditions leaves unresolved risk in the workflow.
Spej advises pausing automation of decisions with real consequences when reliable inputs or accountable ownership are missing. A decision record supports investigation and review; it doesn't guarantee regulatory compliance. A carrier should assess its own obligations through the appropriate internal and professional review processes.
FAQ
What is a trusted data layer for insurance AI?
A trusted data layer records where operational information came from, who owns it and how it was checked. It also defines freshness and permitted use. Spej advises making those records available within the workflow so reviewers can investigate conflicts. The layer can span existing systems; it doesn't require one particular product.
Can insurers trust AI decisions?
Trust needs to be established for a defined workflow through evaluation and continuing review. Verified inputs help, although AI can still produce an incorrect interpretation. Spej recommends retaining human approval for consequential decisions, showing reviewers the supporting records and tracking errors and overrides before extending an agent's authority.
Must carriers replace legacy systems before agentic AI?
Newman describes preserving existing investments while connecting systems and making their data usable. That is an integration approach, with feasibility assessed for each carrier. Spej's advice is to check access, record matching, permissions and policy versions for the selected workflow before deciding whether existing systems need changes or replacement. Insurance Journal
How should a P&C carrier start?
Choose a bounded workflow with a named business owner and a clear final action. Verify its inputs, define how missing or conflicting records reach a person, and name the approver. Establish a baseline for handling time, rework and review effort, then assess the workflow's results before expanding its scope.
Plan the first workflow
Bring one workflow and its data gaps to a conversation with Spej. Spej's project-based services include an AI Roadmap starting at $10K and Design & Build starting at $20K. AI Office is starting at $10K a month. Scope determines final pricing; the initial discussion should establish which work, ownership and evidence your carrier needs.