WhitepaperPrepared for: Executive Leadership and Boards of Mutual Insurers

The Core System Advantage

Why Your PAS Provider is the Right Advisor on AI

Executive Summary

AI advice belongs where the work already happens

Every P&C insurer is now being advised on artificial intelligence. Consultants, point-solution vendors and brokers all arrive with a point of view. The scarce commodity is not enthusiasm for AI. It is advice that is accountable for the outcome once the pilot ends.

This paper argues that the vendor best placed to advise a P&C insurer on AI is the provider of its core system — the policy, billing and claims platform where underwriting and settlement actually happen. Not because core providers know more about models, but because they own what AI needs: the data of record, the workflow the decision sits in, and the audit trail the regulator will ask for.

Advice divorced from those three assets can only produce recommendations. Advice attached to them produces shipped capability — and carries the consequences if it is wrong.

The right AI advisor is not the one with the best model. It is the one that has to live with the result inside your renewal cycle.

What this whitepaper covers

Why AI advice fails when it is separated from the system of record
How data gravity, workflow context and auditability concentrate advisory credibility in the core provider
A decision framework for testing any vendor’s claim to advise you on AI
Why Cognition+ is structurally suited to this role for North American mutuals

The Inflection

AI has moved from pilot to production

The experimentation phase is closing. Celent reports that 48 percent of insurers now run generative AI in production rather than in pilot.¹ The board question has shifted from whether to invest to how to industrialize what already works.

Production changes the advice required. A pilot needs a model and a use case. A production capability needs data lineage, exception handling, audit logging and change control — core system disciplines, not model disciplines.

What production AI actually demands

Clean policy and claim data
Workflow integration
Role-based access
Human-in-the-loop review
Audit logging
Version control
Exception routing
Regulatory reporting
Reversibility

Almost none of that list is about the model. Nearly all of it is about the system the model is asked to operate inside — which is why advice sourced from outside that system consistently underestimates the work.

The hard part of insurance AI is integration and governance.

1. Celent, GenAI in Insurance: Global Year in Review – 2025, 11 January 2026.

The Advisory Question

Four candidates for the AI advisor role

Four kinds of vendor now offer P&C insurers guidance on AI: the consultant, the point-solution vendor, and the core system provider. Each brings real capability. Only one of them is already accountable for the transaction the AI is meant to improve.

The consultant view

“We will define your AI strategy.”

Strong on framing. Departs before the second renewal cycle.

The hyperscaler view

“We will provide the models.”

World-class platform. No view of your policy record.

The point-solution vendor knows one function deeply but sees only the slice of data it is given — and each added vendor is another contract and another seam in the audit trail.

What insurers need

Their real data, not a sample
The workflow, not the slide
Production, not a pilot
Governance a regulator accepts
Change their staff can absorb
Economics that last five years

What the core provider has

The system of record
The workflow every decision uses
The users who must adopt it
The audit and retention layer
A multi-client view of what works

Should AI advice come from a party that observes your operation, or from the party that already runs it?

The Hidden Work

Most AI value is won before the model runs

An AI demonstration is fast. An AI deployment is not. The gap between them is filled with work that belongs to the core system: reconciling coverage codes, resolving duplicate contacts, standardizing loss causes, and agreeing what a field means across thirty years of policy history.

Industry research consistently identifies data quality and governance — not model performance — as the most-cited barrier to enterprise AI implementation:

Fragmented policy, claim and contact data across systems and eras
Undocumented business rules embedded in legacy code and staff memory
No consistent identity for the insured across products and channels
No lineage showing where a value came from or who last changed it

Who can actually resolve these

The vendor that owns the data model
The vendor that owns the workflow
The vendor that owns the migration history
The vendor accountable for the record

An advisor cannot fix a data problem it has no access to.

Advice you cannot implement is not advice. It is a diagnosis without a treatment.

Data Gravity

AI is only as good as the record beneath it

Artificial intelligence has no independent knowledge of an insurer. Everything useful it says about a risk, a claim or a customer is inherited from the record it is grounded in.

In a P&C insurer, that record lives in the core system — and it does not travel well. Exporting it to a separate AI environment produces a copy that is immediately stale, partially governed and disconnected from the transaction that created it.

Grounding: the model reasons over the live policy or claim, not an extract taken last quarter.
Context: the same record carries coverage, endorsement history, prior losses and correspondence together.
Action: the output returns into the workflow as a task, a note or a decision, not as a report someone must re-key.
Evidence: every AI-assisted step is written to the same audit trail as every human step.
Reversal: a wrong output can be corrected through the same controls that govern any other transaction.

This is why the center of gravity for insurance AI is moving back toward the core platform. The advisor that already sits at that center can reason about what is achievable with an insurer’s actual data, on their actual release schedule — rather than what is achievable in principle.

Wherever the data and the workflow already live is where AI has to be advised, built and governed.

Governance and Regulation

Regulators are asking core-system questions

AI supervision has matured quickly. The NAIC Model Bulletin on insurers’ use of AI, adopted in December 2023, expects documented governance, risk controls and testing.¹ Colorado’s SB21-169 requires testing of external data and algorithms.² In Canada, OSFI’s Guideline E-23 extends model risk governance ahead of 2027.³

Read carefully, these regimes ask operational questions rather than technical ones: which record was used, who reviewed the outcome, how was the decision explained to the policyholder, and can it be reproduced two years later. Those answers are produced by the core system, not by the model.

What regulators actually ask for

Lineage — the data behind each AI-assisted decision
Oversight — evidence a person reviewed and could override
Reproducibility — the same inputs yielding the same explanation
Monitoring — testing for drift and disparate outcomes
Residency and retention — where records live and how long

Where advice breaks down

A vendor that cannot write to the audit trail cannot evidence the control it recommended.

1. NAIC, Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023.
2. Colorado Division of Insurance, Senate Bill 21-169, Protecting Consumers from Unfair Discrimination in Insurance Practices.
3. OSFI, Guideline E-23: Model Risk Management (revised), effective May 1, 2027.

The Economics of Advice

Who absorbs the cost of getting AI right

Advisory models differ less in what they recommend than in who carries the cost when the recommendation proves harder than expected.

Advice billed by the hour

Cost rises with complexity
Discovery is re-purchased
Integration is a change order
Success measured at handover
Maintenance becomes yours

Advice inside the platform

Cost amortized across clients
Discovery is already current
Integration is a release
Success measured in production
Maintenance is the vendor’s

Every dollar spent proving what is possible is a dollar not spent making it work in production.

Outcomes, Not Projects

AI should reduce specialized work, not add it

Insurance software has always let ordinary users do expert work: rating engines instead of manual calculation, workplans instead of adjuster memory. Good AI advice continues that trajectory; poor advice reverses it, asking insurers to staff an AI discipline before any frontline benefit appears.

The people whose productivity actually matters

CSRs
Underwriters

Adjusters
Accounting staff

An advisor inside the core platform delivers to those users directly, in screens they already use. An external advisor must first persuade someone to build that surface — and the insurer funds it.

What changes on a CSR’s screen, and in which release?

If the honest answer is a date rather than a demonstration, that date is the earliest your AI ambition can begin.

Why Your PAS Provider

AI delivered inside the workflow

The Cognition+ approach to AI is deliberate: use best-of-breed AI infrastructure, apply insurance domain knowledge on top of it, and deliver the outcome inside the platform experience. Two capabilities illustrate the pattern, and both are advisory positions expressed as product rather than as recommendation.

Cognition+ Help Agent answers natural-language questions grounded in approved Cognition+ documentation rather than generic web content.
Risk Intelligence is called by the platform during the normal underwriting workflow and returns structured risk observations, inconsistencies and exposure considerations natively in the UI.
Both are grounded in the live record, so nothing is exported, copied or re-keyed to make AI work.
Model selection remains a platform decision, insulating clients from a volatile vendor market.

Advice Expressed as Product

The live record

Policy, billing and claim data in the core system

→

Help Agent

Answers grounded in approved documentation

Risk Intelligence

Risk observations returned during underwriting

→

Inside the workflow

Delivered natively in the UI — nothing exported or re-keyed

From system of record to decision support, without leaving the platform.

Neither capability asks an insurer to become an AI builder. Neither requires a prompt library, an orchestration layer, or a new specialist role. That is what advice looks like when the advisor is also the implementer: the recommendation and the release are the same artifact.

Apply domain knowledge to specific workflows, and deliver outcomes inside the platform experience.

Decision Framework

Five tests for any AI advisor

The argument of this paper is structural, not promotional. Any vendor — including a core system provider — should be required to clear the same five tests before its AI advice is treated as credible.

Apply them to every party currently offering you a point of view on AI. The tests are deliberately about accountability rather than capability, because capability is now widely available and accountability is not.

The five tests

Access — does the advisor reach your live record, or a copy?
Workflow — does the output land where the work is done?
Evidence — is every AI step written to your audit trail?
Delivery — is the recommendation tied to a named release?
Continuity — is the advisor still accountable in year three?

How to read the result

A vendor that clears all five can advise you. A vendor that clears three can inform you. Know which you are buying.

Conclusion

Advice that ships

P&C insurers will not be short of AI advice in the years ahead. They will be short of advice that survives contact with a live policy record, an examiner’s question and a renewal deadline. That is a narrower field, and the core system provider stands in the middle of it.

What advice from outside the core system can offer

A strategy, a roadmap and a business case.
Benchmarks drawn from other markets and industries.
Model and infrastructure capability at global scale.
Deep expertise in a single function or process.
An honest view — and no accountability for delivery.

What advice from inside the core system adds

Grounding in the insurer’s actual data and workflow.
Governance evidence produced as a by-product of the transaction.
Delivery on a release schedule, with the vendor carrying the risk.

Insurers do not want an AI strategy. They want better underwriting, faster claims and a lower expense ratio.