The measured size of it
- 65%
- Agents using AI for work
- up from 37% in 2025
- 18%
- Principals with defined AI policy
- 22%
- Principals trusting AI with data
The pilot passes and the client file still fails, because the comparison the model produced cannot be tied to a clause, a page and a named reviewer. That is the failure worth designing against. It rarely presents as a wrong answer; it presents as an answer nobody can evidence three months later, when the client disputes an exclusion or the carrier asks how the risk was presented. Generative tools made the problem acute because they produce fluent output at speed, but the underlying issue predates large language models. Broker data sits in agency management systems, carrier portals, spreadsheets and inboxes, and those sources do not agree on what a field means. Standardising that meaning, exposing it through interfaces a model can call, and recording every call is the architecture work. Everything downstream is procurement.
Agent use has outrun agency control, and the gap is now measured
Liberty Mutual's 2026 Independent Agency Growth Study, reported by Insurance Business, puts the divergence in one dataset: of nearly 1,200 US agency principals and staff surveyed, 65% of agents said they had used AI for work in the past year, against 37% in 2025, while only 18% of principals said their agency had a well-defined AI policy and 22% said they trusted AI with client and business data. Staff adoption nearly doubled in twelve months. Oversight did not move.
The UK reads similarly. The Agent Bureau cites the DSIT UK Business Data Survey 2026 for a figure of 20% of AI-using finance and insurance firms holding a formal written policy, and builds its managed service around an advice boundary set by a senior manager, with recommendations and binding left with authorised staff and customer records kept on UK-hosted infrastructure. Deloitte UK's February 2026 perspective describes the same unevenness from the market end: some insurers scaling AI across servicing, fulfilment, underwriting and claims while others remain in pilot stagnation, and UK brokers repricing sharply when AI comparison and planning tools launched.
Read together, those sources say something narrower than rising adoption. They say client data is already moving through tools the firm cannot enumerate, and that the constraint on scaling is not model capability but the absence of a record. A principal who cannot list which tools touched a renewal file cannot supervise the work, cannot answer an E&O query about it, and cannot tell a carrier what the model was given.
Where the time goes, and which workflows earn the first integration
Up to 70% of a broker's working day goes on administration, according to Insurance Business's account of Cara's operating data, and that proportion is the honest starting point for any architecture decision because it says where the return sits. Rekeying data across carrier websites, generating ACORD forms, building coverage comparisons, checking policies and producing proposals are document-heavy, rule-bound tasks that do not need licensed judgement at every step. Cara puts a workflow that traditionally takes 90 minutes of manual work at roughly two minutes on its platform.
The order matters more than the tool. Agiliux's analysis of broker workflows draws the distinction that decides most implementations: an AI-native platform runs document processing, bordereau management and compliance checks on one data model, while a legacy system with bolt-on AI automates each task separately and leaves the joins between them to a person. Its worked example is a bordereau arriving as a PDF with renamed columns and fields in a different order from last month. Extraction is the easy part. Reconciliation, validation against treaty terms and exception flagging are where the days go. The same source names API readiness as the technical prerequisite and data quality as the honest self-assessment, because a model trained on inconsistent history reproduces the inconsistency at speed.
Druid AI's guide to agentic deployment in insurance adds the control that has to exist before scope expands: an orchestration layer mediating between specialised agents, memory that captures human overrides and feeds them back, and human approval checkpoints on high-stakes actions such as claim denials or large payouts. In brokerage terms, that is the point where a task stops and a licensed person takes over. Firms that can draw that line precisely can automate more because the boundary makes the automation defensible to a carrier, a client and eventually a court.
The four records a workflow must write before it touches a client file
Every AI step that reads or produces client-facing content should leave a retrievable record of what it read, what it concluded, how confident it was and who accepted it. That makes review a control. A minimal schema, stored with the case in the broker management system:
{
"case_id": "PL-2026-04417",
"workflow": "coverage_comparison",
"model": {"name": "vendor-document-model", "version": "2026-04", "prompt_hash": "9f2c..."},
"sources": [
{"doc_id": "wording_2026_acme.pdf", "page": 10, "clause": "3.4",
"field": "business_interruption_limit", "value": "500000 GBP"},
{"doc_id": "schedule_2025_acme.pdf", "page": 2, "clause": null,
"field": "business_interruption_limit", "value": "250000 GBP"}
],
"output": {"row": 4, "finding": "limit increased", "confidence": 0.91,
"flags": ["wording superseded 2026-01"]},
"review": {"required": true, "reviewer": "j.singh", "role": "account_manager",
"decision": "accepted_with_amendment", "timestamp": "2026-05-11T09:42:00Z"},
"destination": "client_proposal_v3.pdf"
}
Three properties of that record do the work. The source array carries a document identifier, a page and, where it exists, a clause for every field the output relies on, so a reviewer opens the wording. The confidence and flags fields force the tool to express uncertainty instead of resolving it silently, which is precisely the behaviour that separates a usable document workflow from a professional indemnity event. The review block names a person, a role and a timestamp, which is what converts an audit trail into evidence for a file review or a Consumer Duty conversation.
The Agent Bureau frames the same requirement from the governance side: a log showing which workflows touched a customer file, held so that a file review is a record. Whether that record is produced by a platform, a managed service or an internal build is a commercial question. Whether it exists is not.
Three routes into the book, compared on what they can prove
There is no dominant default to copy. Agents by AIQ's 2026 feature comparison notes that the top ten agency software vendors hold only 20.7% of market revenue and the leader, Vertafore, holds 4.63%, while vendors frame platform selection as a five to seven year architecture decision in which a wrong choice can force a replatform within 18 months. Fragmentation at that level makes integration a permanent condition, so the route a firm takes should be chosen for the evidence it can produce.
| Route | Where the intelligence sits | What the audit trail proves | Where a licensed person signs | First failure you are likely to see |
|---|---|---|---|---|
| Point tool bought by a team | Inside the vendor's product, outside the book | That a document was processed | At output review, if a reviewer is assigned at all | Data rekeyed back into the AMS; joins held in one person's head |
| AI-native platform | In the same data model as placement and compliance records | That ingestion, validation and output shared one source | At the placement or advice decision | Scope creep into lines the data model does not cover |
| Managed agent layer | In agents configured around the existing AMS | That a chase, capture or draft step ran, with the AMS still the record | At recommendation and binding, kept out of scope | Boundary drift between administrative chase and advice |
The diligence questions follow the table. Insurance Business reports that Cara's vendor checklist starts with where client and policy information goes and whether it trains models, moves to SOC 2 and HIPAA reports that should be read, and then asks whether the tool can trace a finding back to the source document, flag uncertainty, integrate with the agency management system and enforce access controls across teams. Model choice is a separate decision and belongs per task: Insurance Business's 2026 guide to brokerage models notes that context window and accuracy matter more to brokers than speed, that the leading platforms differ in character, and that the harder question is what the firm is asking a model to do with client data and whether the governance framework is ready for the answer.
The competitive argument has already moved. Zywave's 2026 Broker Services Survey, as Insurance Business reports, found for the first time that failure to adopt modern technology and AI has entered the list of top reasons employers give for switching brokers. That turns architecture into a retention question. An account manager freed from rekeying can carry a larger book, which matters where talent is the binding constraint, but only if the output they review is traceable enough to stand behind with a client. The agencies that get this right treat the data contract, the API surface and the review record as assets that appreciate and get refreshed as models change.
Sources
- The AI gap inside agencies that brokers need to close
- Which AI is right for your insurance brokerage? A practical guide to the models that matter in 2026
- How AI Transforms Insurance Broker Workflows in 2026
- Agentic AI in insurance: A guide for brokers and insurers
- Feature Comparison of Insurance Broker Software Options
- Managed AI for UK Insurance Businesses
- AI's Wake-Up Call: Reinventing Insurance and Wealth Management Business Models
Source: The AI gap inside agencies that brokers need to close, Insurance Business





