McKinsey's Four-Subsector AI Map Exposes a Delegated Authority Audit Gap

McKinsey's four-subsector AI analysis splits impact by mechanism: broker AI shortens the submission-to-quote cycle, MGA AI increasingly decides which risks bind and at what price, software-provider AI reshapes the rating engines underneath all three, and TPA AI touches the claims data actuaries rely on for reserving. US direct premium placed through MGAs grew roughly 14% annually, from $47 billion in 2020 to $97 billion in 2024 (McKinsey, February 2026), a pace carrier oversight capacity has not matched.

Tracking delegated authority filings and MGA capacity reports across 2026, the recurring blind spot is structural rather than technological: carrier actuarial teams built their audit playbooks around reviewing a human underwriter's judgment, a file note, a referral rationale, a manager's sign-off. None of that playbook transfers cleanly to reverse-engineering an MGA's black-box risk selection model once loss ratios start drifting from what the binding agreement assumed. McKinsey's report is the first to name that four-way split explicitly across brokers, MGAs, software vendors and TPAs; what it does not do is tell a carrier actuary what to actually ask for in the next program review.

Four Subsectors, Four Different Actuarial Exposures

McKinsey's framing treats brokers, MGAs, software providers and TPAs as facing structurally different AI transitions, and each maps to a distinct point in the actuarial workflow rather than a single generic "AI in insurance" risk. Brokers have already used AI to automate submissions, match carrier appetite faster, and surface cross-sell opportunities; agentic tools are moving toward handling simple renewals with limited human review (McKinsey, February 2026). That shortens the pipeline feeding an underwriter, but it does not itself change what risk gets bound, so the actuarial exposure is closer to a volume and mix question than a selection-quality one.

MGAs sit at the opposite end of that spectrum. McKinsey's view is unambiguous: "MGAs that consolidate their proprietary data and generate clear customer insights will be valuable partners to carriers and brokers" (McKinsey, February 2026), which is another way of saying the MGAs pulling ahead are the ones whose AI models are making the actual bind-or-decline call using data the capacity-providing carrier does not itself hold. That is a direct risk-selection exposure, not an efficiency one, and it lands squarely on the appointed actuary who has to opine on reserves for a book they did not underwrite.

Software providers are the layer underneath all three other subsectors, and McKinsey frames the shift there as a move away from monolithic AI systems toward a modular approach where carriers plug purpose-built models into core systems without a full re-platforming (McKinsey, February 2026). That matters to actuaries less directly, but it is the plumbing determining whether an MGA's or TPA's model output can even be traced back to a specific rate filing or reserve assumption once carriers stop running one vendor's black box end to end. TPAs, finally, have the deepest transaction-level claims data of any of the four subsectors and the clearest tension: efficient AI claims processing directly touches the loss data reserving actuaries treat as ground truth, at the same time as the TPA's own commercial incentive to invest in that efficiency is weak, a point developed below.

MGA Premium Growth Is Outrunning Carrier Oversight Capacity

McKinsey's $47 billion to $97 billion trajectory, a roughly 14% compound annual growth rate from 2020 through 2024 (McKinsey, February 2026), is not the only recent count of how fast delegated authority is scaling, and the separately reported figures tell a consistent story from a different angle. AM Best's 2025 market segment report put total US direct premium sourced from managing general agents and other delegated underwriting authority enterprises at $108.7 billion, up 17.8% from $92.3 billion in 2024 and more than triple the roughly 5% growth rate of the broader US P&C industry (AM Best, June 30, 2026). Two independent counts, one investor-facing and one rating-agency-facing, both show the delegated channel compounding at double-digit rates for at least five consecutive years while carrier oversight infrastructure was largely designed for a slower, more concentrated distribution model.

14%

Annual growth rate in US direct premium placed through MGAs, 2020 to 2024 (McKinsey).

$108.7B

2025 delegated authority direct premium reported by AM Best, up 17.8% year over year.

75%+

Share of 2025 delegated authority agreements where the MGA holds underwriting authority, not just binding or claims authority (AM Best).

The capacity side has grown to match, but not necessarily the oversight side. More than three-quarters of delegated authority agreements in 2025 gave the MGA underwriting authority itself, a higher share than the prior year (AM Best, June 30, 2026), meaning an increasing majority of the growth McKinsey is describing sits behind agreements where the party making the AI-assisted risk selection decision is, by contract, not the party bearing the ultimate underwriting loss. AM Best's own report is titled "Managing General Agents Adapt to Changing Demands and Added Scrutiny," language that reads less like a celebration of the growth McKinsey is describing and more like a warning that capacity providers are starting to price in the oversight gap. When a carrier's own actuarial staff cannot independently reconstruct why a specific risk was accepted, the industry-standard response, tighter loss ratio caps, shorter contract terms, more aggressive profit-commission clawbacks, treats a symptom the McKinsey report identifies as a subsector-wide data and technology asymmetry, not a program-specific underwriting failure.

What a Carrier Actuarial Team Needs to Audit an AI-Driven MGA

A carrier actuarial team auditing a human underwriter has a well-worn toolkit: file notes, referral memos, a underwriting manual the underwriter is expected to follow, and a manager's sign-off trail for anything outside authority limits. None of that transfers directly to an MGA whose bind decisions run through a proprietary model, and building the equivalent toolkit for an AI-driven program means specifying, in the binding agreement itself, three categories of data the capacity provider has historically not required.

The first is loss ratio segmentation by underwriting channel. A capacity provider that cannot distinguish AI-assisted bind decisions from manually reviewed exceptions within an MGA's book has no way to isolate whether emerging adverse development traces to the model's risk selection logic or to the ordinary variance in a specialty book. That segmentation requires the MGA to tag every bound policy at the point of bind, not retroactively reconstruct the tag from underwriting notes months later when a loss ratio review flags a problem.

The second is a model change-control log tied to the underlying rate filing. When an MGA retrains or materially adjusts the model driving its risk selection, the capacity provider's actuarial team needs to know the date of that change and be able to test whether business bound after the change is performing differently from business bound before it, in the same way a pricing actuary tracks a rating plan revision against subsequent loss experience. Without a change-control log, a capacity provider reviewing a year of loss data has no way to know whether it is looking at one underwriting process or three sequential versions of one, silently stitched together in the aggregate.

The third, and the one McKinsey's TPA analysis makes most urgent, is claims data reconciliation when the entity servicing losses is itself running AI on the loss data the actuary depends on. If a TPA's AI is triaging, coding, or auto-adjudicating claims on an MGA-bound book, the reserving actuary is now two AI layers removed from the underlying loss event: one model decided what risk to accept, a second model is shaping how that risk's losses get recorded. A capacity provider's reconciliation process needs to sample claims across both layers, not just audit the MGA's underwriting file, to have confidence the loss triangle it is reserving against reflects the actual claims experience rather than an AI-mediated summary of it.

McKinsey subsector	Primary AI mechanism	Direct actuarial exposure
Brokers	Submission automation, appetite matching, agentic renewals	Volume and mix shift feeding underwriting, not selection quality
MGAs	AI-assisted or AI-driven bind and quote decisions	Risk selection logic outside carrier's own model validation
Software providers	Modular rating engines replacing monolithic platforms	Traceability of model output back to filed rates and reserve assumptions
TPAs	AI-assisted claims triage, coding, and adjudication	Integrity of the loss data feeding reserve triangles

The TPA Revenue Model Problem, and Why It Matters for Expense Assumptions

McKinsey's TPA analysis surfaces a structural tension actuaries building expense assumptions need to price into their fee schedules directly. Most TPA commercial arrangements still run on head count, activity-based constructs, or cost-plus economics, explicitly or implicitly, which means that when AI genuinely reduces the labor required to process a claim, the TPA's own revenue falls under its existing contract even as its service quality improves (McKinsey, February 2026). McKinsey's own language on the point is direct: it is "not yet clear how TPAs will reliably monetize AI-driven efficiency gains," and under head-count or cost-plus models, "automation can actually pressure top-line revenue even when performance improves" (McKinsey, February 2026).

That paradox has a concrete pricing consequence. An actuary building expense assumptions for a TPA-serviced book on the old cost-plus logic, where claims-handling expense scales roughly with headcount and transaction volume, will overstate future servicing costs on a book the TPA is quietly automating, and will simultaneously understate the risk that the TPA underinvests in the AI tooling that would have improved claims-data quality, because that investment does not pay for itself under its own fee structure. McKinsey expects the next phase of the TPA subsector to be defined "less by whether TPAs adopt AI (they will) and more by how they evolve their pricing models and competitive positioning" (McKinsey, February 2026). Carriers renegotiating TPA contracts around outcome-based or per-claim fee structures, rather than headcount-based ones, are effectively realigning the TPA's commercial incentive with the claims-data integrity the reserving actuary needs, which makes the fee-structure clause as much an actuarial control as a procurement one.

Software Providers Set the Ceiling on What Actuaries Can Even Validate

The software-provider layer McKinsey describes, carriers moving from monolithic AI systems toward modular architectures built on open standards, determines whether an actuary auditing an MGA or TPA's AI output can trace that output back to a specific model version and a specific rate filing at all. A monolithic, vendor-locked system that bundles submission intake, risk scoring, and bind logic into one opaque pipeline gives a carrier's actuarial team almost nothing to validate beyond the final decision; a modular architecture, where the risk-scoring component is a separable module with its own version history and its own documented training data lineage, is what makes the change-control logging described above possible in the first place. Carriers evaluating whether to build this validation capability in-house or acquire it are facing the same build-versus-buy calculus playing out elsewhere in the underwriting stack, most visibly in Duck Creek's acquisition of Send Technology's orchestration engine, which absorbed a platform already routing $26 billion in gross written premium into triage and quote-decision logic just ten weeks after its own launch.

Data-layer investments elsewhere in the industry point the same direction. EXL's acquisition of iMerit's data-annotation and training-layer business is a bet that the quality of the labeled data feeding an insurance AI model, not just the model architecture itself, is where competitive and audit value will concentrate, which is precisely the layer a carrier actuary needs visibility into before it can trust an MGA's or TPA's model output. A modular, well-documented data layer is what lets a validation actuary ask "what specific dataset trained this risk score" and get an answerable response instead of a vendor's marketing description of the pipeline.

Why This Matters for Actuaries

The practical task for a carrier actuarial team overseeing AI-driven delegated authority programs is building the audit checklist McKinsey's subsector analysis implies but does not itself provide: loss ratio bands segmented by AI-assisted versus manually underwritten origin, a model change-control log cross-referenced against the binding agreement's rate structure, and a claims-data reconciliation process that samples across both the MGA's underwriting AI and the TPA's claims AI rather than treating either as a black box to be trusted on the strength of the loss ratio it reports. None of that is new actuarial theory; it is the same segmentation, validation, and reconciliation discipline reserving and pricing actuaries already apply to their own books, extended to a delegated relationship where the underlying data and model no longer sit inside the carrier's own systems.

That gap is also where the near-term consulting opportunity sits. Carriers newly reliant on MGA and TPA AI outputs they cannot independently validate are increasingly commissioning targeted reserving reviews and pricing audits specifically to close it, work that requires an actuary comfortable both with traditional loss development analysis and with asking a vendor pointed questions about training data lineage and model change history. As MGA premium keeps compounding toward and past $100 billion and TPA contracts get renegotiated around AI-driven efficiency, the carriers that build this audit capability before a loss ratio drifts, rather than after, will be the ones setting the terms of the next renewal rather than reacting to it.