AI Underwriting:
Where It Actually Changes. From Insurance Forms to Investment Decisions

AI underwriting is the use of machine learning, and increasingly agentic AI, to evaluate risk inputs (documents, structured data, third-party signals) in order to support or automate an accept, decline or price decision. That is the working definition, and it holds whether the thing being underwritten is an insurance policy, a consumer loan or a private transaction.

The core is risk evaluation at machine scale. A traditional underwriter reads a file, weighs the evidence against a set of rules and a body of experience, and reaches a judgement. AI underwriting compresses the reading and the first-pass evaluation: a model ingests the inputs, extracts what matters, and produces a structured assessment a person can act on. In its lighter form it augments the underwriter; in its more automated form it clears routine cases and routes the hard ones to a human. What it does not do is remove judgement: the decision to accept risk, and on what terms, still rests with a person and a governed set of controls.

In practice, AI underwriting follows a consistent three-stage arc.

The first is document triage. Underwriting runs on paperwork (applications, statements, contracts, reports), and the first job is to read and organise it. Models extract the relevant fields from unstructured documents, flag what is missing, and turn a pile of files into structured inputs. This is where most of the early time saving comes from, replacing hours of manual reading with a first pass measured in minutes.

The second is risk evaluation. With the inputs structured, the model assesses them against the relevant risk factors: scoring, pattern detection, comparison against historical cases, and the surfacing of anomalies a tired reviewer might miss. The output is an assessment, not a verdict: a ranked and evidenced starting point for the decision.

The third, and the newer one, is decision support and agentic automation. Where the first wave of these tools answered questions, the current wave acts: agentic AI executes multi-step tasks across systems, pulling data, cross-checking it and drafting the assessment, with limited supervision. For low-risk, high-volume cases that can mean an automated decision within a governed boundary; for everything else it means a far more complete brief landing on a human's desk. Across industries the sequence holds: triage, evaluate, decide.

Read almost any current explainer on AI underwriting and the setting is the same: insurance, mortgages, consumer credit. The authors are insurtech and lendtech vendors describing their products, insurers announcing tools, and consultancies framing the transformation for those markets. It is a coherent, well-covered conversation, and for those industries exactly the right one.

But it stops there. The entire public discussion treats underwriting as the scoring of a policy or a loan, a high-volume and relatively standardised decision about a known type of risk. That is fair as far as it goes: it is where AI underwriting was deployed first, because it is where the volume and the structured data already live.

What the framing leaves out is the more interesting part. Triage the documents, evaluate the risk, support the decision: that sequence is not confined to policies and loans. It describes any process where evidence has to be assembled and weighed before capital is committed, and there is one large domain where the description fits precisely, yet where the conversation has barely started.

Underwriting, at its root, is not an insurance word. It is the act of deciding whether something deserves capital, and on what terms. Reframed that way, it is the central question of investment itself, and AI underwriting becomes far more than a faster claims desk: it becomes the machine-assisted assessment of whether an asset or a deal merits an allocation.

In investment and private-markets diligence, the arc maps almost one to one. Deal sourcing replaces application intake: AI screens a far wider field of off-market opportunities than a deal team could canvass by hand. Diligence replaces document triage, except the documents are a data room and the reading runs to weeks. Evaluation replaces risk scoring, against the questions a private transaction turns on. This is the territory of ai in private credit underwriting, where the decision is whether a borrower and a structure justify the capital and the terms, and where the quality and speed of that assessment is the whole competitive question.

Allocator-side credit underwriting, fund and manager underwriting, deal-level diligence: none of it appears in the insurance-and-lending literature, yet all of it is underwriting in the original sense. It connects directly to adjacent private-markets work, from private credit as an asset class to tokenised private credit, where what an on-chain instrument represents must be verified before it can be underwritten at all.

Here the distinction that matters is between AI as a feature and AI as infrastructure. Most of what is marketed as AI underwriting is a capable model bolted onto an existing workflow, still a tool a person reaches for; the harder version rebuilds the process around the system, so the assessment is produced by the machine and signed off by the human.

This is where IMS Group, a technology-forward private markets investment group and partnership of family offices, has built rather than bought. Through its in-house lab, IMS Labs, the Group runs Cortex: a 41-agent underwriting engine (an "Agent Mesh") that serves as the private-credit intelligence layer beneath its diligence. On the public figure IMS Labs states, Cortex compresses a two-week diligence cycle to roughly 90 minutes end-to-end. The point is not the headline time but what it implies: underwriting has been re-engineered as an operating system, not added as a chatbot.

Cortex is one of three platforms IMS Labs runs in production, alongside Onyx (the patent-pending authentication and truth layer for real-world assets, with digital twins and tokenised ownership) and Nimbus (the agentic commerce layer). For the underwriting question, Cortex is the relevant one: it is the operationalisation proof the vendor conversation asserts but rarely shows, set out by IMS Labs (imslabs.ai) and examined further alongside agentic AI as an operating system.

The question the public conversation keeps returning to is whether AI is going to replace the underwriter. Posed as a yes-or-no it produces a defensive answer. The more useful framing treats this as a reallocation of judgement rather than its removal: more of the work shifts to the model, while the decision itself stays with a person.

What AI absorbs is the triage and the first-pass evaluation: the reading, the extraction, the routine scoring that consumes most of an underwriter's hours and least of their skill. What it does not absorb is the genuinely hard decision: the ambiguous case, the structure with no clean precedent, the judgement call where the evidence points two ways. As the machine takes the volume, human attention moves up the difficulty curve toward exactly those decisions; the role does not disappear, it concentrates on the part that always carried the skill. With only a small minority of firms having scaled an AI agent into live production in any function, the near-term reality is augmentation in most places and full automation in few.

The risk sections in most AI underwriting explainers read as a compliance checklist: model bias, data quality, the EU AI Act, the need for explainability. The items are right, but the framing undersells them. For an allocator, explainability and auditability are not a box to tick; they are the condition under which the output can be trusted at all.

An underwriting decision that cannot be explained cannot be governed, and a decision that cannot be governed cannot carry capital. That is why a credible operator can show where a person reviews and overrules the model, how its reasoning is documented, and how the assessment would stand up to an audit. Cortex reflects this in its design: the agent mesh assembles the investment-committee decision package, and a human investment committee reviews and signs it off, so the machine produces the package and a person retains the decision. Investment-grade governance is not an add-on to AI underwriting; it is the evidence that the capability is real rather than asserted.

The underwriting shift is not an insurance footnote. The insurance-and-lending coverage describes the first place AI underwriting was deployed, not the limit of where it applies. Read in its original sense, as deciding whether something deserves capital, underwriting is how investment decisions are made, and that is where the machine-assisted version is now arriving. To see how an operator approaches it in practice, read across the AI-in-investing cluster, from the agentic shift to how the firm runs its own lab.

This article is provided by IMS Group for information purposes only and does not constitute investment advice, an offer, or a solicitation. Figures are sourced and dated where cited; forecasts are projections and stated as ranges.

This article is provided by IMS Group for information purposes only. It does not constitute investment advice, an offer, or a solicitation. Figures are point-in-time and projections are estimates.