Agentic AI in Finance: The Chatbot Era Is Over and the Operating System Has Arrived

Agentic AI in finance is autonomous software that plans, executes and validates multi-step financial tasks, working toward a goal across tools and data rather than returning a single response to a single prompt. The defining feature is the loop: an agent sets out the steps a task requires, carries them out, checks its output against the objective and revises, all without a person driving each move. That is what separates an agent from a model that predicts the next token on request.

The architecture is usually described as multi-agent, or an agent mesh: not one large model doing everything, but specialised agents that each own a part of the task and hand work between them. One reads the source documents; another extracts the relevant facts; another reasons over them; a senior agent reviews and can override. The pattern matters because financial work is rarely a single step. Pricing a private credit position, screening a deal or reconciling a close is a sequence of dependent judgements, and a system that runs the whole sequence is a different proposition from one that helps with a single link.

This is why the term agentic AI in financial services now appears in the language allocators use: it names the shift from software that assists to software that executes.

The cleanest way to understand agentic AI is against the thing it is replacing. The agentic AI vs generative AI distinction is not a matter of degree; it is a difference in what the software is for. Generative AI, and the chatbots and copilots built on it, is assistance. You ask, it answers: a drafted memo, a summarised filing, a reply to a question. It stops when the response is delivered. Most firms have captured some of that value already, but it waits to be asked and does not act on what it produces.

Agentic AI is execution. Given an objective rather than a prompt, it works through the multi-step path to reach it, calling tools, reading data, taking action and validating the result as it goes. Robotic process automation, the older technology it is sometimes confused with, can also act, but only along a fixed, pre-scripted route that breaks the moment reality departs from the script. An agentic system reasons about the path instead, which lets it handle the variation real financial work throws at it.

This is why "operating system" is the right metaphor, and the agentic AI vs generative AI gap is the gap between a feature and a system. A chatbot is a window bolted onto an existing process; an agentic layer coordinates the work. When research, screening, diligence and monitoring all run through agents that plan and execute, AI stops being a tool the firm reaches for and becomes part of how the firm operates.

Read the coverage of agentic AI in finance and it is mostly about finance functions: the autonomous accounting close, KYC and compliance, banking operations, fraud detection. Those are not where a private markets investment group competes. The agentic AI use cases that matter to an allocator sit in the investment process itself. Three stand out, and a private-markets line runs through each.

Agentic AI for deal sourcing

Deal sourcing is the first and the most overlooked. Origination rarely gets a mention, which leaves ai deal sourcing close to uncontested ground. In private markets the edge is proprietary: seeing opportunities other allocators never see. Agentic systems widen that aperture, screening a far larger field of off-market situations than a deal team could canvass by hand and surfacing the few worth a partner's attention. The agent does not decide what to pursue; it does the finding and filtering that would otherwise cap how many opportunities a team can assess.

Agentic AI in diligence and underwriting

Diligence and underwriting are the second, and the most consequential. A private transaction carries weeks of document review where a public trade carries a screen and a price, and that work is exactly what an agent mesh is built to compress. Agents read the data room, extract the facts that bear on the decision, reason over them and assemble the case for an underwriter to judge, turning research into the inputs for an underwriting decision rather than a pile of unread PDFs. This is the heart of the buy-side wedge, treated in full in where AI changes underwriting. The compression is real, but the decision stays human: the system produces the package; a person prices the risk.

Portfolio monitoring is the third. Private holdings do not reprice on a public tape, so the discipline is to read the signals that do exist, in operating data, covenants and counterparties, continuously rather than at the next quarterly valuation. Agents suit that always-on watch, flagging drift and exception as they appear instead of at period-end. In each case the capability that elsewhere gets pointed at the back office is pointed instead at investing.

Autonomy raises the stakes on governance, and the right register is an allocator's risk discipline, not a compliance vendor's checklist. The first concern is the one the field names most often: data. Agentic systems are only as good as the information they act on, and most financial data is fragmented and uneven in quality. An agent acting confidently on bad data is a worse outcome than a slow analyst, so getting data clean and connected is a precondition, not a detail.

The second is oversight. An autonomous system that takes multi-step action needs a human who reviews, validates and retains the authority to overrule it, designed in from the start and not bolted on afterwards. In an agent mesh this usually takes the shape of a senior reviewer sitting above the working agents, keeping the consequential judgement with a person. The third is auditability and provenance: being able to trace what an agent did, on what evidence and why. Provenance is also where the agentic layer meets the infrastructure beneath it, because a verified record of what an asset is becomes an input the agents can rely on. That problem, authenticating the real-world asset under a token, is taken up in the truth layer that authenticates tokenised assets. None of this is a brake on agentic AI; it is the condition for using it on real capital.

So far this has stayed advisory: what firms could do, framed by vendors and consultants. The more useful vantage point is an operator's, the one IMS Group offers through IMS Labs, the Group's in-house engineering arm. The facts that follow are public, from the live IMS Labs site.

Cortex is the clearest illustration. On the public record, IMS Labs describes Cortex as a 41-agent underwriting engine built on its Agent Mesh, the private-credit intelligence layer that turns research into underwriting decisions. The headline the site puts on it is "2 Weeks to 90 Minutes": end-to-end underwriting time compressed from a two-week cycle to roughly ninety minutes. That figure illustrates what an agentic operating layer does to the shape of the work, not a performance guarantee, and the underwriting judgement stays with a person. Cortex is one of three platforms IMS Labs runs in production, alongside Onyx, the patent-pending truth layer that gives real-world assets digital twins and verifiable provenance, and Nimbus, an agentic-commerce ecosystem. The posture worth borrowing is the one that ties them together: every platform runs inside the Group before it goes anywhere else. What earns the claim is that it runs in production. The detail sits with IMS Labs and its public site, imslabs.ai.

The direction is clear; the magnitude is contested. Industry research has repeatedly found that only a small minority of firms have scaled an AI agent into genuine production in any function, even as the ambition has become near-universal, which puts the agentic shift early on its curve rather than finished. Over the next cycle the dividing line in investing is unlikely to be whether a firm uses AI, a question already settled, but whether it operates an agentic layer that runs. Estimates of the gains vary widely and should be read as ranges, not promises; what is not in dispute is the shape of the move, from assistance to execution and from a feature to an operating system.

For a private markets investment group, that move is a capability to build, because the places it changes the work, sourcing, diligence, underwriting and monitoring, are where private-markets returns are made. This page is one part of the Group's wider reading of AI in investing; to see the thesis built and run, not just described, explore IMS Labs.