AI and the Future of Private Markets

Introduction

As at October 2026.

This guide describes how artificial intelligence (AI) enters the analysis of private-market investments. It sets out the mechanisms by which AI changes a business's costs, competition and risks; what has tended to happen when earlier general-purpose technologies spread; how the theme touches diligence in each private asset class; and the questions diligence puts to companies, managers and lenders about it. AI capabilities and the rules that govern them change quickly, so the guide is built on mechanisms and questions, and its few dated facts carry their dates and sources.

What the term covers

In this guide, AI means software systems that learn patterns from data and produce outputs such as predictions, classifications, text, code or decisions. The term spans older machine-learning tools, used for years in credit scoring, fraud detection and quantitative trading, and generative models that produce text, images and code. A further step is systems that carry out sequences of tasks with limited human direction, often called agents.

Canadian securities regulators use a definition close to this one. In a December 2024 staff notice, the Canadian Securities Administrators (CSA) described an AI system as a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions (CSA Staff Notice and Consultation 11-348, December 5, 2024).

For diligence, the useful distinction is between three kinds of business: one that uses AI in its operations, one whose product is AI, and one whose product is work that AI can perform. Each carries a different exposure, and a single company can be all three.

How AI changes a business

AI reaches a company's value through three channels: its costs, its competitive position and its risks.

Costs

The most direct effect is on the cost of work that consists of reading, writing, classifying, summarizing, coding or answering routine questions. Where AI performs part of that work, or lets fewer people do more of it, the labour cost of each unit of output falls.

Three features of these savings matter for analysis. First, they come with new costs: software licences, usage-based fees to model providers, computing costs, integration work, and the staff who supervise and check outputs. A cost line that was mostly salaries becomes partly a vendor bill, and a vendor can reprice its bill. Second, savings arrive unevenly. Fitting AI into existing systems, data and workflows takes time and money, and many first deployments are pilots that leave the cost base unchanged. Third, a saving that every competitor can buy is a saving that competitors can pass on as lower prices. The worked example below shows what that does to profit.

Competition

AI changes competition by changing what is hard to do. Where a business's advantage rests on work that is expensive or slow to perform, such as drafting, analysis, translation, routine software or customer support, AI can lower the barrier to entry and narrow the gap between an incumbent and a new competitor. Where an advantage rests on things AI does not easily reproduce, such as proprietary data, an installed customer base, licences, physical assets, distribution or trust built over years, AI can widen the gap, because the holder of those assets can adopt the same tools.

Pricing models are part of the picture. A business that charges per user, per seat or per hour of work is exposed in a particular way: if its customers need fewer users or fewer hours to do the same work, its revenue can fall even while those customers are well served.

A third effect is dependence. Many companies build their products on models supplied by a small number of large providers. The provider sets the price, the terms of use and the pace at which the underlying capability changes. A feature that distinguishes a product in one year can become a standard capability of the underlying model in the next.

Risks

In the same December 2024 notice, CSA staff gave examples of AI-related risks that reporting issuers should consider and disclose where material. The categories work as a frame for any business, public or private:

Two further risks recur in private-market diligence. Model drift is the deterioration of a model's performance as the data it meets in use moves away from the data it was built on. A credit or pricing model that worked well can become less reliable without any visible failure. Explainability is the degree to which a person can understand why a system produced a given output. Where a decision has to be justified to a regulator, a borrower, an auditor or an investor, a system whose reasoning cannot be traced is a source of risk in itself.

Worked example

A hypothetical services business has revenue of $10.0 million a year, labour costs of $6.0 million and other costs of $2.0 million. Its operating profit is $2.0 million, a margin of 20%.

It adopts an AI tool that costs $0.5 million a year and reduces labour costs by $1.5 million. Total costs become $4.5 million of labour, $2.0 million of other costs and $0.5 million for the tool: $7.0 million. The assumptions are that the tool's effect on costs is the same in every case, sales volumes do not change, and only prices differ.

The cost saving is the same $1.0 million in all three cases with the tool. What differs is who keeps it. If prices hold, the business keeps all of it and profit rises by half. If competitors adopt the same tool and compete on price, the saving passes to customers: a 10% fall in prices leaves profit where it started, and a 15% fall leaves it below where it started.

The last column has a second reading. A competitor that did not adopt the tool would face the same 15% lower prices with its old cost base: revenue of $8.5 million against costs of $8.0 million, a profit of $0.5 million. In that case adoption protected the business rather than enriching it. A plan that counts AI savings as added margin therefore carries an assumption about prices and competitors, and diligence asks what it is.

What has tended to happen when a general-purpose technology spreads

Economists call a technology general-purpose when it is used across most of the economy, improves over time and prompts further inventions built on it. Steam power, electricity, the computer and the internet are the usual examples, and AI is widely discussed as another. Earlier cases describe patterns that have tended to recur, and they suggest what to watch.

Gains have arrived later than the technology

Economic historians who studied electrification found that much of its productivity gain came only once factories were redesigned around electric motors, long after the motors were available. A similar lag was observed with computers. The pattern points to a distinction diligence can draw: between a business that has added AI to an existing process and one that has reorganized the process around it.

Much of the benefit has passed to customers

When a cost-reducing technology is available to every firm in a market, competition has tended to pass a large part of the saving on as lower prices or better products, as in the worked example. Lasting gains for producers have tended to sit with those holding something the technology does not supply: scale, data, distribution, brand or a protected position.

Investment has run ahead of revenue

Earlier technology waves drew heavy investment into the companies most associated with the theme, at valuations that assumed rapid adoption, and into many more companies than the eventual market sustained. A small number of outcomes accounted for much of the value created, and many companies failed. Investors who committed later in a popular theme have tended to pay higher prices for the same kind of exposure.

Labels have followed the money

Strong investor interest in a theme gives issuers a reason to describe themselves in its terms. The CSA notice addressed this directly. It described "AI washing" as inaccurate, false, misleading or embellished claims about the use of AI, made to attract investors, and noted that such claims can appear in marketing materials, offering documents, term sheets and client agreements and may be misleading or amount to a misrepresentation under securities law (CSA Staff Notice and Consultation 11-348, December 5, 2024).

What to watch

Pattern recognition points to signs rather than outcomes. The signs commonly watched include:

• whether AI spending shows up as lower costs or higher output in reported results, or stays in pilot budgets;
• whether savings in an industry appear in margins or in prices;
• which revenue models charge by user, seat or hour in markets where AI reduces the users or hours needed;
• how concentrated a portfolio's or a lender's exposure is to businesses whose product is work that AI performs;
• how dependent companies are on a few model providers, and how those providers' prices and terms change;
• how AI regulation develops in the jurisdictions where a company sells.

A theme of this kind can move faster or slower than these signs suggest, and the same sign can be read in more than one way.

AI inside the manager

Fund managers use AI in their own work: screening potential investments, reading data rooms and contracts, monitoring portfolio companies and borrowers, supporting valuations and drafting investor reports. These uses raise questions of their own.

• Confidentiality. Data rooms, borrower files and investor records are confidential. Where a manager uses an outside provider's model, diligence asks where that information goes, whether it is retained, and whether the provider can use it to train its systems.
• Oversight. A tool that summarizes a contract can miss a clause, and a model that flags borrower stress can fail to flag it. The question is who checks the outputs and who is accountable for a decision that relied on them.
• Valuation. Where models support the valuation of private holdings, the questions are the ones asked of any valuation: what the inputs and methods are, whether they are applied consistently, and who reviews them independently.
• Claims. A manager that presents AI as a source of advantage is making a claim that can be tested. For investment funds that use AI in pursuing their objectives, the CSA notice said staff expected the offering documents to define what the manager means by "AI", to explain how it is used in managing the portfolio, and to avoid vague and unsubstantiated statements.

How AI enters diligence in each asset class

The theme reaches each private asset class through the businesses or assets underneath it. The sections below describe where it enters the analysis for each class.

Private equity

A private-equity fund buys control of businesses and works to increase their value over a holding period of several years. AI enters in two ways. The first is exposure: how the mechanisms above bear on each portfolio company's revenue and costs, particularly companies that sell software priced per user or services priced per hour. The second is the value-creation plan. Managers commonly build operational improvements into their plans, and AI-driven cost savings are now among them. As the worked example shows, such a plan rests on an assumption about prices, and the exit rests on how a buyer would view the company's exposure at the end of the holding period.

Private credit and mortgage investment corporations

A lender is exposed to AI through its borrowers and through its own processes. On the borrower side, the question is whether AI weakens a borrower's revenue over the life of the loan. In July 2026 the Bank for International Settlements estimated that business development companies, a US vehicle for private lending, had lent about US$115 billion to software firms, about a fifth of their lending. It found that revenue uncertainty from generative AI had not yet affected those loans, and noted that narrower credit spreads had reduced the buffers available to absorb losses (BIS Bulletin No. 128, AI disruption in private credit: exposure to software firms in BDCs, July 14, 2026). Concentration of this kind is something a lender's portfolio reports can be read for.

On the process side, lenders use models for credit decisions, pricing and monitoring. When some lenders assess risk with better information, the borrowers they decline or price higher tend to seek out lenders with less information. This is adverse selection, and it applies to a mortgage investment corporation as to any lender. Diligence asks how a lender's underwriting works, what data it relies on, and how its loss experience compares with what its underwriting expected. Where automated tools make decisions about individuals, privacy law applies. Quebec's private-sector privacy law, for example, requires an organization to inform a person when a decision about them is based exclusively on automated processing of their personal information.

Real estate

In real estate, AI enters through operations, valuation and demand for space. Operators use it for tenant communications, maintenance scheduling, energy management and lease administration, which bears on operating costs. Automated valuation tools are used alongside appraisals, which raises the question of what a reported value rests on. Demand for some kinds of space is tied to employment in tasks that AI changes, so diligence on a property asks who the tenants are, what work they do in the space and how long their leases run.

Venture capital

In venture capital, AI is a feature of many new companies rather than a separate category. Diligence on an AI-related company commonly asks what makes it defensible when the underlying models are available to its competitors: proprietary data, distribution, integration into customers' workflows, or a regulated position. It asks how dependent the company is on one model provider, what its computing costs are for each dollar of revenue, and how those costs change as it grows. Venture outcomes are highly dispersed, with a few companies accounting for most of a portfolio's value, and the pattern of investment running ahead of revenue bears on the prices paid.

Infrastructure

Infrastructure assets such as toll roads, airports, ports and water systems are long-lived and often operate under contracts or regulation. Operators use AI for predictive maintenance, monitoring and scheduling. The diligence questions are mainly operational: how models are used in running physical systems, how operational technology is protected from cyber attack, and who is accountable when an automated system fails.

Hedge funds and liquid alternatives

Quantitative strategies have used machine learning for many years. With analytical tools, an edge has tended to erode as the tools spread. When many managers use similar data and models, their positions can converge, and crowded positions can unwind together. Diligence asks what is distinctive in a manager's data or methods, how models are monitored for drift, when human judgment overrides them, and whether the manager's description of its use of AI is specific enough to test.

The Canadian context

Securities law already reaches claims about AI: the CSA notice set out how existing disclosure and misrepresentation rules, which it described as generally technology-neutral, apply to the use of AI. Privacy law reaches automated decisions about individuals, as in Quebec's rule above. And Canada's domestic market is small relative to the markets many Canadian companies sell into, so a Canadian company's AI products can fall under other jurisdictions' rules, including the European Union's AI Act, whose obligations phase in over several years.

Diligence on AI: what the questions test

Diligence on the theme uses the same approach as any other diligence: verification against documents, numbers and references. The questions test five things about a company, and the same five apply, one level up, to a manager's portfolio or a lender's loan book.

• Use. Where AI is used today and what it has changed in costs, output or revenue. A use that shows up in the numbers differs from one that appears only in a presentation.
• Exposure. Which revenue lines consist of work AI can perform, and how the company charges for them.
• Position. What holds the company's advantage in place if competitors adopt the same tools, and whether it holds the rights to the data it relies on.
• Dependence. Which outside models it relies on, on what terms, and what a change in a provider's price or product would do to its costs.
• Governance. Who oversees its AI systems, how errors are found and corrected, and what incidents have occurred.

For a specific investment, the questions also test the base case: whether it holds if AI capabilities advance faster or slower than expected, and whether a plan's savings survive if prices fall. For a manager, they add the manager's own use of AI, set out above, and how its investment theses account for AI over a fund's life, commonly about ten years. For a lender, they add the lender's own models: how they were built, how they are validated and monitored for drift, and when people override them.

How the questions are received is part of the evidence. A specific, documented answer is one kind of response; jargon in place of detail is another.

Questions to ask

• Where does the company use AI today, and what has it changed in reported costs, output or revenue?
• Which revenue lines consist of work that AI can perform, and how are they priced: by user, seat, hour, transaction or outcome?
• What protects the company's position if competitors adopt the same tools, and does it hold the rights to the data it relies on?
• Which outside AI providers does the business depend on, on what terms, and what would a change in their prices or products do to its costs?
• Who is accountable for the company's AI systems, how are errors found and corrected, and what incidents have there been?
• How do the manager's investment theses account for AI over the life of the fund, and how much of the portfolio sits in businesses whose product is work AI performs?
• How does the manager use AI in screening, diligence, monitoring, valuation and reporting, and where does confidential information go when it does?
• How much of a value-creation plan's margin gain comes from AI-driven savings, and what does it assume about prices?
• For a lender, what share of the loan book is to borrowers whose revenue AI could weaken, and how are its own credit models validated and monitored for drift?
• What exactly does an offering document mean by "AI", and can each claim about its use be substantiated?

What remains uncertain

As at October 2026, several things that matter to the theme are unsettled. The pace at which AI capabilities improve has been hard to anticipate. Adoption depends on organizational, regulatory and social factors as much as on what the technology can do. How the gains are shared between companies, workers and customers is unresolved, and it bears on demand across the economy. Regulation differs by jurisdiction and continues to change.

Key terms

• Due diligence: the investigation of an investment, its manager and its documents before a commitment is made.
• General-purpose technology: a technology used across most of the economy that improves over time and prompts further inventions built on it.
• AI washing: inaccurate, misleading or embellished claims about the use of AI, made to attract investors.
• Model drift: the decline in a model's performance as the data it meets in use moves away from the data it was built on.
• Explainability: the degree to which a person can understand why an AI system produced a given output.
• Adverse selection: the tendency of borrowers declined or priced higher by better-informed lenders to seek out less-informed ones.
• Business development company: a US investment vehicle, listed or unlisted, that lends to and invests in private companies.

Related reading

• Venture Capital
• Private Credit
• Hedge Funds & Liquid Alternatives