Private equity has absorbed a significant amount of AI vendor attention in the past two years. Most of it has been poorly calibrated — vendors projecting capabilities onto use cases without understanding the regulatory environment, or building tools that work in demonstration and fail in production on deal documents.
This is a grounded view. Where AI creates genuine operational leverage in PE, where the constraints make cloud tools unworkable, and where AI does not change anything meaningful.
Where AI creates genuine leverage
Due diligence. This is the clearest application, and the one with the most direct time-cost impact. A mid-market fund running ten to fifteen deals a year spends a substantial fraction of analyst capacity reading data rooms — financial statements, customer contracts, employment agreements, regulatory filings. The reading is not the work. The analysis is. A system that can answer compound questions across an entire data room — "what is the customer concentration in the top five accounts, and has it been disclosed consistently across all representations in the data room?" — eliminates most of the reading time and replaces it with analysis time.
The constraint is immediate and hard: data room documents contain MNPI. This is not a grey area. The documents cannot leave the firm's environment. The due diligence AI use case is entirely dependent on on-premise deployment.
Portfolio monitoring. Portfolio companies produce reporting packages, board decks, and management accounts. A fund with a large portfolio reviews a significant volume of this material on an ongoing basis. AI systems can track metrics across the portfolio, flag variances against plan, and surface the companies and metrics that require attention. This is a monitoring and triage function — the model does not decide what to do about a variance, but it can ensure the variance is not missed in a large portfolio.
Portfolio company documents also contain sensitive financial information, though the MNPI constraint is less acute than in the deal process. The appropriate architecture depends on the sensitivity of the specific materials.
LP reporting. Generating LP reports is time-intensive and largely formulaic — the structure is consistent across quarters, the data sources are defined, and the variation is in the specific numbers and commentary. AI can draft LP reports from structured data inputs, reducing the time cost of report generation substantially. This application is typically appropriate for cloud tools, as the input data is controlled and the output is a document the fund is producing, not client data being transmitted.
Where it doesn't change anything
Deal sourcing. AI can screen a large universe of potential targets against defined criteria — revenue range, geography, sector, financial profile. That is a data retrieval and filtering task, and it is appropriate for AI. The judgment about which targets are actually interesting — which businesses have defensible market positions, which management teams are credible, which industries are entering a cycle that creates acquisition opportunity — is not. Deal sourcing quality is determined by investment thesis quality. AI does not generate investment theses.
Investment judgment. The decision about whether to invest in a company at a given price with a given capital structure involves judgment about business quality, market dynamics, competitive position, and management capability that does not reduce to a query on a database. AI tools that claim to support investment decisions are doing something different — they may surface relevant information, flag risks found in documents, or identify comparable transactions. That is useful. It is not the investment decision.
LP relationships. The relationship between a GP and its LPs is built on trust, track record, and personal connection developed over years and across market cycles. AI does not create or maintain that relationship. It can help prepare materials for LP meetings, organise communication history, and ensure no LP is neglected in periodic outreach — all useful operational functions. The relationship itself is human.
The MNPI constraint, specifically
This deserves direct treatment because it is the constraint that makes most available AI tools unsuitable for the highest-value PE use cases.
MNPI — material non-public information — is generated throughout the deal process. Target company financials, management projections, deal terms, data room documents, and internal deal analyses are all MNPI from the moment of acquisition through the time the information becomes public. Controls around MNPI exist because trading on it is a securities violation, and the controls require that MNPI not be transmitted to parties who should not have it.
An external AI model is a third party. The enterprise agreement between the fund and the AI provider does not reclassify the provider as an internal party. When deal documents are processed by a cloud model — even through an enterprise API with zero data retention — they have been transmitted outside the firm's controlled environment. That is a controls failure.
The solution is not to avoid AI for due diligence. The solution is to run the model inside the environment where the MNPI lives. An on-premise system with no external API calls processes MNPI on the fund's own hardware. The controls are maintained because the data has not moved.
This is the architecture that makes PE-specific AI viable. It is also why most general-purpose AI tools are not the right starting point for PE firms — they are built for use cases where the data can leave the organisation, and PE due diligence is a use case where it cannot.