From potential to performance: Using gen AI to conduct outside-in diligence

Before they can commit to major investments—under tight timelines, with incomplete data, and amid rising expectations of value creation—leaders must conduct a rigorous outside-in diligence process, with strong benchmarking, experienced judgment, sharp analysis, and pattern recognition to model the upside.

The outside-in diligence process used to require weeks of manual effort from a diligence team—sourcing public data, mining the seller’s data room, scraping external signals, triangulating expert input, and stitching together all those insights. Now, gen AI is changing the equation. Gen AI tools can take the first pass—synthesizing vast amounts of public and proprietary data, identifying trends and outliers, and even proposing hypotheses that analysts might not have considered. Some tools even provide preliminary insights on value to be unlocked within the asset.

The result: faster insight generation, broader scope, and sharper strategic clarity.

Yet, for all its potential, many organizations have only just started using gen AI to conduct diligence and other critical business processes. The technology is still new. Implementation is uneven. And very few have cracked the operating model to consistently create impact from their gen AI deployment.

Our experience supporting a range of gen-AI-enabled diligences, from public company transformations to portfolio company transformations and turnarounds to pre-investment value creation assessments, points to five common challenges organizations face—from underleveraged proprietary data to inconsistent prompting structures—and just as many opportunities to raise the bar. In this article, we discuss how to successfully incorporate gen AI into a diligence team’s outside-in analysis. These insights can help leaders move faster, think more deeply, and make better decisions—and the lessons hold true, whether a team is diagnosing an opportunity to transform or pressure testing a target’s growth thesis.

Five ways to improve diligence using gen AI

By deploying gen AI to collect and curate inputs, among other tasks, analysts can focus their time and effort on steering the analysis and sharpening the implications. To unlock gen AI’s full potential, however, diligence teams will need to do more than simply plug in the tool. Like any expert system, these technologies perform best when properly trained, precisely prompted, and paired with experienced judgment. Specifically, teams should aim to strengthen their gen AI capabilities in the following five areas.

1. Customize models using proprietary data

Gen AI models are only as good as the data they are trained on. While off-the-shelf models can provide useful answers to general questions, the most powerful applications emerge when companies can train them on their own proprietary information and experienced judgment. Diligence teams can develop a competitive advantage by systematically capturing and curating their own data sets. However, not all data is created equal. As a first step, leaders must pinpoint the unique data that underpins their organization’s competitive edge to date, including customer-level transaction histories, synergy realization rates from prior M&A activity, and throughput logs for plants. By training gen AI on the organization’s institutional knowledge, diligence teams can shorten the time required to extract valuable insights from various data sets and quantify opportunities in a more accurate and reliable manner.

Consider a software-as-a-service (SaaS) company that was assessing an acquisition. It used a gen AI model that was trained on proprietary customer and sales data and, in doing so, spotted users who weren’t taking full advantage of features they already had access to. Based on past patterns, the gen AI model predicted which of these customers were likely to upgrade or spend more with the acquired company—helping the SaaS business identify and quantify revenue opportunities that other bidders overlooked.

Similarly, when evaluating the scale of a potential transformation for a global oil and gas player, a McKinsey diligence team was able to draw on its anonymized data set of more than 1,600 enterprise transformations and track performance across more than 500,000 initiatives. In this way, the team was able to identify top performance levers, reasonable sizing estimates, and implementation timelines quickly—compressing what would typically take weeks of manual analysis into days.

While proprietary data is one of the most powerful enablers of gen AI’s impact, it is also among the hardest to activate. Companies often struggle with fragmented, unstructured data and unclear pathways to make this data usable. High-value sources—such as diligence archives, integration playbooks, and pricing benchmarks—must be cleaned, tagged, and secured for retrieval before they can be integrated into an outside-in analysis. After companies have identified their proprietary data, they should determine the best path to integrate the data within their diligence processes—leveraging, where needed, the suites of tools (gen AI and otherwise) available in the public domain.

2. Optimize peer set and benchmark selection

Peer comparison is at the heart of most outside-in operational analyses, but it is often more art than science. Diligence teams must balance the need to select a peer set and benchmarks that are focused enough to preserve insights but broad enough to capture emerging and potentially disruptive competitors.

Sophisticated diligence teams are now deploying gen AI to scan industry databases, earnings transcripts, and even patent filings to construct dynamic peer sets based on product overlap, similarities in cost structure, or go-to-market models. Beyond suggesting names of relevant companies, these tools also offer deeper analysis into why these peers are worth reviewing—based on, for instance, customer segments, product mix, and supply chain structure. These insights and additional levels of detail paired with experienced judgment can yield breakthrough value improvements and opportunity generation within a given portfolio asset.

One company, in analyzing a medtech acquisition target, uncovered a set of comparable Asian market players that traditional screening had missed. The company leveraged a gen AI agent to scan global databases, regulatory filings, and local-language press for signals such as overlapping product portfolios, approval pathways, and shared distributors—details that are often buried outside standard industry codes. The agent was able to uncover peers that revealed far greater margin potential and market share growth for the medtech company than initially expected.

Gen AI tools can also help diligence teams rapidly iterate across multiple peer sets, testing how benchmarks vary depending on the inclusion or exclusion of certain players across different geographies. By pairing this agility with a clear set of rules, teams can calibrate their conclusions, communicate a range of outcomes with greater confidence, and make peer selection more science than art. This starts with reviewing prior peer set and benchmarking decisions, locking in the selection criteria, and embedding them as custom instructions for agents or prompts.

3. Construct prompts like a product manager, not a search bar

We frequently observe diligence teams and other users of gen AI making the mistake of treating it like a smarter search engine—firing off short, unstructured prompts and reverting to manual practices when an abysmal result is invariably produced. In practice, the quality of gen AI’s output is directly tied to the quality of the prompt it receives. Effective prompts clearly specify the question, the data sources to be used, the constraints to observe, and the priority of various hypotheses. They also anticipate follow-ups, allowing the model to stay within a cohesive analytical frame.

Diligence teams that invest time up front to craft thoughtful, structured prompts get answers more effectively tailored to aspects of the process—and refine those prompts further through iteration after testing.

For example, a team evaluating the efficiency of a company’s customer acquisition costs (CAC) didn’t just ask its GPT, “How efficient is this company’s CAC?” Instead, they framed the question as: “Compare the company’s CAC to the median of peers with a B2B SaaS model, with less than $500 million in annual recurring revenue, with multichannel go-to-market, using public financials and investor presentations from the past 12 months.” This added context gave the generic GPT the precision needed to tailor its output—producing a benchmark table, qualitative comparisons, and industry-specific insights. This made the analysis more relevant, reliable, and grounded in the company’s specific context.

Context and role give the gen AI models a clear lens and sharpened focus for the diligence task. When teams specify who the model is acting as (role) and what constraints and background apply (context), the gen AI tool can adopt the right mindset, draw on relevant knowledge, and produce actionable results that match the diligence team’s objectives. For example, an analyst evaluating a company for a potential private equity (PE) investment retrieved a nuanced, insightful perspective once it shifted its prompt from “Give me an analysis of this company” to a more comprehensive request: “Act as a due diligence analyst evaluating a target company for a potential PE investment. Focus your review on its EBITDA margins, working capital efficiency, and capital expenditures over the past three years, using publicly available financial statements and analyst commentary. Highlight any anomalies or trends that could affect its EBITDA quality and cash flow sustainability.”

To get started, the team should focus on what a “good” prompt looks like and formalize it. They can start by building a shared prompt library with clear role definitions, context parameters, and examples tailored to common diligence tasks. This mutual understanding will make the prompts sharper, the output more consistent, and help teams speed up the analysis.

4. Build specialized agents for specialized tasks

Leading diligence teams are starting to develop specialized gen AI agents for specific tasks, often integrating them into cohesive, end-to-end workflows that can enhance the overall diligence process. Such agents perform best when focused on a single domain with the right contextual training; they aren’t generic “answer bots” but purpose-built teammates with clearly defined roles, inputs, and constraints.

Again, diligence teams will need to invest time up front—this time to map out each specific agent’s responsibilities, data sources, target users, and desired outputs, and then create a framework that keeps the agent focused and able to collaborate with other agents as needed. Such an approach will make it easier for diligence teams to manage validation outputs from gen AI agents and reduce hallucinations.

One application for leveraging gen AI agents was seen in a leading diligence team’s approach to peer selection. To reach its goal of identifying comparable peers for a niche company, the team built a specific peer selector agent that sifted through hundreds of pages of filings, investor presentations, and market commentary. This agent’s output was then passed along to downstream agents to produce a comprehensive investment thesis of a company—compiled in a matter of days, not weeks.

To get started, diligence teams must take stock of their repeatable processes—even those with multiple steps or handoffs—and pinpoint where one or more gen AI agents could streamline, accelerate, or elevate the work.

5. Treat gen AI as an amplifier, not a decision-maker

One of the biggest risks in using gen AI for diligence is mistaking fluency for accuracy. The technology produces confident, well-articulated outputs, but that polish can mask serious flaws if the underlying data is poor, misaligned, or incomplete. We have seen diligence efforts where ungoverned gen AI tools generated peer sets that ignored business model nuances, surfaced cost estimates disconnected from operational realities, or hallucinated metrics from misinterpreted text.

To avoid these issues, some diligence teams are supporting strong prompts with strong oversight—treating gen AI not as a decision-maker, but as an amplifier of both insight and error. They are requiring human oversight of gen AI models in higher-risk areas, logging and auditing gen AI models’ behaviors, and isolating certain environments—through private cloud deployments or firewalled systems—to protect sensitive data and preserve client confidentiality.

Our experience working on thousands of transformations points to the importance of embedding structured checks into diligence workflows. Doing so can reveal common pitfalls—for example, in one recent case, a systematic check conducted by a gen AI tool caught overstated synergies during the assessment of a potential transformation.

Such a governance layer is fast becoming a best practice—not just for risk mitigation but also to build trust in the results that gen AI delivers. Training teams on AI’s limitations is a critical first step in establishing this governance layer—followed by a clear, organization-wide mandate to implement risk-based oversight before any gen AI tool goes live.

How to get started

The use of gen AI in outside-in analysis holds great promise, and to fully realize this potential, diligence teams can begin integrating gen AI into their process through five practical steps:

• Inventory and prepare your proprietary data. Identify the data sets that give your organization a competitive edge—past deal outcomes, synergy models, pricing benchmarks—and clean, tag, and secure them so they can train custom gen AI agents.
• Codify and test your best prompts. Build a reusable library of structured, best-practice prompts for common diligence questions. Specify role, context, constraints, and data sources, and refine them through testing to create a consistent foundation for analysis.
• Pilot targeted, high-impact agents. Inventory your set of repeatable processes, and identify those that bring the highest value yet currently require the highest effort to implement. Start with two or three task-specific agents—such as competitor scanning, market sizing, or synergy sizing—and integrate them into existing workflows. Keep scope tight to maximize accuracy and learning.
• Appoint an AI champion and shape the model. Designate a leader to coordinate across diligence, data, and risk teams; steward gen AI adoption; and evolve methods over time. Along the way, make practical calls on where to build your own capabilities versus tapping into proven external tools.
• Establish a disciplined feedback loop. Regularly review agent performance, retrain with fresh data, and adjust prompts or workflows to reduce errors and improve relevance—building both accuracy and trust in outputs.

Taken together, these practices suggest a broader shift: Gen AI is not just a new tool—it requires a new operating model to get the most out of it. In this model, the core diligence team plays the role of orchestrator, continuously designing, refining, and integrating gen AI agents into the analysis workflow. Data engineers ensure that relevant data sets—both public and proprietary—are curated and updated. Analysts craft and iterate prompts like product specs. Knowledge management teams help capture what works so it can be reused on future deals. Risk teams set guardrails that keep gen AI safe, ethical, and compliant.

This new model is fast, scalable, and adaptive. It reduces manual work, expedites some analyses, and shifts the focal point from human involvement to applying judgment, structuring the problem, and orchestrating the work.

Gen AI is poised to replace much of the manual lifting involved in completing outside-in diligence. Firms that gain the most during this transition will be those that adapt the fastest—building institutional know-how, training models on proprietary data, and reimagining the analyst’s role as a gen AI orchestrator. And the payoff will be faster diligence, deeper insight, greater agility, and more confident decision-making.