“This is an AI that has been trained on all of our investment committee papers over the last 13 years. So, it’s ingested everything – all the deals that we’ve shown the committee. And interestingly, it has the perspective on what happened to the deals that we did, but also the deals that we didn’t do.” – James Brocklebank – Co-chair of Advent International
The central vulnerability in institutional investing is not a lack of information but a failure to remember, interrogate and learn from past decisions at the moment new capital is committed. Files are archived, partners move on, and narratives about past wins and misses harden into anecdotes rather than data. In private equity, where each fund may back only a few dozen companies over a decade, this amnesia is particularly costly: every decision implicitly rests on a tiny sample of lived experience, filtered through selective memory and shifting market regimes. The attempt to encode an entire firm’s investment history into a working artificial intelligence is, at heart, an attempt to eliminate that amnesia and to systematise learning from the full record of what was done – and what was passed over.
Private equity investment committees are designed as collective safeguards against error. Senior partners gather around a table, review thick papers laying out a thesis, risks, scenarios, and proposed structures, and interrogate the deal team until they are satisfied that the risk-reward trade-off justifies action. Formally, this process is meant to be rigorous and dispassionate. In practice, it is constrained by time, cognitive bandwidth and the tacit hierarchies of a partnership. Only a subset of past cases can be recalled in detail, and those that come to mind are often the most spectacular successes or failures, not the quiet, representative middle. An AI that has ingested every investment committee paper over more than a decade intervenes directly in this mechanism: it sits, figuratively, at the same table, but with an eidetic memory of every prior deal.
The factual context is highly specific. Over 13 years, a large global private equity firm will typically generate hundreds of investment committee papers spanning different sectors, geographies and economic cycles. Each paper encapsulates the state of knowledge and conviction at a point in time: management projections, competitive dynamics, due diligence findings, capital structures, exit scenarios and downside cases. Some of these deals were approved and executed, producing a realised track record of internal rates of return, multiples of invested capital, write-offs or restructurings. Others were rejected and never funded, but their subsequent trajectories can still be observed via public information, market data or later approaches. Training an AI on this full corpus turns the investment committee’s hidden archive into a structured dataset of assumptions and outcomes.
What gives this approach its distinctive bite is the claim that the AI holds a perspective on both the deals that were done and those that were rejected. Conventional performance analysis focuses on the realised portfolio, dissecting which transactions created or destroyed value and why. The opportunity cost of deals passed over is usually discussed qualitatively, if at all. With a model that has systematically absorbed the records of rejected deals, it becomes possible to ask: when the committee declined an opportunity, how did that company actually perform in the market? Were there patterns in the kinds of deals that were reflexively dismissed but later succeeded under different owners? Conversely, did the committee correctly avoid a set of structurally flawed situations that later underperformed? The AI’s value lies less in clever pattern recognition and more in reconstructing this counterfactual history that human participants cannot reliably hold in their heads.
From a technical standpoint, building such a system resembles constructing a domain-specific large language model or retrieval-augmented assistant specialised on deal memos. The raw material is unstructured text: multi-page internal papers, perhaps augmented by spreadsheets and slide decks. These need to be cleaned, anonymised where appropriate, and linked to outcome data: entry valuations, revenue and EBITDA paths, leverage levels, realised returns and holding periods. A typical pipeline might convert each paper into text embeddings and index them, enabling the model to retrieve the most relevant precedents when a new memo is submitted. On top of this, supervised or reinforcement learning could be used to align the model’s outputs with the kind of questioning and scepticism the committee wants – highlighting, for example, where current assumptions about margin expansion or multiple arbitrage resemble past cases that disappointed.
Although this configuration rarely requires explicit equations, the conceptual structure is similar to mapping each historical deal i to a feature vector x_i (sector, geography, leverage, growth assumptions, sponsor plan) and outcome y_i (return metrics, qualitative success or failure). The new deal is another point x_{\text{new}}, and the AI’s task is to estimate properties of y_{\text{new}} by learning from the joint distribution of past (x_i,y_i). Unlike a traditional predictive model, however, a language-based system can expose not just a numeric forecast but a narrative: “this assumption resembles deal A, B and C, which underperformed for these reasons”. That narrative interface is precisely what makes the tool usable in a committee that operates through argument and persuasion rather than pure quantitative scoring.
Strategic tension: judgement versus automation
Deploying an AI in this way exposes a powerful tension at the heart of private equity. The business sells itself on human judgement: the ability of partners to see around corners, interpret nuance in management teams, and navigate complex regulatory or technological change. At the same time, investors demand repeatability and process discipline, especially for firms managing tens of billions in commitments. Codifying 13 years of decisions into a model that interrogates new deals is a step towards industrialising judgement – turning what were once informal mental checklists into a systematised set of questions and warnings.
That shift cuts both ways. On the one hand, an AI grounded in the firm’s own history can act as a powerful guardian against overconfidence and fad-chasing. When enthusiasm builds for a hot subsector that has already burned the firm in previous cycles, the model can surface those scars instantly: pointing to prior memos with similar narratives that ended badly. It can also challenge cognitive biases such as confirmation bias, by recalling contrary evidence the team has quietly de-emphasised. On the other hand, there is a risk that partners begin to treat the AI’s prompts as an authoritative view rather than a stimulus for debate. If the tool repeatedly flags risks around duration, leverage or customer concentration, committees might mechanically shy away from any deal that looks remotely similar to a past failure, even when the factual circumstances differ.
This tension is particularly acute because the dataset itself encodes past biases. If, for example, the firm historically underweighted technology investments and overindexed to business services, the AI’s training material will reflect that skew. It may implicitly learn that certain sectors are “not for us” simply because internal memos framed them more sceptically or lacked later outcome data. The model is also only as good as the analysis that went into past papers; if bad assumptions were never corrected in the documentation, the AI may take them at face value. The danger is that, without careful design and governance, the system could become a sophisticated echo chamber – amplifying, rather than interrogating, the partnership’s existing worldview.
Why private equity is a fertile testbed
Private equity’s structural features make it unusually well-suited to this form of AI augmentation. First, the ratio of documentation to decisions is high: a single investment committee paper can run to dozens of pages, capturing granular thinking about a company at a moment in time. That means the training data are rich, domain-specific and already curated around a common template, which improves both model performance and comparability across deals. Second, the feedback horizon, while long, is ultimately discrete: each deal produces a realised return profile and qualitative assessment. Over 13 years, a global firm will have observed multiple full cycles of entry, ownership and exit, providing the model with completed outcome labels rather than perpetually open-ended forecasts.
Third, the environment is competitive and information-sensitive but still human-scale. A firm might review hundreds of potential investments in a year, but only a fraction make it to a full committee, and fewer still are executed. This creates a natural filter: the AI is trained not on every pitch deck that crosses a junior associate’s desk, but on the subset that senior leadership deemed worthy of intensive scrutiny. Finally, the governance structure of an investment committee – with clear accountability, minutes and voting records – makes it feasible to embed a tool that formally “speaks” during deliberations, providing a traceable record of what it raised and how the committee responded.
The specific personality and track record of the individual championing such a system also matters. James Brocklebank has been associated with a strategy that embraces complexity: targeting situations where operational improvement, carve-outs and regulatory nuance create barriers to entry for more formulaic capital1,3,6. In that context, an AI trained on the firm’s own intellectual history is not a substitute for this complexity-driven edge but an amplifier. It can scan across hundreds of prior carve-outs, regulatory approvals or multi-jurisdictional integrations to highlight patterns that a single partner, however experienced, might miss. At the same time, someone whose philosophy treats complexity as an opportunity rather than a threat is more likely to view the AI as a challenging partner – a source of discomfort as well as reassurance – than as a box-ticking device.
Debates and objections
Despite the appeal, there are serious objections to this approach, both technical and philosophical. One objection is that the past 13 years may be a misleading guide to the next decade, especially in a world reshaped by digital platforms, geopolitical risk and monetary regime shifts. Training an AI on this period risks baking in patterns that were contingent on cheap capital, benign inflation and a particular form of globalisation. Critics would argue that by systematising these patterns, the firm could become less, not more, adaptive to regime change. Defenders might respond that the AI’s value lies less in predicting returns mechanically and more in surfacing how the firm reasoned under different conditions – including moments when earlier assumptions broke down. The model can be tuned to highlight where underlying macro regimes differed, prompting committees to ask whether current conditions are analogous or structurally novel.
A second objection concerns explainability and responsibility. If an AI trained on internal memos expresses scepticism about a deal, to what extent should that sway the committee? If the deal later fails and the AI had raised red flags, limited partners might ask why its warnings were not heeded. Conversely, if the AI favoured a deal that underperforms, was there a failure in model governance? The more central the tool becomes, the more it becomes part of the fiduciary chain. That raises questions about validation, monitoring and documentation familiar from algorithmic trading and credit underwriting, but less tested in discretionary private equity. Some observers fear a creeping “automation of blame”, in which partners seek refuge in the model’s outputs to justify conservative decisions, reducing entrepreneurial risk-taking.
Third, there are concerns about confidentiality and security. Investment committee papers contain highly sensitive information: detailed financials, customer lists, trade secrets and personal data about executives. Training an AI on this corpus implies that all of these details now live inside a system that needs to be carefully isolated, audited and controlled. Even if built entirely on-premise or within a tightly ring-fenced environment, questions remain about who can query it, whether their prompts are logged, and how to prevent unintended leakage between deals or to external counterparties. The technical challenge is not just to make the model smart but to embed it within a robust operating framework that satisfies legal, regulatory and reputational constraints.
Finally, there is a cultural objection. Investment partnerships are built around human apprenticeship: junior professionals learn by watching how seniors debate, where they push hardest, and how they navigate ambiguity. Introducing an AI “voice” into that process might skew attention towards what the machine highlights rather than what the most experienced partner finds troubling. Some fear that future generations could grow up over-reliant on prompts generated by the system, losing the hard-won intuition that comes from wrestling personally with messy, incomplete information. Others counter that younger professionals are already fluent in using sophisticated tools and that the real risk is leaving them with inferior, ad hoc aids while pretending that the old ways of doing committee work remain fit for a more complex, data-rich environment.
Why this matters beyond one firm
The broader significance of training an AI on 13 years of investment committee papers extends well beyond a single private equity house. It offers a template for how other long-horizon, high-consequence decision-makers could institutionalise learning. Any organisation that makes episodic, complex decisions – central banks setting policy, boards approving major acquisitions, regulators ruling on systemic cases – generates a similar trail of internal documents. Most of that trail is effectively inert: stored for compliance, occasionally resurfaced during crises or post-mortems, but not actively mined in real time when new decisions are taken. Turning that archive into a live analytical layer could profoundly change how institutions remember and reason.
In finance specifically, the approach draws a sharp line between generic AI assistance and firm-specific “memory machines”. Off-the-shelf models trained on public data can summarise earnings calls or generate draft memos, but they cannot know how a particular partnership thinks about risk, what trade-offs it has historically accepted, or which blind spots have been most costly. By contrast, a model that has ingested all of a firm’s past decisions becomes a reflection of its internal culture, with the power both to reinforce and to challenge that culture. Competitively, firms that succeed in building such tools may enjoy a compound advantage: every new deal and outcome improves the dataset, which in turn sharpens the model’s ability to interrogate the next wave of opportunities.
There is also a subtle shift in the nature of due diligence itself. Traditionally, diligence focuses outward: on the target company, its market and its risks. An AI trained on internal committee papers shifts part of the exercise inward, forcing the sponsor to perform due diligence on its own prior judgement. Instead of asking only “what could go wrong in this company?”, the team must also ask “when we last saw a situation that felt like this, how did we misjudge it?”. This reflexivity is uncomfortable but healthy, especially in an industry where success can breed complacency and narratives of inevitability around past wins.
Over time, such systems could also reshape stakeholder expectations. Limited partners might begin to ask not only for track records and team biographies but for descriptions of how AI-augmented governance works in practice. Questions about model bias, validation and override policies could become part of standard due diligence questionnaires. Regulators might inquire how algorithmic inputs are documented in investment decisions, particularly where pension savers or retail investors are indirectly exposed. Meanwhile, portfolio companies might gain or lose confidence depending on whether they perceive that critical decisions about their future are influenced by a tool trained on past deals in other sectors or regions.
Ultimately, training an AI on the investment committee’s full history forces a confrontation with a deeper question: how much of what a successful private equity firm does can be codified? Some aspects – statistical patterns in leverage tolerance, sector rotation, or sensitivity to macro variables – are clearly amenable to analysis and automation. Others – the chemistry between a deal partner and a founder, the instinct that a management team will rise to a challenge, the significance of subtle shifts in regulatory tone – resist formalisation. The power of the described system lies not in collapsing this distinction but in making it visible. Each time the AI raises a pattern and the committee chooses to override or reinterpret it, the partnership is forced to articulate why this situation is different.
In that sense, the long-run impact of such a tool may be less about the specific deals it nudges towards acceptance or rejection, and more about the discipline it imposes on how a firm explains itself to itself. By preserving a rich, interrogable memory of both executed and declined deals, including their eventual outcomes, the AI becomes a standing invitation to revisit comfortable narratives, to test folk wisdom against evidence, and to re-examine how appetite for risk has shifted over the cycles. For an industry built on the promise of superior judgement, the willingness to subject that judgement to systematic, machine-assisted scrutiny may prove to be the real competitive differentiator.
References
1. James Brocklebank – Advent International – 2026-04-20 – https://www.adventinternational.com/our-team/james-brocklebank/
2. ‘Excited and terrified’: One of private equity’s top investors built an AI … – 2026-05-26 – https://fortune.com/2026/05/26/james-brocklebank-advent-international-goldman-sachs-ic-robot/
3. ‘Complexity is Our Friend’: James Brocklebank on Advent’s Private … – 2026-05-26 – https://www.goldmansachs.com/insights/goldman-sachs-exchanges/complexity-is-our-friend-james-brocklebank-on-advents-private-equity-strategy
4. James Brocklebank on Advent’s Private Equity Strategy – YouTube – 2026-05-26 – https://www.youtube.com/watch?v=bGz3AgFPCs8
5. Advent’s James Brocklebank on the PE opportunities in Europe’s … – 2026-05-26 – https://www.youtube.com/shorts/3mfnWzOXgyc
6. James Brocklebank: Complexity Is Our Friend – Inside Advent’s … – 2026-05-26 – https://finance.biggo.com/news/6cbd95c24b18346f
7. Advent’s James Brocklebank on why complexity can … – YouTube – 2026-05-26 – https://www.youtube.com/shorts/L_qjsbrS4H0
8. James Brocklebank on the Shifting Tectonic Plates of Private Equity – 2025-01-07 – https://www.adventinternational.com/news/james-brocklebank-on-the-shifting-tectonic-plates-of-private-equity/
9. James Brocklebank on Advent’s Private Equity Strategy – Podwise – 2026-05-26 – https://podwise.ai/episodes/8072639
