When an investment fund decides to take on artificial intelligence, it almost always makes the same choice, and makes it without noticing. It starts with itself. Investment teams get trained, memos come faster, models land in an afternoon. That is the right first step. It is also the smaller of the two.
Tandem works with funds and their portfolio companies on AI applied to operations. This article describes the two levels of the subject, what each one actually returns, and the order in which to open them. It is not about finance in general, it is about private equity.
The two levels of AI in a fund
The first level is the fund. Analysts, portfolio directors, operating partners. AI speeds up the work of finding, assessing and monitoring deals. The second level is the companies owned. There, AI changes how the business actually runs.
Confusing the two is the most common mistake, and an expensive one. A partner processing board decks faster gains hours. A portfolio company whose teams build their own tools and rewrite their own processes is a different business by the time you sell it. Both are worth doing. Only one shows up in the exit price.
Where does private equity actually stand on AI?
The French market remains solid and selective. In 2025, private equity players invested 36.4 billion euros across 2,904 companies and infrastructure projects, raised 42.9 billion euros, up 10 %, and exited 14.0 billion euros at historical cost (France Invest and Grant Thornton, study published in March 2026). The number of transactions is falling. In a market where exits are negotiated harder, what justifies a multiple becomes the central question.
On AI, the profession starts from behind. The French financial markets authority study published in February 2026 observes that adoption is mostly led by traditional asset management, and that managers specialised in private equity are only beginning to look at it. Across asset managers generally, the AI budget stays under 1 % of IT spending.
On the portfolio side, the number that matters is this one. About 20 % of portfolio companies have operationalised a generative AI use case and see concrete results. The figure comes from Bain & Company's Global Private Equity Report 2025, based on a survey of institutional investors run in September 2024. A majority of the others are still testing.
In other words, four portfolio companies out of five have nothing measured yet. For a fund holding twenty companies, that is sixteen open files and an advantage still available.
What AI prepares at each step of the investment cycle
At the deal desk, gains are immediate and easy to feel, which explains why everyone starts there. AI reads a data room, drafts an investment note, summarises a management call, builds a first version of a model and stress-tests assumptions. What took an analyst a week fits into an afternoon.
The most interesting gain at fund level is not speed, it is reach. A fund screening a few hundred files a year and discarding most of them can point AI at the top of its funnel. Seeing more targets at constant cost, when the binding constraint is capital rather than deal flow, has a direct effect on selection.
The Swedish fund EQT has held that logic for a long time. Its internal tool, Motherbrain, launched in 2016 and now covers the full cycle, from sourcing to due diligence and value creation (tech.eu, November 2025). Alexander Fred-Ojala, Head of AI at EQT Ventures, sums up the purpose of these in-house tools in one line, to give the firm an edge and boost productivity. Ten years of a head start on this subject is not a detail.
The investment memo is not a prompting problem
Many teams get stuck in the same place. They write longer and longer instructions to get a memo in the right format, and the output stays generic. The problem is not how the request is phrased. It is the context the team makes available.
The real lever is to place your investment thesis, your selection criteria, your past memos and your document templates, once and for all, in a space the tool can read. From then on the request fits in two lines. I detailed that shift in my newsletter edition on Claude Cowork, and the site keeps a full version of it in our practical Cowork guide.
The level above is called a skill, and it is the format that makes a fund genuinely efficient. A skill describes the house procedure for one type of deliverable, the committee memo for instance, with its structure, its mandatory sections and its prohibitions. Written once, it applies to every file. That is the subject of our article on the end of the prompt as the unit of work, and of my post on the four levels.
On a data room, the perimeter matters more than the model
This is the question that always comes first at a fund, and it is legitimate. You handle confidential memoranda, models, shareholder agreements and portfolio company data. Handing those documents to a tool is not something you improvise.
The right answer is not to pick the most reassuring model, it is to define the perimeter. An agent reading documents has rights over a specific folder, and those rights are decided before the first session. Three folders are enough, and two of them are read-only.
- What is decided before the first session. The professional plan chosen, the hosting location, and the vendor's written commitment not to reuse your data for training.
- What is decided file by file. Which documents enter the perimeter, who puts them there, and when they leave. A closed data room is a data room removed from the workspace.
- What is never delegated. The legal qualification of flagged clauses. The tool surfaces departures from the standard, the lawyer decides. That boundary is the same one described in our article on AI in finance.
Why the real pocket of value sits in the portfolio companies
Speeding up the deal desk is worth doing. It does not move the multiple. The value that appears at exit is created inside the companies owned, and the funds pulling ahead know it.
The contrast is stark once you name it. Take a portfolio company whose process owners build their own tools. They automate their own work and ship without waiting on an IT queue. That business no longer resembles the one bought three years earlier. That is a portfolio-wide capability, not a head-office convenience.
The organisational models observed in the market vary, and that is instructive. Bain & Company's 2025 report describes three ways of mobilising a portfolio, at Vista Equity Partners, Apollo Global Management and Hg. What they share is not the tool chosen. It is that the fund pushes a common capability rather than one isolated project per company.
Optimise before you automate, the discipline that separates value from theatre
There is a failure mode worth naming before any rollout. A team gets excited, picks a process, and automates it as is, mess included. Automating a broken process only makes the mess run faster.
The discipline is to fix first. Ask whether the step still serves anything, cut what no longer does, simplify what remains, and only then automate. That work is unglamorous. It makes the difference between a portfolio company that genuinely runs leaner and one accumulating brittle automations nobody dares touch.
This is also why judgement matters more than tooling. Knowing what to automate, and above all what to leave alone, is the skill that transfers. We detail the four levels of automation and when each becomes relevant in our article on workflows, assistants and agents, and the general failure mode in our review of companies that see no return on their AI.
How fast a deployment pays for itself in a portfolio company
On our assignments in funds and their portfolio companies, payback on the initial investment usually lands under six months, and sometimes under two. I gave that order of magnitude in a conversation for the La Reprise podcast in spring 2026. It applies to well-chosen augmentation projects, not to an information system overhaul.
| Before deployment | After deployment | |
|---|---|---|
| Volume | Items processed in a typical month | Items passed through the flow |
| Time | Average time per item, actually timed | Residual time on reworked cases |
| Quality | Error rate observed today | Rework rate after control |
| Cost | Loaded hourly cost of the function | Run cost, inference and supervision |
The left-hand column is the one almost nobody fills in, and the only one that makes the gain demonstrable. A week of sampling is enough. After deployment that starting point cannot be reconstructed, and the committee debate becomes a matter of opinion.
What AI will not do in an investor's place
An honest article on the subject also says where the subject stops. Four things do not move, and none is solved by a better model.
- Investment conviction. It is the product the fund sells to its subscribers. A well-made summary does not replace it, it frees time to build it.
- The relationship with management teams. A management meeting does not prepare itself and cannot be delegated. A good share of won deals is decided there.
- Negotiation. A tool can list departures from market standard and propose alternative drafting. It does not hold a balance of power in a room.
- Responsibility. It stays entirely with the fund and the authorised person, whatever tool prepared the document.
A fifth limit is harder to accept. There is no off-the-shelf transformation manual a fund could buy and roll out across its portfolio. That is not a reason to wait, it is a reason to start. The teams that get good do so by practising on real files. The others look for a method that is not coming.
Where to start when you run a fund and twenty portfolio companies
Three moves, in this order, and none requires a blanket mandate. First train the investment team on real work, diligence, memos, modelling. It is fast, low-risk, and it gives you the standing for what follows.
Then pick one willing portfolio company, and point the same capability at its operations. A management team asking for the subject is worth infinitely more than one you have to drag. The first case must be a success you can show to the other nineteen.
Finally treat the subject as a capability you are building, not a project you are finishing. Keep what works as reusable procedures, train people rather than tools, and expand from the wins. The full approach sits on our AI agency for investment funds page, and the scoping method in our AI audit method.
Two articles complete this one without repeating it. For the broad sector frame, the state of adoption in France and the applicable regulatory calendar, see our article on AI in finance. For the advisory side, from teaser to signing, the one devoted to AI in a mergers and acquisitions mandate.
One last rule, valid everywhere. The winning fund is not the one reading its deals fastest, it is the one whose portfolio companies sell better. Everything else is preparation.
