The first artificial intelligence use case in a private equity fund is chosen on five criteria. They are how often the work happens, where the freed time goes, data access, the value of a reviewed first version and the build cost. The right candidate is a repeated, document-heavy queue of work with a named business owner. The investment decision is never part of it.
The question comes up everywhere, and rarely with a method. Bain & Company measured it in its Global Private Equity Report 2025. Only about 20 % of portfolio companies had put a generative AI use case into operation, with concrete results.
Tandem works with funds and their portfolio companies on this choice. This article gives the grid we apply, the ideas we drop straight away, and what has to be written before building. It complements our article on AI at fund and portfolio level, which covers the order in which to open the two levels.
Start from a queue of work rather than a tool
A good first workflow is described by a queue of work, never by a tool category. A finance copilot or a support agent does not say which work disappears.
Four elements must be known before keeping an idea. You need the trigger (an incoming email, a file received, a deadline), the sources used, the deliverable a person reviews, and the person who decides what happens next. If the team cannot name them, the workflow is not ready. Plugging a model into a fuzzy process only produces, faster, a result nobody knows how to check.
To get these answers, Tandem interviews the people who do the work every day. Those most comfortable with AI make good relays. They are rarely the ones re-keying purchase orders or hunting for information across three shared folders.
Who owns the first AI use case, at fund and portfolio level
The owner changes with the ground. Inside the fund, it is usually IT or a person named as AI sponsor. In a portfolio company, it is more often the operating partner, or the person at the fund in charge of portfolio companies.
Tandem starts with the fund as often as with a portfolio company. We currently work with funds such as Newfund, InfraVia, Kartesia and Épopée Gestion, and with companies in their portfolios.
One point comes back in almost every engagement. A department head struggles to automate their own scope, since it can mean shrinking their team. The first workflow therefore has two owners, a sponsor who arbitrates and a business owner who signs off deliverables day to day.
Five questions to score an AI use case in private equity
Tandem's grid combines our experience of the same workflows across sectors with the company's own context. An idea that pays off in a growing company can be worthless in a company preparing its exit.
- Is the work frequent enough? A rare decision, however important, makes a poor first case. Fifteen sales proposals a month are handled with a well-configured assistant, and automation is kept for volume.
- Does the freed time have somewhere to go? A salesperson who gets back a third of their time calls more prospects if the pipeline follows. A support team that is already well sized will not handle more tickets.
- Is the data accessible, and which source is authoritative? If two sources disagree with no named arbiter, the agent flags the gap instead of inventing a reconciliation.
- Is the first version useful without being final? The right first deliverable is a draft reviewed by a person, such as a completed file or a pre-filled order. AI prepares most of the work, the person keeps the decision.
- What does the build cost? The score takes into account technical feasibility, the number of systems to connect and the duration. Going live takes weeks, with real work between business teams and the technical team.
Each criterion is scored from 1 to 4. The scores group into two axes, value and feasibility, and the priority score is their product, out of 16.
| Criterion | What is looked at | Axis |
|---|---|---|
| Impact | Time saved, risk reduced, revenue secured | Value |
| Reach | Number of people and teams involved | Value |
| Return on investment | Annual gain compared with build and running costs | Value |
| Complexity | Number of systems, edge cases, build duration | Feasibility |
| Data readiness | Accessible sources, stable formats, identified owner | Feasibility |
The full approach, from process mapping to the order of rollout, is described in our AI audit method.
What the grid produced at an industrial portfolio company
This autumn, Tandem audited a portfolio company of a venture capital fund, a manufacturer with three plants in France, the United States and China. Around thirty interviews and about 150 questionnaire answers brought up more than 150 ideas.
The first sort removed the ideas a script or an existing tool could solve, and those out of scope. About 110 remained, grouped into 15 families and then scored. Three projects came out on top.
- Reading incoming documents, starting with customer purchase orders. Several hundred arrive each month as PDFs, and each one takes around twenty minutes of re-keying into the ERP.
- Sales steering, with a daily alert on sales forecasts that have no matching order.
- A multilingual troubleshooting assistant for operators, connected to the manuals and the maintenance history.

The chief operating officer then corrected the grid. A project scored too low moved up at their request, the manual entry of product sheets by R&D, with close to a hundred attributes per product. That bottleneck was delaying customer projects. The grid structures the discussion, the executive has the final word on impact.
The same executive found our return on investment estimates too cautious. An underestimated ROI makes a committee doubt instead of reassuring it. Tandem now presents gross figures, as ranges, with the assumptions in an appendix.
Ideas to drop before any scoring
Some ideas leave the list before being scored. The audit of another portfolio company, a European energy distributor owned by a private equity fund, gave several examples.
- A tool is being rolled out on the same scope. The upcoming transport management system was going to cover delivery notes, so automating their reading would only have served for a few months.
- The ERP changes within the year. Invoicing stayed a priority, and its workflow was designed for the new ERP.
- The volume does not justify the effort. One of the countries had only a dozen customers.
- The team involved does not cooperate. Without data or a contact person, a project does not start, whatever its score.
- The process changes every quarter. A company still experimenting has no interest in freezing a workflow, and a well-configured assistant serves it better.
- The need is common to every company. If a market product already covers it, buying it and doing the integration work is the better call.
Before promising anything, Tandem also tests feasibility on a sample. On that engagement, a seven-day forecast of tank levels reached 80 to 90 % accuracy with simple moving averages. Beyond seven days, it fell back towards 50 %. That test set the scope of the first version.
At fund level, the first workflows that keep coming back
A fund has no call centre and no tickets, but it receives a lot of inbound requests by email. Decks, deal flow files and investor questions form regular queues of work. The first workflows that pass the grid look alike from one fund to the next.
- Pre-screening a deal. At Newfund, Tandem built six agents on Dust, one per analysis dimension (technology, product, team, market, exit). They prepare the material for the investment memo, and the investor keeps the thesis.
- Side letters and recurring clauses. At Kartesia, a single Claude skill covers the side letter clause libraries of around ten funds. It indicates which fund a deal can enter given investor restrictions, and it adds a new side letter when one arrives.
- Sourcing. At Épopée Gestion, our work focused on making the existing sourcing workflows reliable.
- Reporting and monitoring. AI consolidates what comes up from portfolio companies, prepares investor communications and retrieves the history of a company held for five years.
The investment decision never goes first, nor does anything that commits the fund without review. On tooling, our article on Claude Skills for finance details what the skills published by Anthropic already cover.
The first workflow rarely stands alone. At a European private equity fund, dozens of use cases were narrowed to a short list ranked by impact and feasibility. More than 100 staff were trained, and around fifteen agents and skills run in production.
Measure the gain in a unit an executive can check
The baseline is recorded before building, in volumes, time and errors. Once the workflow is live, it cannot be rebuilt.
Two adjustments were asked of us during final presentations, and Tandem now applies them everywhere. The gain is expressed in full-time equivalents rather than hours or euros, because an executive checks “0.6 FTE” by multiplying it by a loaded cost. Gains are also sorted by category, between real cash, avoided costs, freed time and one-off gains such as lower working capital needs.
Gains vary with the job and the complexity of the task. On some highly repetitive flows, the team involved gets back more than three quarters of the time it spent on them. On others, the time gain stays small and the project is justified by quality or avoided risk.
The one-page brief to write before building
When a candidate comes out on top, Tandem writes a one-page brief before the first line of code. It holds six sections.
- The scope, with one queue of work, one trigger and one business owner.
- The inputs, meaning the systems, fields and access rights, as well as what the agent must never touch.
- The output, a structured deliverable with its sources and mandatory fields.
- The rights, meaning what the agent does alone and what stays reserved for a person.
- The exceptions, when a document is missing or two sources disagree. The agent stops, says why and hands over to whoever can decide.
- The measurement, with the baseline, the gain indicators and a quality indicator such as the correction rate.
The inputs are best thought of as a new hire's first days, with their methods, their access and their documentation. I summarised these three building blocks in my skills, MCP and RAG infographic in October 2026.

The first version then runs alongside the current process, on normal cases and on difficult ones.
Widen the agent's rights one step at a time
A first version never covers the whole use case, and this has to be said before signing. Otherwise the client believes the workflow will be complete within a few days. Tandem plans iteration days for each use case and shows them in the quote.
The agent's rights then widen in stages. It prepares a file, then routes simple cases, then updates a reversible internal status. Each stage has its indicator and a dashboard with execution logs. OWASP ranks excessive agency among the main risks of applications built on language models.
Monitoring also catches simple mistakes. On one of our engagements, an old version of a skill, still published under the same name, was answering instead of the new one during testing.
Which AI use case should a private equity fund start with?
Start with a frequent, document-heavy, reviewed queue of work, carried by a sponsor and a business owner. At fund level, deal pre-screening or side letters almost always pass the grid. At portfolio level, reading incoming documents often comes out on top.
Tandem recommends a full audit of the fund or the portfolio company. It gives a coherent rollout order and the baseline figures. When that order already exists, what remains is to choose the first version, decide how to approach it with the teams, and write it on one page.



