AI value creation in portfolio companies: from fund strategy to measurable EBITDA impact

Why AI programmes stall in portfolio companies

We keep seeing the same pattern.

The investment thesis names AI as a value lever. The portfolio company starts a pilot. Twelve months later there is a dashboard, a proof of concept, and not one line in the reporting pack that moved.

The technology is rarely the reason.

The reason is that nobody described the path from pilot KPI to the P&L. A chatbot answering tickets does not reduce cost while service shift planning stays unchanged. A scoring model does not raise revenue while sales still prioritises by instinct.

The second reason is ownership. AI lands with IT because it sounds technical. The value lever sits in a business function, and that function never asked for the pilot.

The third reason is time. A four to five year hold period does not tolerate programmes that think in quarters. What produced nothing in 100 days usually produces nothing in 300 either.

Two layers: the fund itself and the portfolio

AI value creation in private equity touches two separate systems. They get mixed together routinely, and that costs clarity.

Layer one: the fund as its own operation

A fund is an information-processing business with a small team. That makes it the cheapest place to build real experience, because the data is manageable and the users sit next door.

Three applications carry the most weight early.

Sourcing and screening. Market monitoring, signal extraction from public sources, pre-sorting targets against fund criteria. A model makes no investment decision. It reduces the share of the team’s week spent on pre-qualification.

Due diligence support. Making data rooms searchable, extracting contract clauses, flagging gaps between the management presentation and the numbers. The benefit is completeness under time pressure, not a replacement for judgement.

Portfolio reporting. Collecting monthly metrics from heterogeneous systems, normalising them, and drafting commentary. Many funds pay for this today in analyst hours and spreadsheets.

Layer two: the portfolio

This is where the value actually is. And here the rule is simple. A fund does not have one AI programme. It has as many as it has holdings. Each has its own systems, its own data quality, its own people.

A shared platform across every portfolio company sounds like scale and usually breaks on heterogeneity. What scales is the method, not the software.

From fund theme to company project

Between the thesis in the investment committee and a running system in the portfolio company sits a translation that almost nobody makes explicit.

The thesis is written at fund level and reads like a category. Sales automation. Back office efficiency. AI as a margin lever.

A portfolio company cannot implement a category. It can implement a process, with an owner, a budget, and a date.

Making that translation is the job of an AI operating partner, whether or not the role carries that title. It comes down to three decisions.

Who carries the outcome? Not who runs the project, but in whose targets the metric sits. If the lever is in sales, the initiative belongs to the sales lead. IT provides access and operations, not the thesis.

What gets dropped? A mid-sized portfolio company can carry one or two initiatives of this kind at once. A list of nine use cases is not a prioritisation. It is a deferred decision.

When do we stop? The stop condition belongs in the same decision as the start. Without it, initiatives continue because they were once approved.

A narrow role at fund level still helps. A shared method, a shared set of assessment questions, and one place where the fund records what worked in which holding. What does not help is mandating tools centrally.

Value creation levers by function

AI value creation reduces to three categories: lower cost, higher price or revenue, defended multiple. It only becomes concrete function by function.

Sales and go-to-market. This is the lever with the shortest distance to the result. Research, enrichment, first-touch drafting, follow-up, CRM hygiene. A worked example for scale: if a team of eight sales reps spends half its day on research and documentation, recovering a third of that is roughly one full-time person per week returned to selling. Whether that holds in a given company has to be measured, not assumed.

Back office. Invoice processing, purchase order matching, master data, expenses. High document throughput and stable formats are the precondition. The effect shows up in cycle time and error rate, often before it shows up in headcount.

Pricing. Discount discipline, differentiation by customer segment, detecting margin leakage in the installed base. Pricing levers act fast because they need no volume. They do need clean transaction data and support from the CEO.

Customer service. Deflection of standard requests, suggested replies for agents, automatic case summaries. The value lands in capacity planning, not in the bot.

Operations and supply chain. Demand forecasting, maintenance, visual quality inspection. Longer lead times, but harder for a competitor to copy.

Where LLM agents fit

Agents make sense when a task spans several steps across several systems, occurs often, and has a checkable result. Quote generation, supplier requests, moving data between CRM and ERP.

They do not make sense when an error is expensive and hard to spot, or when a rule solves the task more cheaply.

We build the evaluation harness before the rollout. Without test cases and an error budget, an agent in production is a bet.

A 100-day playbook

What follows is not a project plan. It is an order of operations that stops effort from arriving before clarity.

Day 1 to 20: assessment. An AI readiness assessment in the portfolio company answers four questions. Which processes consume the most time? Which data exists, and at what quality? Which systems can be connected? Who inside the company can own an outcome?

Day 20 to 30: prioritisation. Every candidate is scored on two axes. EBITDA impact and data readiness. High impact and ready goes first. High impact and unready gets a data project, not an AI project. The rest waits.

Day 30 to 60: pilot. One use case, real data, real users, production systems. No sandbox. The pilot has a baseline measured beforehand and a stop condition.

Day 60 to 90: scale. The pilot rolls out to the whole team, or it is shut down. Both are acceptable outcomes. A pilot that keeps running without scaling is the most expensive ending.

From day 90: governance. Access rights, logging, handling of personal data, accountability for errors, documentation for the eventual buyer’s diligence. Setting up governance early is cheaper than retrofitting it during an exit process.

Measurement: from pilot KPI to EBITDA

The bridge between a technical metric and a financial one has to exist before the pilot, not after it.

A simple scheme works. The pilot measures an operational quantity, say cases handled per staff hour. That quantity links to an economic one, say labour cost per case. That in turn touches a P&L line.

The third link is the one that matters. Time saved becomes cost saved only when a decision realises it. Either the capacity is redirected into revenue, or it is removed, or it absorbs growth without new hires. If that decision is never made, the effect stays invisible.

For the exit story, repeatability counts as well. A buyer values a documented, running process differently from a one-off saving.

Three measurement mistakes we see often

Baseline taken after launch. If the starting point is only recorded after rollout, the effect can no longer be evidenced. The baseline belongs in the week before the pilot.

Usage treated as success. Number of users and number of requests describe adoption, not value. Both rise even when the output is unusable.

Double counting across initiatives. Two projects that save the same hour save it once. With several parallel initiatives in one function, the sum of the individual benefits is almost always too high.

What the reporting pack needs

One line per initiative is enough. Status, baseline, current measurement, linked P&L line, next decision point. If that line cannot be filled in, either the measurement or the value lever is missing.

What a fund needs from an engineering partner

We are a small senior team and we work accordingly.

Few experienced people, not a pyramid. In a 200-person portfolio company a large consulting team is a foreign body. Two or three people who build the thing get further.

Weeks, not quarters. An assessment takes us weeks. A first production use case follows in the same quarter.

Deployment, not decks. We deliver running systems inside existing environments, connected to CRM, ERP, and the data that is already there, with tests and a handover to the internal team. A slide deck is a by-product.

Handover as the goal. The portfolio company should be able to continue without us. That serves the fund too, because dependency on an outside vendor shows up in a buyer’s diligence.

Closing

AI in a portfolio company is not a technology topic you can delegate. It is a value creation topic with a technical implementation.

The route there is unglamorous. Understand where the time and the margin sit. Prioritise by impact and data readiness. Get one use case into production. Measure it. Then take the next one.

If you have a portfolio company where AI is in the thesis and missing from the reporting, write to us at [email protected]. We will look at the case and tell you whether we see a lever.