Three levers actually move EBITDA through AI

Most AI initiatives start out of curiosity, not out of a P&L line. Someone tried a language model, built a demo, got budget for a pilot. That's not a criticism. That's how most technologies enter a company. The problem shows up later, when the pilot has to go into the next budget cycle and nobody can say which cost center or revenue line it actually moves.

A pilot with no assigned P&L impact competes in the budget process against every other initiative that can show a number. It almost always loses. The result is an organization that has launched many AI experiments and scaled few of them. The reason wasn't that the technology didn't work. Nobody defined beforehand which lever it was supposed to move.

That order can be reversed, and doing so is exactly what separates companies where AI shows up in the results from those where it doesn't. The starting point is "which P&L line do we want to move, and what's the shortest path there", not "what can we build with AI". The technology choice comes after.

EBITDA isn't an arbitrary metric for this exercise. It's the figure a purchase-price multiple gets applied to, the figure a covenant measures, and the figure an operating partner is meant to read progress off of. An AI project that doesn't connect to any of these dials might be strategically interesting. It's just hard to defend when the next quarterly review comes around.

That doesn't mean every AI initiative needs a euro-denominated business case on day one. It means you should know, before you start, which P&L or balance sheet line it touches (personnel cost, gross margin, working capital), even if the exact size can only be pinned down reliably after a pilot has run.

The most direct lever sits in functions where headcount hours scale linearly with case volume: customer service, order processing, first-level support, content production, data entry and maintenance. An agent that takes over recurring processing steps shows up directly in personnel or outsourcing cost.

The math behind it is simple enough to build together with any function head: hours per case, times case volume, times hourly rate, minus the residual capacity that stays needed for review and escalation. An agent rarely replaces the whole case. It shifts human work from processing to checking the exceptions. That residual figure has to be in the case, or you systematically overstate the effect.

The effect only shows up at scale. In the first phase, headcount rarely drops, because the freed-up capacity is initially needed for review and rework. The effect on the cost line appears once case volume grows without processing capacity having to grow with it, a growth function that, without agents, would scale linearly with headcount.

The second lever is less visible, but at sufficient volume at least as large: margin per transaction. Better search relevance, personalization, dynamic pricing and more targeted cross-selling don't have a dramatic effect on any single order. They work across the entire order volume, on conversion rate and basket value.

The same mechanism works in the other direction: an agent that catches fraud patterns or data-entry errors earlier reduces write-offs and reversals. That's margin too, not revenue, and it's routinely overlooked in prioritization discussions because it isn't a new feature, it's a quieter improvement to an existing one.

The reason this lever gets underestimated is attribution. A new campaign immediately gets its own line in performance reporting. Better search relevance spreads invisibly across thousands of individual decisions and only shows up in aggregate gross margin. If you want to capture this lever, you have to set up the measurement beforehand. Otherwise you can't later separate the effect from seasonal or market swings.

The least-discussed lever is capital tied up. Faster order-to-cash cycles, a faster reporting close, less rework from errors caught earlier, and shorter decision paths affect working capital and the cash conversion cycle, figures that matter at least as much to an investor's valuation as operating margin.

This lever is rarely mentioned first because it doesn't look like AI. An agent that shortens invoice checking from days to hours sounds unspectacular. It still moves a number that shows up in every cash flow plan.

For companies with seasonal business or tight working capital, this is often the fastest lever to capture: the underlying processes (invoice approval, inventory reconciliation, reporting consolidation) tend to be well documented and rule-based, because they're already subject to internal controls. That makes them easier to automate than processes with a lot of discretion.

The method is a walk through the P&L, not through the technology catalog. For every major cost line: what share is recurring processing with a recognizable pattern? For every revenue line: what share is decided by matching, relevance or timing? For working capital: where does one person wait on another person's decision, even though the decision is rule-based?

These questions are best answered in a workshop with the people accountable for each function, not with IT alone. The function head knows the actual case distribution; IT knows technical feasibility. Both perspectives together give you a map of levers you can prioritize from.

One warning here: levers often overlap. Faster invoice checking affects both process cost and capital tied up. Add both effects independently and you overstate the total impact. The map should therefore assign each lever to one primary effect, even while documenting side effects on other lines.

The obvious mistake is starting with the most interesting use case instead of the biggest one. An AI-generated marketing video is more visible than an automated invoice check, and moves considerably less EBITDA in most companies. The right order follows case volume and proven cost impact, not novelty value.

That also means taking unglamorous functions seriously: returns processing, complaints management, master data maintenance. They rarely make a good conference slide, but often carry the biggest lever, because their case volume is high and their pattern is recurring.

A chatbot answering common questions on the website is AI theater as long as it runs alongside the existing support team without changing its workload. An agent that independently handles part of the support caseload start to finish and escalates only the exceptions changes the function's cost structure. The difference isn't the technology, since both can run on the same model. It's whether the agent is embedded in the process and headcount planning, or sitting next to it.

For a PE portfolio company, this work pays off twice. A documented, working lever in EBITDA is a result in the current period and an argument in the exit story: a repeatable improvement transferable to other sites or product lines, not a one-time cost cut. Buyers and their advisers check this closely. A lever that only works because of one person's knowledge counts for less in valuation than one documented in process and system.

That's why value creation deserves the same documentation a technical due diligence expects: which lever, what effect, since when, with what operating model behind it. That's the difference between a story that holds up in the data room and one that only sounds good in the deck.