01 · Bankable value
Margin Points
Compounds directly into EBITDA
Largely within operating control
Can carry most or all of the deal return
2.5×
supported mainly by earnings growth
Reinventing Traditional Businesses Through Generative AI
Inflection Ventures partners with legacy service businesses - call centres, BPOs, managed services - to drive 10× value creation in less than 5 years, by shifting from labor-first to AI-first.
Quantifying AI ROI: What 50%+ Cost Reduction Actually Looks Like on a P&L
Quantifying AI ROI: What 50%+ Cost Reduction Actually Looks Like on a P&L
August 26, 2026
⏱ 8 min read
by Inflection Ventures
August 26, 2026︱⏱ 5 min read︱by Inflection Ventures
Shares
Private equity now has to buy its returns out of earnings. A deal that once needed about 5% annual EBITDA growth to return a benchmark 2.5x MOIC over five years now needs 10–12%, because borrowing costs sit in the 8–9% range, leverage runs at 30–40% and purchase multiples are in record territory - Bain's "12 is the new 5" (Bain, Global Private Equity Report 2026). AI is the lever most operating partners have nominated to close that gap. It is not yet showing up in the accounts: 39% of organisations attribute any EBIT impact at all to AI, and most of those put it at under 5% of enterprise EBIT (McKinsey, The State of AI, Nov 2025).
The gap between those two facts is not usually a deployment failure. It is an arithmetic one. Headline cost reduction figures - ours included, are quoted without the denominator that makes them mean anything on a P&L.
Private equity now has to buy its returns out of earnings. A deal that once needed about 5% annual EBITDA growth to return a benchmark 2.5x MOIC over five years now needs 10–12%, because borrowing costs sit in the 8–9% range, leverage runs at 30–40% and purchase multiples are in record territory - Bain's "12 is the new 5" (Bain, Global Private Equity Report 2026). AI is the lever most operating partners have nominated to close that gap. It is not yet showing up in the accounts: 39% of organisations attribute any EBIT impact at all to AI, and most of those put it at under 5% of enterprise EBIT (McKinsey, The State of AI, Nov 2025).
What you will find in this article:
• Why a cost reduction percentage is not yet a return, and what has to be added to make it one
• The full walk-down from headline saving to EBITDA margin points, on a real services cost structure
• The one number that decides whether AI margin alone carries a deal and it is not 50%
• What margin expansion does, and does not, buy at exit
The gap between those two facts is not usually a deployment failure. It is an arithmetic one. Headline cost reduction figures - ours included, are quoted without the denominator that makes them mean anything on a P&L.
What you will find in this article:
• Why a cost reduction percentage is not yet a return, and what has to be added to make it one
• The full walk-down from headline saving to EBITDA margin points, on a real services cost structure
• The one number that decides whether AI margin alone carries a deal and it is not 50%
• What margin expansion does, and does not, buy at exit
Shares
The Governing Idea
The Governing Idea
A 50% cost reduction is not a number an investor can use. The usable number is the margin points that survive two deductions , the share of the cost base the reduction actually touches, and the cost of running the AI itself , followed by an honest statement of what those points buy at exit and what they do not.
Cost reduction is the input. Margin expansion is the mechanism. Multiple expansion is the outcome, and it does not follow automatically from the first two. Most AI business cases stop at the input, present it as though it were the outcome, and leave the mechanism , the only part that shows up in the accounts , undescribed.
A 50% cost reduction is not a number an investor can use. The usable number is the margin points that survive two deductions , the share of the cost base the reduction actually touches, and the cost of running the AI itself , followed by an honest statement of what those points buy at exit and what they do not.
Cost reduction is the input. Margin expansion is the mechanism. Multiple expansion is the outcome, and it does not follow automatically from the first two. Most AI business cases stop at the input, present it as though it were the outcome, and leave the mechanism , the only part that shows up in the accounts , undescribed.
This matters more in the businesses we buy than anywhere else. Labour-first service companies sit in the headcount trap: to double revenue they must roughly double staff, so costs grow as fast as revenue and margins never improve with scale. AI is the only lever that breaks the link between output and headcount. That is precisely why the number attached to it needs to be stated carefully rather than loudly.
This matters more in the businesses we buy than anywhere else. Labour-first service companies sit in the headcount trap: to double revenue they must roughly double staff, so costs grow as fast as revenue and margins never improve with scale. AI is the only lever that breaks the link between output and headcount. That is precisely why the number attached to it needs to be stated carefully rather than loudly.
The Walk-Down
To make this concrete, it needs a real cost structure rather than an invented one. Teleperformance, the largest listed operator in exactly the kind of business this playbook targets, reported 2025 revenue of €10,209m, personnel costs of €6,950m , 68% of revenue , and EBITA before non-recurring items of 14.6% (Teleperformance SE, 2025 Annual Results, 26 February 2026). That shape is the template: two-thirds of revenue is people, and the margin is thin enough that a few points changes the business.
The Walk-Down
To make this concrete, it needs a real cost structure rather than an invented one. Teleperformance, the largest listed operator in exactly the kind of business this playbook targets, reported 2025 revenue of €10,209m, personnel costs of €6,950m , 68% of revenue , and EBITA before non-recurring items of 14.6% (Teleperformance SE, 2025 Annual Results, 26 February 2026). That shape is the template: two-thirds of revenue is people, and the margin is thin enough that a few points changes the business.
Somewhere between a third and 40% convertible is where AI margin alone covers the whole earnings requirement of a 2.5x deal. Below roughly 30% it does not, and the deal needs revenue growth to carry the difference. That threshold, not the 50%, is the number worth diligencing. It is also the number almost nobody measures before signing, which is what Stage 02 of our six-stage model , screening for AI convertibility rather than only financials , exists to fix.
Somewhere between a third and 40% convertible is where AI margin alone covers the whole earnings requirement of a 2.5x deal. Below roughly 30% it does not, and the deal needs revenue growth to carry the difference. That threshold, not the 50%, is the number worth diligencing. It is also the number almost nobody measures before signing, which is what Stage 02 of our six-stage model , screening for AI convertibility rather than only financials , exists to fix.
What Margin Points Buy & What They Do Not
Here the honest answer is less flattering than the marketing version, and it is the most useful part of the exercise.
Across 471 privately held companies backed by private equity in deals completed from 2023 onward, McKinsey found that moving AI into the operating model lifted revenue per employee by roughly 20% , but moved the median revenue multiple only from 13x to 14x. Multiples rose materially only when AI was embedded in the offering itself: 20x at product transformation, 31x at business building, where revenue per employee jumped from $118,000 to $180,000 (McKinsey, Beyond productivity, Jun 2026). The market's differentiation between opportunistic adoption and operating-model AI is, in McKinsey's word, negligible.
What Margin Points Buy & What They Do Not
Here the honest answer is less flattering than the marketing version, and it is the most useful part of the exercise.
Across 471 privately held companies backed by private equity in deals completed from 2023 onward, McKinsey found that moving AI into the operating model lifted revenue per employee by roughly 20% , but moved the median revenue multiple only from 13x to 14x. Multiples rose materially only when AI was embedded in the offering itself: 20x at product transformation, 31x at business building, where revenue per employee jumped from $118,000 to $180,000 (McKinsey, Beyond productivity, Jun 2026). The market's differentiation between opportunistic adoption and operating-model AI is, in McKinsey's word, negligible.
Inflection Ventures underwrites Path B against a target of 10x enterprise value creation in under five years. Stating the ladder honestly is what makes that a target rather than a slogan: 10x is not reachable by cost reduction, and the walk-down above is the first two stages of the climb rather than the whole of it.
4 Ways the Number Gets Overstated
Every inflated AI business case we have reviewed makes at least one of these four errors, and usually two.
Inflection Ventures underwrites Path B against a target of 10x enterprise value creation in under five years. Stating the ladder honestly is what makes that a target rather than a slogan: 10x is not reachable by cost reduction, and the walk-down above is the first two stages of the climb rather than the whole of it.
4 Ways the Number Gets Overstated
Every inflated AI business case we have reviewed makes at least one of these four errors, and usually two.
Underneath all four is a measurement habit: adoption is easy to count, EBIT is not. Only 39% of organisations attribute any EBIT impact to AI, and among the advanced early adopters McKinsey studied in service operations, attributing more than 10% of EBIT to gen AI was the marker of the leading group (McKinsey, From promising to productive, Aug 2024) , a bar most programmes do not clear because most programmes never set it.
Underneath all four is a measurement habit: adoption is easy to count, EBIT is not. Only 39% of organisations attribute any EBIT impact to AI, and among the advanced early adopters McKinsey studied in service operations, attributing more than 10% of EBIT to gen AI was the marker of the leading group (McKinsey, From promising to productive, Aug 2024) , a bar most programmes do not clear because most programmes never set it.
What Is Your Convertible Share?
What Is Your Convertible Share?
Cost reduction is the input. Margin expansion is the mechanism. Multiple expansion is the outcome , and it is the one that does not arrive on its own.
The number that decides whether an AI programme is a value creation plan or an efficiency project is not the reduction percentage. It is the proportion of the cost base that reduction can reach. Everything downstream , margin points, earnings growth, whether the deal clears 2.5x on operations alone , follows from it, and it can be measured before anyone signs anything.
Cost reduction is the input. Margin expansion is the mechanism. Multiple expansion is the outcome , and it is the one that does not arrive on its own.
The number that decides whether an AI programme is a value creation plan or an efficiency project is not the reduction percentage. It is the proportion of the cost base that reduction can reach. Everything downstream , margin points, earnings growth, whether the deal clears 2.5x on operations alone , follows from it, and it can be measured before anyone signs anything.
Bring us a P&L and we will walk it down line by line, name your convertible share, and tell you what it is worth in margin points rather than in percentages. Explore the two tracks at www.inflectionventures.co or connect with the Inflection Ventures team.
Bring us a P&L and we will walk it down line by line, name your convertible share, and tell you what it is worth in margin points rather than in percentages. Explore the two tracks at www.inflectionventures.co or connect with the Inflection Ventures team.
Copyright © 2026 Inflection Ventures. All rights reserved.
Copyright © 2026 Inflection Ventures. All rights reserved.