What Does AI Really Do to Cash Flows? An Operator’s Guide for Private Equity

Field notes from an AI practitioner on turning AI ambition into measurable EBITDA and growth for investors.

Jan 28, 2026

I come at this from the operating side. Over the last two years, I’ve sat with banks, FMCGs, manufacturers, and PE-backed leadership teams. Different sectors, same recurring question: what does AI actually do to our cash flows—and on what timeline?

I’m not a dealmaker, and I’m not trying to tell private equity how to build LBO models. I’m describing what I’m seeing when the conversation gets real—in boardrooms, in operating reviews, and in the gap between ambition decks and P&L reality:

In most boardrooms, AI is now hygiene; the absence of a credible AI trajectory increasingly attracts a discount.
In a five- to seven-year hold, AI will either compress cash flows faster than expected or create value that never made it into the model.
In many processes, AI still shows up as narrative or a generic “operational improvement” line, rather than something priced with the same discipline as leverage, churn, or capex.

According to Bain, the macro context leaves little room to hide. Global M&A in 2024 was roughly USD 3.6 trillion. Buyouts represented about 19% of that (up from 16% in 2023). Buyout-backed exits reached around USD 468 billion across roughly 1,470 deals—a recovery, but still below the five-year average. At the same time, distributions as a share of NAV have fallen by roughly half compared with a decade ago, while buyout AUM has more than doubled.

In plain language: funds are holding more assets for longer, exits are harder, and LPs want a forward-looking value-creation story—one that collides with the AI wave, not just a historical IRR curve built on financial engineering.

If AI is going to affect almost every business you own or underwrite, the question becomes how private equity should actually price AI across sourcing, diligence, value creation, and exit—without turning into a venture fund.

The rest of this piece is my attempt to answer that from the field.

The Pricing Gap: When “AI-Powered” Means Nothing

In practice, I still see three recurring patterns when AI shows up in deal processes.

1) The narrative premium

AI appears in the Confidential Information Memorandum (CIM). Management talks about copilots and automation. The model assumes higher revenue CAGR or an extra 200–300 basis points of margin expansion from “AI efficiency.” AI is treated as “found money”: productivity that flows straight to EBITDA at low execution risk, usually without a clear mapping to P&L lines or adoption constraints.

2) The narrative discount

AI is framed mainly as disruption risk. Teams tag assets as “at risk of AI displacement,” shave the multiple a little or step away, but rarely quantify that risk or test for offsetting AI upside in the same asset.

3) The implicit hygiene test

Finally (this is closer to how investors I have spoken with described their practice), AI is now seen as standard hygiene: the absence of any credible AI trajectory is a red flag that can justify a discount or a faster exit. A demonstrated ability to adopt and scale technology can justify a premium. But even here, AI is often assessed qualitatively (“this management team can change”) rather than priced through a structured lens.

In my view, all three patterns share a problem: they treat AI as a label, not as a variable that depends on business model, infrastructure, corporate culture, and economics.

The divergence is already visible. A 2025 MIT survey shows that Enterprise AI programmes in big organizations suggest that 80–95% of AI projects fail to reach production or deliver measurable business impact. For many traditional companies, AI spend has shown up as capitalized software, consultants, and “innovation labs”—with no visible change in unit economics. The opportunity was real on slides; absorption capacity in data, infrastructure, and company culture was not.

At the same time, some PE owners are starting to treat AI as something to price and operate against, not just to market.

The gap between those stories is not who “uses AI” and who does not, but how.

A Value Equation for AI: Transformation Yield

AI is not a separate asset class. It is a force that reshapes the future cash-flow curve of almost every asset it touches. Sometimes it steepens that curve. Sometimes it flattens it. Sometimes it inverts it.

I am trying make that explicit in underwriting with a simple lens (not entirely mathematical but rather directional):

AI Transformation Yield = (AI Opportunity − AI Disruption) × Absorption Capacity

AI Opportunity is the credible upside from AI use cases: cost reduction, revenue growth, pricing power, data monetization, or multiple expansion.
AI Disruption is revenue erosion or margin compression from substitution, AI-native entrants, or regulatory constraints.
Absorption Capacity (from 0 to 1) is the share of net opportunity the company can realistically capture within the hold, given its industry, stack, organization, culture, and economics. It includes the velocity of change to capture the opportunity or the risk of being disrupted (speed matters!).

Example: suppose base-case EBITDA is 100. AI Opportunity could add 20; AI Disruption could remove 15. If Absorption Capacity is 0.4, net yield is (20 − 15) × 0.4 = 2, not 5.

In plain language: upside minus downside, discounted for execution reality.

Two clarifications that matter:

First, this is not an academic formula. It’s a forcing mechanism for assumptions that are often left implicit, especially around execution risk.
Second, the logic is two-step. Step one is sector-level: is there material AI disruption or upside within our hold, and of what kind? Step two is asset-level: given that context, is this company likely to be a winner, a laggard, or somewhere in between?

The point of an AI Transformation Yield is to make those steps explicit and to make it harder to hide execution risk inside optimistic margin bridges.

To use it, you need structure. That’s where MICE comes in.

MICE: Four Lenses Investors Already Understand

MICE is a practitioner framework designed to be intelligible to both investment and operating teams.

M: Market and Model
I: Infrastructure and Information
C: Capabilities and Culture
E: Edge and Economics

Together, these lenses decompose AI Transformation Yield into questions that are already familiar from commercial and technology diligence.

M: Market and Model

Question: How does AI change demand, pricing power, and the competitive set for this business model in this sector?

This is where many AI pricing errors begin. Teams often treat “AI risk” as a generic tag, when in reality it is a highly uneven function of industry structure and revenue model.

High-disruption examples include businesses that sell routine cognitive labor by the hour (basic BPO, shared services) or undifferentiated content.

High-opportunity examples include mission-critical workflow software with switching costs, industrial platforms where better routing and maintenance improve utilization, and vertical software with proprietary workflows and data.

Investors already apply a version of this logic. When they believe disruption is coming, they map companies along a spectrum—forefront adopter, possible transformer, laggard.

M is that discipline, made explicit for AI.

I: Infrastructure and Information

Question: Does this company have the technical and data spine to do anything useful with AI at pace, or are you really buying a two-year plumbing project?

This is often the hidden constraint on Absorption Capacity.

Stack modernity: modular, API-driven architectures versus monolithic legacy systems.
Data position: clean, governed, unified data versus fragmented, vendor-owned, or compliance-locked data.
Integration readiness: automated workflows and orchestration versus manual processes and point solutions.

Pricing implication: if an asset is presented as “AI-ready” but Infrastructure is weak, you are not buying an AI flywheel—you are buying the right to fund a technology clean-up (!)

That should show up in entry pricing, capex budgets, or milestone-based structures.

C: Capabilities and Culture

Question: Does this organization have the people and operating model to turn AI from PowerPoint into P&L?

This is where the real bottleneck usually sits.

Leadership ownership: CEOs and CFOs accountable for specific outcomes from each AI initiative, not isolated experiments or siloed initiatives.
Basic fluency: managers able to identify use cases; front-line teams able to adopt tools safely and quickly.
Governance that enables: clear policies that unblock operations, with incentives that reward measured adoption.

Empirically, companies that already use data in daily decisions adopt AI far faster than those where data remains a monthly report. When AI transformation succeeds or fails, it usually comes down to management choices, behaviors and incentives.

Absorption Capacity is mostly driven by this dimension. Technology is necessary, but culture and leadership determine whether AI sticks and adoption spreads.

E: Edge and Economics

Question: What does AI do to structural advantage and unit economics?

This is the bridge back to cash flows and multiples.

On the edge side, proprietary data and workflows can compound; distribution and brand can increase long-term value through better targeting and retention; regulated or complex processes can be optimized in ways that are difficult to copy.

On the economics side, the discipline is to locate the P&L line. AI can bend cost curves—often via slower headcount growth rather than layoffs—and lift revenue through higher conversion, pricing power, or new revenue models. The common modelling error is treating AI as a one-off step change rather than a change in slope.

The most common error observed in deal models is the assumption of a singular 300-basis-point margin increase, rather than recognizing a structural alteration in marginal cost or revenue per full-time equivalent (FTE), which is crucial for acknowledging the transformative impact of AI on the economic foundation of the company’s operating model.

Taken together, M, I, C, and E provide a reasonably robust view of:

how large AI Opportunity and AI Disruption really are, and
how much of the net can plausibly be captured within the hold period.

Evidence in the Wild: Negative and Positive Yield

With the AI Transformation Yield lens, recent history looks less random: similar technologies create value where MICE is strong and destroy it where it is weak.

Negative Yield (AI as a value destroyer)

Volkswagen’s Cariad
Volkswagen set up Cariad to build a unified, AI- and software-defined platform across its brands. By 2025, the unit had accumulated an estimated USD 7.5 billion in losses over roughly three years. It contributed to launch delays of key models such as the Porsche Macan EV and Audi Q6 e-tron by more than a year, and triggered around 1,600 job cuts.

From an AI Transformation Yield perspective, this was not a lack of ambition. Market and Model offered genuine upside: software-defined vehicles, over-the-air updates, and data-driven features are real value pools in autos. But weak Infrastructure (an unstable, fragmented stack across brands), weak Capabilities (organizational complexity and slow decision cycles), and unclear near-term Economics pushed the yield negative despite massive spending. The opportunity was real; absorption capacity was not.

Health insurers’ AI denial tools
Several US health insurers now face class actions and regulatory scrutiny over AI-driven claims engines alleged to auto-deny large volumes of claims without adequate human review. Some tools were reported to process denials at roughly “one claim per second.”

Here, AI was positioned as pure OpEx reduction. But weak governance and opaque decisioning (a combination of Capabilities and Infrastructure failures) created regulatory, remediation, and reputational exposure that can easily overwhelm any initial savings. This is classic negative yield in a legacy model: AI improves throughput, but not trust, explainability, or resilience.

Industrial and enterprise AI project failures
Surveys of AI initiatives in large organizations consistently suggest that 80–95% fail to reach production or deliver measurable business value. In traditional manufacturers, logistics players, and utilities, this often shows up as capitalized software, consulting spend, and “AI labs” that never touch the P&L.

In these cases, AI Opportunity existed on paper, but Absorption Capacity—driven by fragmented data, legacy infrastructure, and disengaged business owners—was near zero. In hindsight, yield was negative not because AI was the wrong technology (although it may have been as well!), but mostly because it was priced as an unqualified positive.

Across all three examples, investors implicitly assumed high opportunity, under-modelled disruption and second-order risk, and over-assumed absorption capacity.

What looks like execution failure often starts as mispricing.

Positive Yield (AI as a Force Multiplier)

DHL
Logistics players report that AI-driven route optimization, dynamic scheduling, and warehouse automation can deliver double-digit reductions in transport and handling costs, shorter delivery times, and higher asset utilization. In one documented case, network optimization and load planning produced on the order of a 30% reduction in logistics costs for a traditional operator.

Here, the Market is largely unchanged. The yield comes from strong Infrastructure (route, load, and timing data), improving Capabilities (teams adopting AI-generated plans in daily operations), and better Economics (fewer empty miles, more volume per truck). The result is clear positive yield in a classic PE vertical.

NileDutch
AI case studies in manufacturing logistics describe OEMs and tier-1 suppliers using historical shipment and inventory data to forecast container needs 10–12 weeks ahead, reduce empty repositioning, and cut fleet and storage costs while maintaining service levels.

For a traditional industrial shipper, this is almost pure E (Economics) improvement: the same revenue delivered with fewer assets and lower working capital. The enablers are better Infrastructure (clean, usable supply-chain data) and Capabilities (planners trusting and acting on AI-generated forecasts).

Global CPG and retail supply chains
Across consumer goods and retail, AI-driven demand forecasting and routing have delivered 15–20% reductions in logistics costs, roughly 30% faster deliveries, around 30% lower forecast error, and approximately 15% lower safety stock. One large consumer company cites about USD 300 million per year in inventory savings from such a programme.

BCG’s work on cost transformation similarly profiles a global consumer packaged goods firm using GenAI and advanced analytics to deliver 60–90% efficiency gains on specific marketing and back-office tasks as part of a broader cost-base reset. These are not SaaS firms. They are traditional manufacturers and retailers where AI measurably improves working capital and SG&A.

European energy provider: AI in payments and procurement
BCG also highlights a European energy provider that used AI and GenAI in payment reviews and procurement to detect overpayments and streamline invoice processing as part of a multi-year cost-transformation programme.

In a low-growth, regulated sector, the Market is flat. But modernized finance and ERP foundations, combined with changed operating routines, translate into recurring savings and better margin resilience—positive yield in a utility-like asset.

Common pattern
Across these examples, Market and Model are neutral to favorable. Infrastructure and Capabilities are strong enough to deploy AI at scale. Edge and Economics are amplified—lower cost per unit, higher utilization, and reduced working capital—rather than eroded.

Applying the AI Transformation Yield Across the Deal Lifecycle

The practical question for private equity is how to make this usable across sourcing, diligence, value creation, and exit. The aim is not a new theory, but a discipline that can be baked into Investment Committees’ templates and post-close operating plans.

Sourcing: Avoid, Reprice, or Pursue

At sourcing, you do not need a detailed model. You need to avoid blind spots.

The first screen is sector-level: over a five- to seven-year hold, is AI likely to be a primary driver of disruption or of value creation in this industry?

In practice, sectors tend to fall into three buckets:

Avoid / distress only: AI-driven disruption structurally outweighs opportunity, with no credible pivot inside a PE hold.
Reprice / require a plan: both opportunity and disruption are high; entry pricing assumes you can move MICE scores.
Pursue: favorable Market and Model, with Infrastructure and Capabilities that allow AI to amplify existing Edge.

For assets that pass this screen, require a one-page AI Transformation Yield snapshot in the first IC memo: rough opportunity, rough disruption, an initial MICE view, and a provisional absorption-capacity range. This forces early clarity on where AI actually sits in the deal thesis.

Due Diligence: From Snapshot to Transformation Assessment

Traditional tech diligence asks whether the system works. AI diligence has to ask whether the organization can capture AI value at the pace the model assumes.

Three practical shifts help.

First, quantify opportunity and disruption explicitly. Use customer interviews, win-loss data, and product roadmaps to estimate how AI could change volumes, pricing, and churn. Use job postings, process maps, and competitor behavior to identify where automation or substitution will likely hit first. Build at least three scenarios—AI-accelerated decline, AI-neutral, and AI-compounding—and make visible what has to be true in each.

Second, score absorption capacity using MICE. Build a simple one-to-five score for Market, Infrastructure, Capabilities, and Edge/Economics. Convert that into a rough absorption-capacity percentage rather than assuming you can capture 100% of net upside. An asset strong across all four dimensions might credibly capture 40–60% of net AI upside within the hold. One with weak infrastructure and culture may struggle to realize more than 10–20%, even with external support.

Consider absorption velocity : how fast can the company drive change and capture value until the competition and industry catches up..

Feed those assumptions directly into valuation and deal structure. If the base case assumes 300 basis points of AI-driven margin expansion but absorption capacity is low, you either reduce the assumption or explicitly budget for the transformation work required to unlock it, including leadership change.

Third, map concentration risk and resilience. Identify dependence on single cloud providers or model vendors, exposure to export controls or regulatory change, and realistic switching paths. For some assets, specific remediation actions—model diversification, hybrid or local compute, data-governance upgrades—should be explicit in the 100-day plan.

The output is a one-page AI Transformation Yield section in the final IC memo summarizing opportunity, disruption, absorption capacity, key MICE gaps, and the first three AI workstreams post-close.

Value Creation: Moving the Needle

Across portfolios, the most common pattern I see—and hear from PE investors—is “a hundred pilots, no scale.” AI becomes tourism: demos, proofs of concept, innovation days, and dashboards that never move the P&L.

A more disciplined approach looks like this:

Start at the top of the P&L. Pick three use cases that can move EBITDA or revenue materially within 18–24 months. In many businesses, these sit in engineering or product, back-office operations, and sales or pricing.

Run 90-day industrialized sprints. Set clear metrics upfront. Pair a business owner with a tech lead. Define scale, pivot, and stop thresholds early.

Centralize patterns, decentralize ownership. The GP can standardize governance, vendors, and risk frameworks; share playbooks; and provide scarce talent. Portfolio companies own outcomes.

Invest in people and change. If you do not budget for training, role redesign, and adoption incentives, productivity gains tend to leak into slack rather than margins.

See here the details of my past engagement showing how we combined “HR & Tech” to drive AI transformation at a large multi-national FMCG, headquartered in Singapore.

When HR and Technology Co-Design the Future of Work: How do you grow the business significantly without growing headcount at the same pace?

Exit: Turning AI Into a Multiple

By the time you sell, buyers will ask two questions: how has AI changed the business, and how can we trust the numbers?

Vendor materials should be able to answer both.

They should show before-and-after MICE and Transformation Yield: what changed in AI exposure, infrastructure, capabilities, and unit economics under your ownership.

They should document concrete deployed use cases: dates, ownership, outcomes, adoption rates, cycle-time improvements, and attribution where feasible.

They should evidence resilience and optionality: where models run, how data is governed, vendor concentration, and credible switching paths.

In a world where exits remain constrained relative to AUM, the assets that trade first—and at better multiples—will be those that can evidence AI Transformation Yield, not those that merely mention AI.

The Macro Wildcard: AI Bubble Risk and Geopolitical Fragmentation

Even the cleanest deal model sits inside a macro and geopolitical environment it does not control. For AI, that environment is unusually unstable.

As I have shared in prior articles The Dependency Economy of AI comparing 25 National AI Strategies and Reading the U.S. National Security Strategy as a Technology Blueprint: geopolitics now impacts technology decisions and creates a unique set of new challenges for companies.

Debate around an “AI bubble” has moved from academic circles into mainstream earnings calls and policy discussions. Justin Yifu Lin, a senior Chinese economist and former World Bank chief economist, has argued that the current US-led AI boom shows characteristics of an asset bubble: heavy capital concentration, rising valuations ahead of clear monetization, and infrastructure investment justified by expectations of future productivity rather than current demand. In his view, a correction could coincide with—and actively reshape—China’s next Five-Year Plan, with the state using fiscal and credit tools to accelerate AI infrastructure and industrial deployment even as private capital in the United States retrenches.

Whether one agrees with that specific thesis or not, it creates a scenario matrix that matters for private equity, because AI economics are unusually sensitive to capital cycles, policy, and geopolitics.

One scenario is correction and retrenchment. If valuations correct and capex retrenches in the US, AI infrastructure could become more state driven in some markets. Access to compute, chips and AI heavy components may become more regionally asymmetric.

A second scenario is a continued boom with consolidation. If the boom persists and monetization catches up, model scale and bargaining power may consolidate further among a few hyperscalers, with more intense regulatory and export control scrutiny.

Both scenarios have direct implications for AI Transformation Yield and for each dimension of MICE.

Market and Model. Entire AI subsectors can switch from growth to distress if funding dries up. Other areas (for example AI assisted manufacturing in countries with strong industrial policy) could accelerate.

Infrastructure and Information. Heavy dependence on a single foreign cloud or model provider in one jurisdiction becomes a macro risk parameter, not just an IT detail.

Capabilities and Culture. Boards and investment committees that understand this will expect scenario planning and “sovereignty options” (hybrid or local compute, containerized model stacks, clear migration paths).

Edge and Economics. Assumptions that GPU and cloud prices will fall predictably over the hold period are fragile if export controls tighten or if subsidy regimes shift. In some cases, paying more for diversified and resilient architecture will be rational if it protects continuity and pricing power.

The point is not to forecast bubble timing. It is to make sure each significant AI intensive asset has at least a basic view of how a boom or correction would change its MICE profile and Yield.

Using AI to Improve Your Own Judgment

It’s clear from my conversations with PE professionals that they are also looking at AI as an opportunity to transform themselves as well. As such, two of the most important AI value pools for private equity include:

Knowledge management
PE firms sit on decades of deal memos, diligence reports, and portfolio outcomes. AI can surface patterns across that history—what actually drove margin expansion, where tech debt destroyed value, how sector narratives evolved—far more effectively than traditional knowledge-management tools.

Relationship intelligence
Relationships with LPs, CEOs, bankers, and advisors drive access and allocation. Applied carefully, AI can analyze communication flows and networks to help firms priorities outreach, spot gaps, and deepen relationships more systematically.

The firms that underwrite AI Transformation Yield best will be those that use AI to upgrade their own collective judgment.

Monday Morning Questions

All of this only matters if it changes the questions asked in Investment Committees and portfolio reviews. For each sector and each asset, five questions are worth asking—explicitly and in plain language:

Over the hold period, is AI a net headwind or tailwind for this asset—and what is our rough Transformation Yield today versus under our ownership?
Where are the biggest MICE gaps, and which of them can we realistically move within three to five years?
Are we running a hundred experiments—or three industrialized plays that materially move the P&L?
Do we have the leadership, data maturity, and AI fluency to capture upside—or will gains leak into slack?
If we had to sell in three years, could we evidence on a single slide how MICE and Transformation Yield improved since acquisition?

Resilience is the new multiple. Pricing AI as both a multiplier and a disruption force is now central to underwriting new deals and managing existing portfolio companies.

I believe the AI Transformation Yield can provide a practical compass for deciding where to avoid, where to reprice and fix, and where to double down.

About the Author

Damien Kopp is a Singapore-based strategist specializing in AI, digital transformation, and technopolitics. He is the Managing Director of RebootUp, where he advises governments, financial institutions, and Fortune 500 companies on building resilient, sovereign-ready AI architectures.

With more than 25 years of experience across Europe, North America, and Asia, Damien works at the intersection of technology, geopolitics, and organizational design—helping leaders navigate a world where AI infrastructure has become a strategic dependency.

He is also the founder of KoncentriK, a thought-leadership platform exploring the deeper structural forces shaping the digital future, and an Associate Faculty member at Singapore Management University, where he teaches applied AI to senior executives.

Damien holds an Executive MBA from Kellogg–HKUST and a Master’s in Electronics Engineering from ESME Sudria.

Asia Tech Lens

Discussion about this post

Ready for more?