Artificial intelligence has already cleared its first hurdle: distribution. Generative AI reached mass-market scale at a speed that previous general-purpose technologies never approached. The IMF notes that ChatGPT reached 100 million users within months, compared with years or decades for technologies such as the internet, computers, television, telephones, and mobile phones. That difference matters because it dramatically reduces the friction involved in putting a new technology into workers’ hands. But it also creates a misleading impression of economic maturity. Consumer access can spread almost instantaneously; organizational transformation cannot.

That distinction is now becoming central to private equity. Most sponsors no longer need to be convinced that AI matters. The more difficult question is whether adoption inside a portfolio company produces a measurable change in earnings, cash flow, growth, or ultimately enterprise value. BCG argues that, despite extensive experimentation, relatively few PE firms can yet demonstrate meaningful AI returns across a broad portion of their portfolios. Its explanation is instructive: many businesses have deployed tools, but relatively few have redesigned the operating model around them. Merely distributing access to an LLM can improve individual efficiency; it rarely creates material and measurable enterprise value by itself.

This distinction is becoming more consequential because private equity's broader return environment has changed. McKinsey reports that median buyout entry valuations reached a record 11.8x EBITDA in 2025, above the 2010–2022 average of 9.1x. At the same time, leverage is contributing less to entry structures than historically, while average holding periods have risen to 6.6 years. More than half of buyout-backed portfolio companies had been held for four years or longer by 2025. In other words, sponsors are paying demanding prices while capital remains tied up for longer. Operational performance has therefore become progressively harder to treat as a secondary return lever.

The implication is straightforward: AI is moving from an experimental technology budget into the value-creation plan. The relevant question is no longer whether a portfolio company “uses AI.” It is whether AI changes a recurring commercial or operational decision enough to alter the P&L.

The macroeconomic case for AI ultimately rests on productivity. In the Solow growth framework, technological progress raises total factor productivity—the efficiency with which capital and labor are converted into output. Unlike simply adding workers or physical capital, sustained improvements in productivity can increase output per worker without encountering the same diminishing-return dynamics. AI therefore matters economically to the extent that it increases the productivity of labor, capital, or both.

Yet estimates of the aggregate effect remain unusually wide. The IMF’s work on Europe illustrates the tension between substantial task-level improvements and much more modest economy-wide effects. Experimental studies cited by the IMF have produced productivity improvements ranging from roughly 14% in some occupations to more than 50% in software engineering. But when task exposure, economic feasibility, adoption rates, and sector composition are incorporated into a macroeconomic framework, the IMF’s preferred scenario produces a cumulative European TFP gain of only about 1.1% over five years. Regulation could reduce those gains further.

This is an important warning for private equity. There is a large difference between demonstrating that an employee can complete one activity faster and demonstrating that a company can generate more EBITDA. Time saved can disappear into unused capacity. Faster analysis can create more analysis rather than better decisions. A customer-service agent can reduce handling time without allowing headcount to fall or volumes to rise. A salesperson can produce proposals faster without improving conversion. Productivity becomes financial value only once management redesigns capacity, targets, staffing, workflows, incentives, or pricing around the new capability.

The adoption data reinforces the gap. The World Bank finds that corporate adoption remains far behind individual experimentation and is heavily skewed toward larger companies. Its OECD data showed average firm AI adoption at only around 8% in the earlier survey period, with adoption far higher among enterprises employing 250 or more people. The dispersion across countries is equally substantial: some economies already show large-company adoption approaching half of firms, while adoption among the overall business population remains much lower.

More recent U.S. evidence indicates that the frontier is moving quickly. Federal Reserve researchers estimate that approximately 18% of U.S. firms had adopted AI by year-end 2025, with the adoption rate having increased sharply before a survey-methodology change late in the year. That is rapid progress, but still means the majority of businesses have not yet integrated AI into operations at the firm level.

For sponsors, this creates both an opportunity and a constraint. PE ownership can potentially compress the adoption curve by bringing portfolio-wide expertise, capital, vendors, governance and execution discipline to businesses that would otherwise struggle to build those capabilities independently. But the same evidence suggests that simply purchasing enterprise licenses is unlikely to create differentiation. The scarce capability is increasingly not access to AI. It is the organizational capacity to absorb it.

One of the clearest lessons emerging from AI implementations is that the model itself is rarely the dominant determinant of the outcome.

BCG’s experience with a technology company attempting to increase software-development productivity is particularly instructive. AI tools were already deployed and more than 80% of weekly users had access, yet productivity gains remained limited. The eventual intervention went substantially beyond the coding tool itself: project-management processes changed, incentives were restructured, performance was measured differently, and roles and workflows were redesigned. BCG attributes approximately 70% of the resulting productivity uplift to people, organization and processes, with technology and algorithms contributing the remainder. Developer productivity ultimately increased by roughly 60% compared with around 10% before the intervention.

This is the dividing line between an AI pilot and an AI value-creation program. A pricing algorithm that identifies discount leakage does not create EBITDA. EBITDA is created when its recommendation enters the quoting process, sales managers understand when exceptions are permitted, compensation does not encourage unnecessary discounting, realized prices are tracked against a baseline, and management intervenes when adoption falls.

A churn model does not protect recurring revenue. Value is created when customer-success teams are automatically supplied with prioritized accounts, intervention protocols are defined, outreach happens early enough to matter, and retention is measured relative to a comparable baseline.

An inventory forecast does not release working capital. Value appears when purchasing decisions, order quantities, safety-stock policies and production planning actually change.

This is why BCG separates AI implementation into Deploy, Reshape and Invent. Deploy means making tools available inside an existing operating structure. Reshape means changing workflows, organizational structures, responsibilities and processes so the productivity improvement can be captured in the P&L. Invent goes further, embedding AI into the customer proposition or creating an entirely new business model. BCG argues that most PE portfolio companies remain primarily in the first category.

PwC reaches a similar conclusion from the fund perspective: leading PE organizations increasingly treat AI not as another point solution but as the basis for a new operating model centered on data, analysis and redesigned ways of working.

The economic hierarchy therefore runs:

Tool access → employee adoption → workflow redesign → operating KPI change → financial impact → enterprise value.

Every link matters. The further a program remains from the right-hand side of that chain, the more cautiously a sponsor should count its “AI value.”

Portfolio AI strategies are often built as catalogs of technologies: copilots, agents, predictive models, RAG systems, automation platforms. For value creation, the better taxonomy is financial.

AI can create economic value through four principal channels: revenue growth, margin improvement, cash conversion and decision quality. Each use case should begin with one of those mechanisms and work backward to the workflow and technology required.

BCG’s functional assessment points toward the areas with the clearest near-term opportunity. R&D, sales, marketing, customer service and customer success sit at the top of its priority matrix. The distinction between augmentation and automation is important. Sales and R&D, for example, remain heavily augmentation-oriented because relationship judgment and innovation remain human-intensive. Customer service offers greater direct automation potential because portions of the workflow are standardized and high-frequency.

The build-versus-buy decision should follow the same economic logic. Building internally makes most sense when proprietary data or a proprietary workflow can produce durable strategic differentiation. Buying is more attractive where the process is standardized and implementation speed matters more than customization. Partnership occupies the middle: it can accelerate implementation where internal technical capabilities are limited but substantial integration or sector customization remains necessary.

The mistake is treating this as an ideological choice. A portfolio company does not become more “AI-first” because it built its own model. Proprietary development is only valuable when proprietary capability produces better economics than the available commercial alternative after development cost, maintenance, talent requirements, model risk and time to deployment are included.

Pricing is among the most attractive PE applications because the transmission mechanism to EBITDA can be unusually direct. Models can identify customer willingness to pay, discount leakage, product-level price elasticity, renewal risk, contract inconsistencies, promotional effectiveness and opportunities for differentiated increases.

PwC points specifically to dynamic pricing and improved visibility into customers, SKUs, inventory and margins as examples of how better data infrastructure can transform portfolio-company decision making. It also highlights AI-supported go-to-market optimization, customer analytics and pricing strategy as routes to faster top-line improvement.

Roland Berger provides useful implementation evidence. In one portfolio-company case, dynamic pricing generated $5 million of additional margin within one quarter. The relevance is not the individual dollar figure—which is company-specific—but the structure of the intervention: a narrow use case tied to a measurable financial pool, implemented quickly enough that results could be verified rather than inferred.

Sales productivity follows the same principle. AI can prioritize leads, enrich accounts, summarize calls, generate proposals, maintain CRM records and identify cross-sell opportunities. BCG case experience includes materially faster proposal development and improved win rates from redesigned AI-supported sales processes. The KPI, however, should ultimately be revenue per seller, conversion, pipeline velocity or selling capacity—not prompts generated.

Procurement provides another direct bridge from AI to P&L. Spend classification, supplier benchmarking, contract extraction, price variance analysis and should-cost modeling can rapidly identify opportunities. Roland Berger reports one portfolio implementation that reduced indirect spend by 7% in less than ten weeks.

But “identified savings” should never be equated automatically with EBITDA. A sourcing model may identify a $5 million opportunity while only $2 million reaches the income statement after implementation delays, volume effects, supplier switching costs and operational constraints. The value-creation scorecard should record realized savings, preferably finance-validated, rather than theoretical opportunity.

The same logic applies to margin leakage: unbilled services, unnecessary discounting, incorrect freight charges, warranty leakage, contract noncompliance, labor scheduling and customer-level loss making. These may be less glamorous than building a proprietary AI agent, but for PE they can be superior use cases because the financial pool already exists and the results can be measured against historical transactions.

Collections prioritization, inventory optimization and demand forecasting can generate cash even without raising EBITDA immediately. A model that predicts which invoices will become delinquent allows collections staff to direct attention before payment deteriorates. Better demand forecasting can reduce excess inventory and safety stock.

Again, prediction is only the first half of the intervention. Unless purchasing parameters, collections queues, escalation rules or inventory policies change, the model has produced information rather than value.

Forecasting is particularly important because it sits upstream of many other value levers. Better revenue forecasts affect staffing. Better demand forecasts affect inventory. Better cash forecasts influence liquidity planning. Better pipeline models affect commercial resource allocation.

Roland Berger reports AI forecasting applications that reduced forecast error by approximately 25% while accelerating planning cycles.

The sponsor’s objective should therefore not be “better forecasting” in isolation. It should be identifying which management decisions improve because forecast quality improves.

The available evidence increasingly supports an augmentation-heavy near-term model rather than a simple labor-replacement thesis.

Federal Reserve Governor Michael Barr has noted that financial-sector adoption is concentrated in operational-efficiency applications such as text analysis, classification, internal information search and customer-facing processes. More consequential functions such as credit decision support, fraud detection and financial-market applications are developing too, but require materially stronger controls because outputs must be explainable, precise, replicable and compliant.

The productivity evidence also shows a relationship between how deeply AI is incorporated into work and how much time employees save.

Management, business and finance, computer and mathematical occupations, and other information-heavy professional categories already show measurable time savings associated with higher AI use. But a PE owner should immediately ask a second question: what happens to the saved hour?

There are only a few economically meaningful answers. The employee produces more. The company eliminates an external expense. The business avoids hiring an additional employee. Service improves enough to drive retention. A bottleneck disappears and revenue capacity expands. Or management chooses to reduce staffing.

Anything else is primarily convenience. Company size also matters because adoption capability is not evenly distributed.

The attached evidence points toward stronger labor-productivity effects among larger firms. That is consistent with the World Bank’s finding that AI adoption is significantly higher among larger enterprises. Scale helps because large organizations can afford data infrastructure, specialized talent, implementation teams and governance. Middle-market companies often face the opposite problem: they have enough process complexity for AI to matter, but not enough internal capability to execute transformation independently.

That is precisely where private equity may possess a structural advantage. A sponsor can spread scarce expertise across multiple businesses, standardize vendors, aggregate training costs, create common measurement frameworks and transfer successful implementations between portfolio companies.

But the labor conclusion should not be oversimplified into “AI means fewer jobs.”

The post-ChatGPT divergence between equity-market valuations and aggregate job openings is visually striking, but it should not be interpreted mechanically as causal evidence of AI displacement. The Federal Reserve’s firm-level research reaches a much more cautious conclusion. Across more than one million firms in its dataset, researchers find no evidence thus far that higher AI exposure or firm-level AI investment has reduced subsequent job postings. The measured effects are economically small and, where statistically significant, slightly positive.

That does not imply labor markets will remain unchanged. It suggests the transition is currently more complicated than a one-for-one substitution story.

PE150’s previous analysis identifies an important counterweight to automation risk: occupations with high exposure often also possess significant capacity to adapt. Its analysis finds approximately a 60% correlation between automation exposure and adaptability, suggesting that many highly exposed professional occupations may change materially without disappearing altogether.

For portfolio management, the implication is that workforce planning needs to move beyond crude headcount-reduction assumptions. The more useful questions are which tasks disappear, which expand, what new bottlenecks emerge, where spans of control can widen, and which employee groups can operate at materially higher output with AI support.

AI does not remove management. It raises the standard for management.

PwC describes a familiar middle-market starting point: fragmented applications, inconsistent data, manual workarounds and limited standardization. These weaknesses existed before generative AI, but AI makes their economic cost more visible. A model cannot make reliable customer-level recommendations if product definitions conflict across systems. An inventory optimizer cannot perform properly when lead-time data are stale. A sales agent cannot improve account prioritization if CRM records are incomplete.

The temptation is to conclude that every portfolio company therefore needs a multi-year data transformation before using AI. That is equally mistaken. Data readiness should instead be evaluated against the specific economic use case.

For each initiative, management should ask:

A spend-analysis exercise may work from monthly ERP extracts. Dynamic pricing may require far more frequent data. Predictive maintenance may depend on sensor-level streams. The architecture should therefore be proportional to the economic problem.

This approach also prevents data modernization from becoming detached from value creation. The best sequence is often use case → required decision → required data → minimum viable architecture, rather than architecture first and economic justification later.