Commercial Underwriting of AI-Enabled Value Creation

Commercial Underwriting of AI-Enabled Value Creation: The Cohres AI Value Creation Assessment

August 11, 2026

Artificial intelligence is increasingly moving from a technology consideration into an investment assumption.

Private equity firms are incorporating AI into sourcing, diligence and portfolio value-creation plans. Management teams are linking AI initiatives to revenue growth, productivity, cost reduction, margin expansion and operating leverage. Investors are increasingly required to determine not only whether companies are exposed to AI, but whether AI can materially improve the economics of the businesses they are underwriting.

The evidence, however, reveals an important gap between adoption and realised value.

McKinsey’s 2025 global research found that 88% of respondents reported regular AI use in at least one business function, yet only around 33% of organisations had begun scaling AI across the enterprise. Only 39% attributed any enterprise-level EBIT impact to AI, and most of those reported an impact of less than 5%.[1]

IBM’s 2025 CEO research produced a similar distinction: only 25% of AI initiatives had delivered their expected return on investment, while only 16% had scaled across the enterprise.[2]

The problem is not necessarily that AI fails to create value.

There is increasingly credible evidence that it can.

Apollo has reported meaningful cost reductions across selected AI initiatives within portfolio company Cengage, including lower content-production, lead-generation and customer-care costs. At Barnes Group, Apollo reported a five-times first-year return on an AI application supporting service technicians.[3]

Hg has separately reported more than $130M of EBITDA value-based uplift across its portfolio AI initiatives, alongside AI-enabled products contributing meaningfully to bookings at a number of portfolio companies.[4]

The emerging investment problem is therefore more nuanced.

Some AI initiatives are producing measurable economic value. Many others remain pilots, deployment commitments, productivity improvements or identified opportunities whose financial contribution has not yet been established.

For investors, those categories cannot be treated as equivalent.

A company may have credible AI applications, successful pilots, substantial deployment commitments and a compelling management narrative without demonstrating that those initiatives will create financial benefits at the scale required by an investment thesis.

This distinction is becoming increasingly important as AI enters private-market underwriting.

Deloitte’s 2025 research found that 86% of surveyed US corporate and private equity M&A leaders had integrated generative AI into their M&A workflows, including strategy, target screening and diligence.[5] At the portfolio level, major sponsors are increasingly incorporating AI into value-creation programmes covering revenue acceleration, customer service, software development, supply chains and operating efficiency.

Although, greater institutional adoption does not lower the evidentiary standard required for investment underwriting.

Blackstone has itself cautioned against AI initiatives becoming “performative pilots” that fail to reach EBITDA.[6]

That observation captures the central investment problem.

The relevant question is no longer simply whether a company is adopting AI.

It is: How much of the value expected from AI is credible enough to underwrite before capital is committed?

Cohres developed the AI Value Creation Assessment (AVCA) to address this question.

AVCA is a structured methodology for the commercial underwriting of AI-enabled value creation. It independently evaluates whether AI-enabled growth, productivity, margin expansion, cost improvement and operating leverage assumptions are commercially credible, economically attractive and realistically executable before capital is committed.

The objective is not to determine how technologically advanced a company is.

It is to determine whether investors should place conviction in the economic value AI is expected to create.

From AI Adoption to Investment Value

Investment theses have always depended on assumptions.

Revenue growth, market expansion, pricing power, margin improvement, operating leverage, customer retention, competitive advantage and management execution all influence the value an investor is ultimately willing to place on a business.

AI increasingly sits inside those assumptions.

Management teams may expect AI to automate labour-intensive activities, increase throughput, improve pricing, reduce maintenance costs, improve customer acquisition, accelerate product development or enable a business to scale without proportionately increasing its cost base.

The difficulty is that the path from AI capability to shareholder value contains several separate stages.

An AI application can work technically without being widely adopted.

A successful pilot can fail to replicate across a wider organisation.

Adoption can produce measurable productivity improvements without reducing costs or increasing output.

Operational improvements can be economically real but too small to affect financial performance.

And gross benefits can be partially or completely offset by implementation costs, recurring expenditure, human oversight, organisational constraints or competitive replication.

Research increasingly demonstrates these distinctions.

S&P Global Market Intelligence found that organisations scrapped an average 46% of AI projects between proof of concept and broad adoption.[7] An NBER field experiment involving more than 7,000 knowledge workers found that active users of generative AI saved approximately 2 hours of email time per week, but the study did not find corresponding changes in the quantity or composition of work performed.[8]

Both findings can be positive from an operating perspective while remaining insufficient, on their own, to support an investment assumption.

Two hours of employee capacity released is not automatically two hours of labour cost removed.

A deployed AI application is not automatically an adopted one.

An adopted application is not automatically a profitable one.

The investor therefore needs to follow the entire value-creation chain:

AI capability → operational or commercial improvement → measurable economic benefit → sufficient deployment scale → financial value → investment impact

A weakness anywhere in that chain should affect investment conviction.

This is the distinction between AI opportunity and underwritable AI value.

From AI Opportunity to Underwritable Value

This distinction becomes particularly important when AI forms part of a forward-looking value-creation plan.

Management may identify substantial potential savings across a business. An investor may expect AI to improve margins following an acquisition. A portfolio company may demonstrate a successful pilot that could theoretically be expanded across dozens of locations, business units or customer groups.

Each may represent genuine opportunity.

But opportunity and underwritable value are not equivalent.

An operator can reasonably look at a successful process improvement and ask how it can be replicated.

An investor has to ask an additional question: how much of that replication should enter the investment case before it has occurred?

That difference matters.

A 20% productivity improvement in one workflow may be operationally significant. But whether it contributes meaningfully to EBITDA depends on what happens to the capacity released, the proportion of the cost base affected, implementation and recurring costs, adoption across the organisation and whether the improvement can be sustained.

Similarly, a reduction in customer-service handling time may improve service economics. But the investment consequence depends on whether the company can reduce cost, serve more customers with the same resources, improve retention or generate some other measurable financial outcome.

The operating metric is therefore the beginning of the analysis, not the conclusion.

Investment underwriting requires a higher evidence threshold.

An investor needs to understand what has already been achieved, what remains an identified opportunity, what depends on future execution and how sensitive the investment case is to those assumptions.

This is the basis of commercial underwriting of AI-enabled value creation.

Rather than asking how many AI initiatives exist or how sophisticated the underlying technology appears, commercial underwriting asks whether the expected value can reasonably enter the investment case.

AI opportunity is not the same as underwritable AI value.

The Evidence Threshold for AI Is an Investment Question

The growing body of evidence around AI produces apparently contradictory conclusions.

AI can reduce costs, improve productivity, accelerate revenue generation and create meaningful financial value.

At the same time, many initiatives remain difficult to scale, their financial contribution remains difficult to attribute and productivity improvements do not always translate directly into earnings.

Both can be true.

The distinction depends on the evidence supporting the individual investment case.

This is why the appropriate question for an investor is not whether research proves that AI creates value in general.

It is whether the evidence available for a particular company supports the AI-enabled value assumed within its investment thesis.

A demonstrated 15% reduction in a defined operating cost should carry a different evidentiary weight from an identified opportunity to reduce that cost.

A pilot showing productivity improvement should carry a different weight from an application replicated across the organisation with measurable financial attribution.

An announced deployment should carry a different weight from sustained adoption.

Furthermore, an AI capability available equally to competitors should carry a different valuation implication from one strengthened by proprietary data, embedded workflows or operating advantages that are difficult to replicate.

Commercial underwriting therefore requires investors to move beyond a binary assessment of whether AI “works”.

The relevant task is to determine what has been demonstrated, what remains assumed, how economically material the distinction is and how much conviction the evidence deserves.

That is the analytical purpose of the Cohres AI Value Creation Assessment.

The Cohres AI Value Creation Assessment

AVCA evaluates AI-enabled value creation through six interconnected dimensions:

AI Investment Thesis
Commercial Credibility
Economic Attractiveness
Execution Feasibility
Competitive Durability
Investment Risk & Sensitivity

Together, these dimensions move the analysis from identifying an AI claim to determining its relevance to investment conviction.

1. AI Investment Thesis

The first question is not whether the company uses AI.

It is how AI enters the investment thesis.

AI may be central to expected returns, support particular elements of a value-creation plan or represent incremental upside that is not required for the underlying investment to succeed.

This distinction determines the appropriate level of diligence.

The assessment identifies the principal AI-enabled value-creation claims and determines their relationship to expected revenue growth, productivity, margins, operating leverage, valuation and ultimately investment returns.

Particular attention should be paid to dependency.

If a substantial portion of expected margin expansion requires successful AI deployment, evidence supporting that deployment should receive materially greater scrutiny than an initiative representing optional upside.

The operating question may be whether a particular AI initiative can improve a process.

The investment question is whether the improvement is sufficiently material, scalable and evidenced to influence expected returns.

This distinction prevents technological possibility from becoming an investment assumption without an appropriate evidentiary bridge.

The most important test is therefore not whether AI presents an opportunity, but how dependent the investment case is on that opportunity being realised.

That leads to a critical downside question: Does the underlying investment remain attractive under conservative AI assumptions?

2. Commercial Credibility

Once the AI investment thesis is understood, the next question is whether the underlying claims are supported by credible evidence.

Management presentations can describe potential applications.

Pilots can demonstrate technical capability.

Deployment announcements can indicate strategic intent.

None independently establishes economic value.

Commercial credibility therefore examines the progression from claimed opportunity to demonstrated operating evidence.

This includes evidence of actual deployment, user or customer adoption, realised operating improvements, repeatability across locations or customer groups and consistency between management’s narrative and observable results.

A useful discipline is: Installed ≠ adopted ≠ economically successful.

Each represents a different evidentiary threshold.

This matters particularly when management extrapolates from a small number of successful deployments.

A pilot may demonstrate that a process can be improved under controlled conditions. It does not necessarily demonstrate that the same improvement will survive different workflows, operating environments, employee behaviours, customer groups or data conditions.

S&P Global’s finding that organisations scrapped an average 46% of AI projects between proof of concept and broad adoption illustrates why this distinction matters.[7]

For an investor, the implication is not that pilots should be disregarded.

It is that the proportion of expected value underwritten should reflect the maturity of the evidence.

Where evidence remains limited, the appropriate conclusion is not necessarily that the initiative will fail.

It is that the expected value has not yet earned sufficient conviction to be treated as part of the base investment case.

3. Economic Attractiveness

Commercially credible AI applications must then translate into financially meaningful outcomes.

Depending on the business, those outcomes may include increased revenue, improved conversion, higher retention, better pricing, lower labour requirements, reduced energy consumption, lower maintenance costs, increased throughput, improved asset utilisation or higher margins.

The relevant operating metric will differ across sectors.

The analytical principle does not.

Gross theoretical opportunity should not be confused with economic value captured by the business.

This is where an operator-informed investment perspective becomes particularly important.

If an AI application reduces a task from 10 hours to 6, the operational improvement is measurable.

Although, an investor still needs to determine what happens to the 4 hours released.

If the company can eliminate an external cost, avoid incremental hiring, increase throughput, serve additional customers or redirect scarce skilled labour toward higher-value activities, the productivity improvement may translate into economic value.

If the released capacity remains unused and the cost base does not change, the financial value may be substantially lower than the productivity statistic implies.

The same discipline applies to revenue.

Higher customer engagement is not revenue.

Improved lead qualification is not revenue.

Faster product development is not revenue.

Each may create a credible pathway to revenue, but the investor needs evidence connecting the operational improvement to the financial outcome being underwritten.

Apollo’s disclosed portfolio examples demonstrate that this bridge can be established. At Cengage, selected AI applications reportedly reduced content-production costs by 40%, lead-generation costs by 15–20% and customer-care costs by 15%. At Barnes Group, an AI-supported service application reportedly produced a five-times return on investment in its first year.[3]

Those are materially different evidence points from an identified opportunity or an announced deployment.

Implementation costs, recurring technology expenditure, organisational requirements, capital investment, adoption constraints and deployment limitations must also be incorporated.

The central question is therefore: How much economic value can realistically be captured after the costs and constraints required to produce it?

Where evidence permits, the assessment should connect operating improvements to financial consequences through measures such as payback, ROI, free cash flow, margins and ROIC.

Precision should not be manufactured where evidence is insufficient.

An apparently sophisticated financial estimate built on unverified operating assumptions does not increase investment conviction.

4. Execution Feasibility

An economically attractive use case still creates little investment value if it cannot be implemented repeatedly at the scale assumed by the investment thesis.

Execution feasibility therefore tests whether management can convert AI deployment into financial value across the relevant organisation.

Factors such as management capability, operational readiness, data reliability, technology dependencies, adoption, implementation complexity, deployment history, external dependencies and accountability for benefit realisation can all affect value capture.

The purpose, however, is not to perform an AI maturity or readiness assessment.

The analysis remains investment-oriented.

For example, fragmented data matters not because an investor requires a technical diagnosis of the company’s data architecture.

It matters because inconsistent or inaccessible data may prevent a successful application from being replicated across the business, reducing the proportion of expected savings that can reasonably enter the investment case.

Limited employee adoption matters for the same reason.

The issue is not whether management should introduce another change-management programme. The issue is whether the demonstrated economics of a pilot can reasonably be extrapolated across the wider organisation.

Human oversight also matters where it limits the degree to which apparent automation translates into cost removal.

If an AI-enabled workflow still requires substantial review because errors carry financial, regulatory or customer consequences, the gross productivity opportunity should be adjusted for the continuing human resource requirement.

Execution feasibility therefore connects operating reality with investment consequence.

The relevant questions are not simply:

Can this be deployed?

Can the technology work?

Instead: Can the business repeatedly convert deployment into measurable financial value at the scale and within the timeframe assumed by the investment thesis?

Operational constraints become investment risks when they affect value capture, replication, margins, cash flow, ROIC or the timing of returns.

5. Competitive Durability

AI-enabled value creation must also be considered relative to competitors.

An initiative may produce genuine economic benefits while creating little sustainable competitive advantage if competitors can access similar technology and replicate the same improvements.

This distinction is becoming increasingly important as AI capabilities diffuse across industries.

Stanford’s 2025 AI Index reported that the inference cost associated with GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024, while performance gaps between leading closed and open-weight models narrowed substantially.[9]

The investment implication is significant.

As capable models become cheaper and more accessible, access to AI itself becomes a weaker basis for assuming durable differentiation.

The assessment therefore considers what surrounds the technology.

Proprietary data, embedded workflows, operating knowledge, distribution, customer relationships, switching costs, regulatory competence and the ability to repeatedly improve a process may create more durable economics than the underlying model.

Thomson Reuters provides an instructive example. Its AI products combine third-party frontier models with proprietary professional content, expert validation and integration into legal, tax and compliance workflows. The company reported that AI-enabled products represented 28% of annualised contract value by the fourth quarter of 2025.[10]

The important investment characteristic is not simply that Thomson Reuters uses AI.

It is that AI operates alongside assets and workflows accumulated outside the model itself.

This leads to the central competitive question: Does AI allow the company to earn superior economics, or does it merely prevent the company from falling behind?

The answer has important valuation implications.

An efficiency improvement that becomes industry-wide may protect margins without supporting structurally higher returns.

Conversely, AI combined with difficult-to-replicate data, workflows, distribution or operating knowledge may strengthen existing advantages.

AI itself should therefore never automatically be treated as a moat.

6. Investment Risk & Sensitivity

The final dimension asks what happens when the expected value does not materialise as planned.

AI-related investment assumptions can fail in several ways.

Deployment may take longer than expected.

Adoption may remain limited.

Savings may prove smaller.

Implementation costs may increase.

Human oversight requirements may constrain expected labour efficiencies.

Competitors may replicate the same capabilities.

Benefits may prove temporary rather than durable.

The investment case should therefore be tested against conservative scenarios.

Where AI contributes materially to forecast growth, margins, cash flow or valuation, sensitivity analysis should examine the consequences of weaker value capture.

The Klarna experience illustrates why operating metrics should be interpreted carefully.

The company reported substantial automation within customer service, including significantly shorter resolution times and work capacity equivalent to hundreds of employees. Management subsequently acknowledged that the company may have gone too far in pursuing AI-led cost reduction and shifted emphasis back toward service and product improvement.[11]

The lesson is not that the original productivity improvements were unreal.

It is that operational capacity and sustainable economic value are not necessarily equivalent.

Quality requirements, customer experience, the complexity of remaining work and the continuing need for human capacity can all affect how much theoretical productivity can actually become recurring financial benefit.

This is precisely why investment sensitivity matters.

The most important question remains: Does the underlying investment remain attractive under conservative AI assumptions?

If investment returns depend heavily on successful future AI execution that has not yet been demonstrated, the investor is underwriting execution risk rather than established economic value.

That distinction should be explicit.

Evidence Must Be Classified Before It Is Underwritten

One of the greatest risks in assessing AI value creation is treating fundamentally different forms of evidence as though they are equivalent.

Cohres distinguishes between four categories.

Realised Benefits

Credible evidence that operating or financial improvements have actually been achieved.

Examples may include demonstrated cost reductions, realised revenue contribution, measurable margin improvement, verified productivity converted into financial benefit or established project-level returns.

Identified Value Opportunities

Potential areas of value creation that have been identified but not yet demonstrated or fully quantified.

These may be commercially credible while remaining dependent on future execution.

Expected Annual Potential

Estimates of financial benefits that could occur if required implementation, adoption and deployment assumptions are achieved.

These figures can be useful for understanding potential materiality, but they remain forecasts rather than realised economics.

Deployment Commitments

Announced plans, contracts, budgets, rollouts or implementation targets that demonstrate intent.

They do not demonstrate adoption or economic value.

These categories should not be collapsed into a single headline number.

A €50M identified annual opportunity is not equivalent to €50M of realised recurring savings.

Nor should an announced deployment be treated as evidence of adoption, or adoption as evidence of financial contribution.

This classification creates an important bridge between operating evidence and investment underwriting.

The investor can acknowledge the full potential value pool without placing equal conviction in every component of it.

From Evidence to Investment Conviction

AVCA is designed to produce an underwriting conclusion rather than a list of AI initiatives.

Conviction should increase where there are:

realised financial benefits
measurable operating improvements
demonstrated adoption
repeatability
deployment beyond pilots
credible project economics
successful replication
clear financial attribution
conservative assumptions

Conviction should decrease where:

management claims remain unsupported
opportunities are presented as realised benefits
pilots are extrapolated to scaled deployment
deployment is treated as adoption
implementation costs remain unclear
benefit attribution is weak
expected value depends heavily on future execution
aggressive extrapolation is required
the financial contribution remains uncertain

This does not require artificial certainty.

Investment underwriting frequently operates with incomplete information.

The objective is to distinguish what the available evidence supports from what the investor is still being asked to believe.

Where evidence is incomplete, investment conviction should remain incomplete.

Screen the Claim → Test the Claim → Underwrite the Claim

Not every investment requires the same depth of AI analysis.

AVCA is therefore modular and operates within Cohres’ existing investment research process.

The progression is: Screen the claim → Test the claim → Underwrite the claim

Focused Screening

At the Focused Screening stage, AVCA should normally remain concise.

The central question is: Is AI material to the investment case, and are the underlying value-creation claims sufficiently credible to justify further diligence?

The objective is not to prove the entire AI value-creation case.

It is to identify materiality, understand the principal claims, distinguish early evidence from narrative and identify issues that could materially strengthen or weaken investment conviction.

Where AI is immaterial, further AVCA work may not be required.

Where it is material but poorly evidenced, that itself becomes a diligence finding.

Expanded Evaluation

At the Expanded Evaluation stage, the evidentiary burden increases.

The central question becomes: Which AI-enabled value-creation assumptions appear credible, how economically meaningful could they be, and what remains unverified?

The analysis begins connecting use cases to operating improvements and operating improvements to potential financial consequences.

The distinction between realised benefit and expected potential becomes increasingly important.

At this stage, Cohres should form preliminary conviction while remaining explicit about what is known, what can reasonably be inferred and what still requires verification.

Commercial Due Diligence

Where AI is sufficiently material to the investment thesis, AVCA can form a specialist module within broader Commercial Due Diligence.

The question becomes: Which AI-enabled value-creation assumptions can be underwritten, which should remain potential upside, and which should be rejected or subjected to further verification before capital is committed?

The evidentiary standard is highest here because the conclusions may influence forecasts, valuation, downside analysis and ultimately the investment decision.

This is where operational observations must become investment conclusions.

A scaling constraint is relevant because it affects the value that can enter the forecast.

A data limitation is relevant because it affects replication.

Limited adoption is relevant because it affects the probability that pilot economics become enterprise economics.

Implementation cost is relevant because it affects net financial benefit and return on invested capital.

The purpose is not to create an implementation roadmap.

It is to determine what the investor can reasonably underwrite.

AVCA and Commercial Due Diligence

AVCA complements rather than replaces Commercial Due Diligence.

The distinction is straightforward: Commercial Due Diligence evaluates the investment. AVCA tests the AI assumptions inside the investment thesis.

Traditional commercial diligence may determine whether a market is attractive, whether demand is sustainable, how the company is positioned competitively and whether management’s commercial forecasts are credible.

Where AI materially affects those forecasts or the post-investment value-creation case, an additional underwriting question emerges: How much confidence should the investor place in the value AI is expected to create?

AVCA provides a structured method for answering that question while keeping the analysis anchored to the investment decision.

Many investments will not require an AVCA.

It becomes relevant where AI materially supports expected growth, productivity, margin expansion, cost improvement, operating leverage, scalability or another important component of the investment thesis.

Applying AVCA: Heidelberg Materials

Cohres’ first public application of the AI Value Creation Assessment examines Heidelberg Materials, demonstrating how the methodology can be applied to a traditional industrial company rather than an AI-native software business.

This distinction is deliberate.

In an industrial company, AI value may emerge through process optimisation, predictive maintenance, energy efficiency, asset utilisation, quality improvement or other productivity gains rather than through the sale of an AI product.

The investment logic nevertheless remains the same.

The assessment considers Heidelberg Materials across all six AVCA dimensions:

AI Investment Thesis
Commercial Credibility
Economic Attractiveness
Execution Feasibility
Competitive Durability
Investment Risk & Sensitivity

It also distinguishes between realised benefits, identified value opportunities, expected annual potential and deployment commitments.

The purpose is not to prescribe how Heidelberg Materials should implement AI.

It is to determine how an investor should interpret the evidence surrounding AI-enabled value creation and how much of that expected value deserves investment conviction.

The full public assessment is available here: Cohres AI Value Creation Assessment — Heidelberg Materials
https://cohres.com/wp-content/uploads/2026/08/AI-Value-Creation-Assessment-Heidelberg-Materials.pdf

The public assessment is intentionally more extensive than the AVCA module expected within a typical Commercial Due Diligence engagement because it serves as the first public demonstration of the methodology.

The Evidence Threshold Is Changing

AI will increasingly appear within investment theses even where the underlying company is not an AI business.

That creates both opportunity and underwriting risk.

Investors who ignore AI-enabled value creation may underestimate genuine sources of productivity, growth and operating leverage.

Investors who accept AI narratives without sufficient evidence may capitalise benefits that remain dependent on unproven execution.

The appropriate response is neither automatic skepticism nor automatic optimism.

It is underwriting discipline.

The presence of AI does not establish investment value.

A credible use case does not establish financial materiality.

A successful pilot does not establish scalability.

Deployment does not establish adoption.

Productivity does not automatically establish financial benefit.

And identified opportunity does not establish realised economic value.

As AI becomes increasingly embedded within assumptions about growth, productivity, margins and operating leverage, investors will need to distinguish between AI opportunity and underwritable AI value.

That is the purpose of the Cohres AI Value Creation Assessment.

The question is not whether AI can create value. It is how much of that value an investor should be willing to underwrite before capital is committed.

References

[1] McKinsey & Company. The State of AI: How Organizations Are Rewiring to Capture Value / 2025 State of AI research.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[2] IBM Institute for Business Value. 2025 CEO Study, 2025.
https://www.ibm.com/thought-leadership/institute-business-value/en-us/blog/2025-ceo-blog-article

[3] Apollo Global Management. Building AI Capabilities into Portfolio Companies at Apollo, June 2025.
https://www.apollo.com/insights-news/news/2025/06/building-ai-capabilities-into-portfolio-companies-at-apollo

[4] Hg. Hg Launches Catalyst, a New AI Incubator, 2026.
https://hgcapital.com/insights/hg-launches-catalyst-a-new-ai-incubator

[5] Deloitte. 2025 GenAI in M&A Survey, October 2025.
https://www.deloitte.com/us/en/about/press-room/deloitte-survey-genai-in-mna.html

[6] Blackstone. AI at Scale: A Conversation with Blackstone’s CTO and Global Head of the Operating Team, May 2026.
https://www.blackstone.com/insights/article/ai-at-scale-a-conversation-with-blackstones-cto-and-global-head-of-the-operating-team/

[7] S&P Global Market Intelligence. AI Experiences Rapid Adoption but with Mixed Outcomes, 2025.
https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning

[8] National Bureau of Economic Research. Generative AI at Work / field research involving 7,137 knowledge workers, 2025.
https://www.nber.org/papers/w33795

[9] Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2025.
https://hai.stanford.edu/ai-index/2025-ai-index-report

[10] Thomson Reuters. Company disclosures concerning CoCounsel adoption and FY2025 results, 2026.
https://ir.thomsonreuters.com/news-releases/news-release-details/one-million-professionals-turn-cocounsel-thomson-reuters-scales

[11] Reuters. Sweden’s Klarna shifts AI focus from cost cuts to growth, September 2025.
https://www.reuters.com/business/swedens-klarna-shifts-ai-focus-cost-cuts-growth-2025-09-10/