Enterprise AI in the S&P 500: Measurable Gains, Incomplete ROI Proof

Review period: January 1, 2023–September 19, 2026 Scope: Publicly disclosed AI operating benefits, monetization and investment costs at selected S&P 500 companies, with an index-wide disclosure assessment. Source policy: Company filings, earnings materials and original statements prioritized; institutional research and high-quality reporting used for context and verification.

Executive summary

S&P 500 companies are generating measurable benefits from AI, but public evidence of fully costed financial returns remains much thinner than evidence of deployment, productivity or revenue. Amazon, JPMorgan Chase, IBM and C.H. Robinson provide meaningful examples of quantified savings or operating improvements. However, their disclosures differ substantially in attribution, measurement and treatment of investment costs. (aws.amazon.com)

Torsten Slok’s central concern is therefore justified: investors usually cannot reconstruct the relationship between an enterprise’s AI spending and its incremental cash earnings. But the stronger interpretation—AI is producing no measurable economic value—is not supported. Moreover, AI suppliers disclose substantial monetization, while Meta provides evidence that AI can improve an existing revenue engine rather than merely reduce costs. (apollo.com)

The strongest current evidence concerns bounded, repeatable workflows and established digital businesses. The weakest link is the investment denominator: implementation costs, inference charges, human oversight, depreciation and organizational change are seldom matched to the benefits being advertised. The appropriate conclusion is selective demonstrated value—not yet broadly demonstrated, independently reconstructable enterprise AI ROI. (aws.amazon.com)

1. What Slok and Apollo actually claimed

In his September 11, 2026 Daily Spark, Slok reported the following findings from the S&P 500’s second-quarter reporting season:

Among quantified proof points, 70% concerned costs and 22% revenue. These are disclosure classifications drawn from The AI Value Gap, not independently audited estimates of companies’ economic returns. (apollo.com)

Assessment: persuasive diagnosis, but not proof that returns are absent

Three qualifications matter:

1. A quantified result is not necessarily a financial return. It may measure processing speed, employee hours, adoption or customer behavior.
2. Undisclosed returns are not zero returns. The underlying project explicitly characterizes its scores as measures of disclosed evidence. Its historical observations also have different verification treatment from the latest quarter. (thevaluegap.io)
3. “No AI KPI” is too broad if applied literally to all AI-related disclosure. Accenture reports identified generative-AI revenue, and Salesforce reports Agentforce annual recurring revenue. These are principally supplier monetization metrics, however—not proof that enterprise customers earn an adequate return from using those products. (accenture.com)

Independent institutional work supports the narrower concern. J.P. Morgan Asset Management’s September 2026 review found that 65% of S&P 500 companies mentioned AI on second-quarter calls, but only 2% quantified AI-productivity effects on profits. That 2% is not the same measure as Apollo’s 2% tracking a metric over time. (am.jpmorgan.com)

There is also a more optimistic reading. Morgan Stanley reported in March 2026 that 21% of S&P 500 companies cited at least one AI benefit and that identified adopters exhibited cash-flow-margin expansion roughly twice the global average. This is encouraging observational evidence, but it does not isolate AI from business mix, management quality or other investments. (morganstanley.com)

Methodological limitation: This report investigates selected public cases rather than independently recoding all 500 companies. It does not estimate the percentage of the index earning positive AI ROI.

2. An evidence hierarchy for evaluating AI claims

The following is this report’s assessment framework—not an accounting standard or a reproduction of Apollo’s scoring system.

For this investigation, a convincing ROI disclosure would identify the baseline or control, incremental cash benefits, upfront investment, recurring operating costs, useful life and quality constraints. Hours saved become cash savings only when they reduce expenditure, avoid necessary hiring or generate additional profitable output. That distinction is central to the assessments below.

3. Concrete enterprise-adoption cases

The figures below are company-reported unless otherwise indicated. “Not disclosed” means absent from the cited evidence—not necessarily unavailable internally.

A. Quantified operating and financial benefits

B. Which of these approach financial ROI?

JPMorgan offers the closest simple comparison—but break-even is not a 100% net return. Using the approximately $2 billion annual benefit and expense figures, the implied current benefit-to-cost ratio is about 1:1. The annual net benefit would therefore be approximately zero under a simplified calculation. Historical investment, future benefits and the precise scope of each figure remain unresolved. This is evidence of expense coverage, not a complete lifetime-return calculation. (semafor.com)

C.H. Robinson provides unusually useful operating corroboration. Successive productivity disclosures, staffing data and financial results are more informative than a single pilot announcement. Yet its own investor presentation describes a combined program of waste removal, Lean processes and custom AI. Assigning all operating-income growth to AI would overstate the evidence. (s25.q4cdn.com)

IBM’s large savings claim should not be discarded—but should not be relabeled. It establishes the scale of its transformation program, not AI’s independent contribution. Its Q2 2026 results continued to discuss AI and automation-led productivity, while also showing that consolidated financial outcomes reflect other revenue and expense pressures. (ibm.com)

Walmart illustrates the gap between commercial promise and material earnings attribution. At its April 2026 investor conference, its CFO reiterated the larger Sparky basket and described emerging benefits, but cautioned that they were unlikely yet to be noticeable to investors against a revenue base of approximately $700 billion. (corporate.walmart.com)

C. An important boundary case: Deere’s customers versus Deere’s shareholders

Deere reported that See & Spray operated across more than five million acres in 2025, reducing non-residual herbicide use by nearly 50% and saving nearly 31 million gallons of herbicide mix. Its 2024 disclosure reported more than one million acres and approximately eight million gallons saved. These are repeated, physical-world operating measurements—not simply an AI deployment announcement. (deere.ca)

However, the immediate savings accrue to farmers using Deere’s equipment. They do not establish Deere’s incremental profit after technology development, equipment production and commercial costs. Nor can the change between seasons be treated as a controlled improvement in AI effectiveness: acreage and operating conditions differ. The case supports customer value, not a disclosed Deere AI ROIC. (deere.ca)

D. Where evidence remains at deployment or expectation

HCA Healthcare’s July 2026 strategy publication describes clinical-documentation pilots and a nurse-handoff initiative addressing a workflow occurring 24 million times annually. That is a measure of the opportunity’s scale, not hours or dollars already saved. The publication does not supply a corresponding enterprise earnings bridge. (hcahealthcaretoday.com)

Similarly, 3M’s July 2026 Microsoft partnership outlines intended improvements in decision-making, customer experience and productivity. The announcement itself is evidence of a strategic deployment commitment, not realized financial ROI. Neither example establishes that benefits are absent—only that the cited disclosure does not demonstrate them. (investors.3m.com)

4. Is AI producing new revenue—or mainly reducing costs?

Enterprise adopters: cost and productivity dominate the disclosed evidence

The best quantified adopter cases examined here generally concern existing cost structures: software maintenance, help desks, freight administration and back-office work. Apollo’s finding that cost-related observations represent 70% of disclosed proof points is consistent with this case evidence. It does not mean 70% of AI’s total economic value is cost savings; disclosure frequency is not a dollar-weighted benefit measure. (aws.amazon.com)

Analysis: Cost effects are relatively straightforward to observe when work has a clear baseline—minutes per task, transactions per employee or infrastructure expense. Revenue attribution is harder when AI changes customer behavior alongside pricing, promotions, distribution and product improvements.

Meta: an important revenue-side counterexample

On its Q2 2026 call, Meta attributed an 8.3% increase in ad clicks and a 15.7% uplift in Facebook conversions to advances in its advertising models. This is evidence that AI can improve monetization within an established business rather than simply eliminate labor. The metrics nevertheless do not disclose the incremental profit produced by each additional dollar of AI investment. (s21.q4cdn.com)

The distinction matters because Meta’s Q2 revenue increased 28%, while operating income declined 8%. Legal charges, severance and the wider expense base also affected results; neither the revenue growth nor the profit decline can be attributed exclusively to AI. (investor.atmeta.com)

AI suppliers: substantial monetization, but a different ROI question

Bottom line: AI suppliers already have measurable revenue. That rebuts a claim of no monetization, but does not answer whether their customers’ savings or incremental profits justify what they spend. Supplier revenue is simultaneously part of the customer’s AI cost base.

5. Are the gains large enough to justify the investment?

A. Infrastructure providers: the investment hurdle is very large

J.P. Morgan Asset Management reported that consensus estimates for 2026 AI capital expenditure had risen from approximately $650 billion at the beginning of the year to $800 billion by its September review. This is a forecast, not completed spending, and should not be confused with a measured AI investment total for every S&P 500 adopter. (am.jpmorgan.com)

Selected disclosures illustrate both scale and comparability problems:

The near-term cash-flow consequences are visible. Amazon reported negative $7.6 billion of trailing-twelve-month free cash flow, despite $161.4 billion in operating cash flow, with the investment increase primarily reflecting AI. Meta’s Q2 free cash flow was approximately $784 million. These figures demonstrate the cash burden of investment, not necessarily that the investments have negative lifetime value. (ir.aboutamazon.com)

Microsoft also reported a 65% Microsoft Cloud gross margin in Q4, down year over year, partly because of AI infrastructure investment and increased usage. Strong sales and utilization therefore need to be evaluated alongside servicing costs—not just seat counts. (microsoft.com)

One further accounting trap: Amazon’s Q2 net income included $53.4 billion of non-operating pretax other income, primarily from its Anthropic investments. That should not be mistaken for cash operating returns from enterprise AI deployment. (ir.aboutamazon.com)

B. Ordinary adopters: smaller capital requirements, opaque operating costs

Enterprise adopters need not replicate a hyperscaler’s data-center buildout. Nevertheless, the available spending figures often cover much more than AI:

• JPMorgan: Its February 2026 company update projected $19.8 billion in annual technology spending and identified about $600 million in efficiencies, only some AI-related. Neither amount should be relabeled as exclusively AI spending or savings. (jpmorganchase.com)
• Bank of America: Its April 2025 disclosure cited $13 billion in annual technology spending, including approximately $4 billion for new technology initiatives in 2025. It did not isolate the investment supporting its help-desk and coding results. (newsroom.bankofamerica.com)
• C.H. Robinson: Q2 2026 capex was $18.2 million, with $65–75 million expected for the year. Those are total-company amounts and exclude the information needed to reconstruct AI-related operating expenses. Dividing operating-income improvement by capex would therefore produce a misleading “AI return.” (s25.q4cdn.com)
• Accenture: FY2025 investment of $3.3 billion covered acquisitions, research and development, and employee learning. It is not an AI-only cost figure that can simply be compared with its $2.7 billion in identified AI revenue. (accenture.com)

Analysis: For adopters, the missing denominator is often operating expenditure rather than conspicuous capex. A complete project assessment should include:

• Software subscriptions and inference/API consumption;
• Integration, data preparation and legacy-system changes;
• Internal engineering and specialist personnel;
• Testing, security, monitoring and human review;
• Training, workflow redesign and transition costs;
• Maintenance investment and the cost of unsuccessful experiments.

These are items investors should request, not an assertion that every company incurs each cost in the same proportion.

C. A defensible investment test

For a bounded deployment, this report recommends evaluating:

Incremental cash benefit = cash costs avoided + incremental contribution profit − incremental AI operating costs.

Then compare those cash benefits over time with upfront implementation costs and subsequent capital requirements, discounted at an appropriate hurdle rate. Revenue should enter at its incremental contribution margin—not dollar for dollar. Labor capacity that is merely freed should be distinguished from realized expenditure reduction.

Applied to the disclosed cases:

• JPMorgan: Its reported annual comparison suggests expense coverage, but not an independently reconstructable lifetime return. (semafor.com)
• Amazon’s modernization: The disclosed dollar benefit is meaningful, but without the project’s cost and allocation methodology, an exact ROI or payback period cannot be verified. (aws.amazon.com)
• IBM: Aggregate productivity savings are material; AI’s share of both benefits and costs remains unresolved. (ibm.com)
• Hyperscalers: Monetization is substantial, but the reviewed disclosures do not isolate sufficient AI-specific profit and invested capital to establish that the entire buildout earns an adequate return. (microsoft.com)

No case in this review supplies every ingredient needed to independently reconstruct company-wide AI ROIC. That is a disclosure conclusion, not a finding that all projects are unprofitable.

6. Cross-sector assessment

Assessment: The leading distinction is less “technology versus non-technology” than measurable, repeatable workflow versus open-ended transformation. C.H. Robinson is a notable non-technology example; broad AI strategies at much larger businesses may still be difficult to evaluate financially.

7. What investors should watch over the next 12–24 months

Analysis horizon: September 2026–September 2028. This is separate from the historical review period.

The following is an investor monitoring framework, not a forecast that each metric will improve.

Two immediate disclosure warnings deserve special attention: Salesforce changed the scope of Agentforce ARR, while Microsoft’s lease-classification change affected reported capex expectations without an equivalent change in investment plans. Investors need comparable economic measures, not merely improving headline indicators. (salesforce.com)

Conclusion

Slok is substantially right about the proof gap, but that gap should not be confused with an absence of economic benefits. Public company evidence already supports real task-level efficiency, material management-estimated savings and growing supplier revenue. The strongest adopter evidence is concentrated in specific software, service and logistics workflows; established advertising businesses also provide credible revenue-side evidence. (apollo.com)

What remains weak is the public reconciliation from these benefits to incremental, fully costed cash returns. Spending totals are often too broad, benefits are frequently counterfactual or bundled with wider transformation, and supplier revenue is sometimes mistaken for customer success. (jpmorganchase.com)

The most defensible current judgment is therefore:

• Measurable AI value: established in selected companies and workflows.
• Persistent operational improvement: increasingly supported, but unevenly measured.
• Positive, fully attributable enterprise-wide financial ROI: still rarely reconstructable from public disclosure.
• Justification of the entire infrastructure investment cycle: not yet settled by the available revenue and cash-flow evidence. (am.jpmorgan.com)

Selected sources

Original claims and institutional context

• Apollo / Torsten Slok: September 11, 2026 Daily Spark and its underlying disclosure framework. (apollo.com)
• J.P. Morgan Asset Management; Morgan Stanley: Index-level disclosure and AI-adopter research. (am.jpmorgan.com)

Primary company evidence

• JPMorgan Chase: Original executive interview, 2026 company-update transcript. (bloomberg.com)
• IBM: Q4 2025 prepared remarks and Q2 2026 financial results. (ibm.com)
• Amazon: Internal modernization case and Q2 2026 earnings release. (aws.amazon.com)
• C.H. Robinson: Productivity disclosures and Q2 2026 earnings materials. (chrobinson.com)
• Bank of America; Walmart: Employee-AI disclosure and company-hosted earnings transcripts. (newsroom.bankofamerica.com)
• Microsoft; Meta: Latest quarterly earnings calls and financial disclosures reviewed. (microsoft.com)
• Accenture; Salesforce: AI revenue definitions and recurring-revenue disclosures. (accenture.com)
• Deere; HCA Healthcare: Physical operating outcomes and deployment strategy. (deere.ca)