Wharton Study: 3 in 4 Enterprises See Positive ROI from GenAI in 2025

In a striking snapshot of enterprise generative AI in late 2025, Wharton Human‑AI Research finds that roughly three in four companies report positive ROI from generative AI projects, and survey respondents name OpenAI’s ChatGPT as the most widely used tool — while Anthropic’s Claude registers far lower enterprise usage than many observers expected.

Background​

Wharton Human‑AI Research (WHAIR) has transitioned its questions to business leaders from curiosity-driven adoption to measured adoption — asking how often firms use generative AI, whether they track ROI, and which vendors and products are actually being put to work inside companies. The center’s multi‑year study, produced with GBK Collective, reports steep increases in routine use (weekly and daily) and shows that a growing share of companies now requires ROI measurement for GenAI investments. That shift — from pilots to line‑item investments — is the context for the headline findings that dominated recent coverage: nearly 72–75% of enterprise respondents report positive returns on their generative AI programs, and the distribution of tool adoption is concentrated among a small set of platforms with clear distribution advantages. These are the results that explain why ChatGPT tops usage charts, Microsoft Copilot shows outsized enterprise traction, Google’s Gemini converts more slowly in mixed environments, and Anthropic’s Claude sits lower than many expected in public‑facing usage metrics.

---

What the Wharton numbers actually say — and what they don’t​

Wharton’s reporting is important because it documents behavior inside real organizations rather than just vendor-reported counts. The principal, verifiable takeaways are:

• A large majority of surveyed enterprise leaders now use GenAI on a weekly basis, and many use it daily; Wharton’s recent releases cite weekly usage in the 70–80% range and daily use in the mid‑40s.
• Roughly three out of four firms that are tracking ROI report positive returns from generative AI projects — a signal that pilots are moving into production with measurable benefits.

Important caveats and limits to interpretation:

• The ROI figure is survey‑based and self‑reported; it reflects managers’ experience and program-level measurement, not audited company financial statements isolated to AI. Treat the 72–75% positive‑ROI headline as directionally meaningful, not as proof of universal profit from AI spending.
• Survey results depend on sampling frame (enterprise leaders, often large companies) and the specific ROI metrics used by respondents. Different sectors, company sizes, and maturity levels show substantial variance in measured outcomes.

Read this as signal, not gospel: enterprises increasingly extract real value from generative AI, but outcomes vary by use case, governance, and instrumented measurement.

---

The leaderboard: who’s winning and why​

ChatGPT — the generalist that refuses to be ignored​

OpenAI’s ChatGPT continues to lead in public usage and referral‑traffic metrics, often commanding a large majority share in third‑party telemetry. Multiple audience and web‑traffic trackers place ChatGPT well ahead of rivals for chatbot referrals and site traffic, a dynamic that translates into massive developer and third‑party ecosystem momentum. That scale creates a self‑reinforcing advantage: broader plugin ecosystems, more integrations, richer community assets, and easier access for employees experimenting with AI at work. Why enterprises lean on ChatGPT:

• Broad capability set — drafting, summarization, coding help, and multimodal features that cover many everyday knowledge‑work tasks.
• Mature developer ecosystem — APIs, plugins, and third‑party tools that make it simple to embed ChatGPT in internal workflows.
• Familiar pricing and admin tiers — consumer and enterprise offerings allow staged procurement and pilot-to‑scale moves.

Tradeoffs: public usage dominance does not replace a need for enterprise contractual protections (data residency, non‑training guarantees) and operational governance.

Microsoft Copilot — distribution beats novelty​

Microsoft’s Copilot family is showing disproportionately strong enterprise adoption relative to its raw public referral share because Copilot is embedded inside the productivity tools organizations already pay for — Word, Excel, Outlook, Teams, and Windows. Embedding AI inside the applications employees open every day drastically reduces adoption friction and makes Copilot a practical productivity multiplier in many Microsoft‑centric firms.
Why Copilot converts enterprise budgets more easily:

• Deep integration with Microsoft 365 and Windows, giving Copilot contextual access to tenant data (subject to tenant governance).
• Familiar procurement and licensing channels, which simplify buying and vendor management for IT teams.
• Enterprise governance controls (through Microsoft Graph, Purview, and tenant-level admin capabilities) that procurement and compliance teams can evaluate against regulatory needs.

Tradeoffs: Microsoft’s licensing complexity and SKU fragmentation can complicate cost forecasting for SMBs and cross‑platform deployments.

Google Gemini — technical strengths, slower enterprise conversion​

Google’s Gemini brings technical differentiators — multimodal reasoning, long‑context capabilities, and deep Google Workspace integration — but converting product strengths into cross‑platform enterprise habit has been slower than expected. Gemini performs best where organizations are already deeply invested in Google Workspace; in mixed environments or Windows‑first firms, Microsoft’s bundling into existing workflows has proven more immediately practical.

Anthropic Claude — a surprising laggard in public usage​

Anthropic’s Claude has been widely praised for safety design and strong long‑form reasoning, and it emphasizes enterprise features such as non‑training contracts. Yet public telemetry and many surveys show Claude considerably lower in visible usage compared with ChatGPT and Copilot. That gap has surprised some analysts given Anthropic’s safety‑focused posture and enterprise messaging. Two important nuances: Claude may have meaningful behind‑the‑firewall deployments that public traffic measures undercount, and enterprise adoption can be slower to appear in public telemetry because many contracts are private.

---

Why distribution and governance matter more than raw model architecture​

The pattern across corporate procurement is consistent: integration beats marginal model improvements when adoption is the goal.

• Embedding an assistant in the daily workflow (e.g., an inline Copilot in Excel or a ChatGPT bot tied to Slack) reduces friction and raises habitual usage.
• Governance and contractual clarity — especially non‑training guarantees, data residency, and SOC/ISO artifacts — matter heavily for regulated sectors; sometimes more than headline capability claims. Organizations buying AI for finance, health, or legal workflows require auditable controls.

Put differently: IT leaders buy tools they can secure, govern, and roll out at scale. Vendors that succeed in enterprise procurement combine technical capability with procurement‑friendly packaging and clear governance artifacts.

---

Verifying the big claims — what independent data shows​

To test the Wharton survey headlines and the Business Insider chart, I cross‑checked multiple data sources:

• Wharton’s own report and associated press materials document the high weekly usage and the statistic that roughly three in four firms tracking ROI report positive returns. These are primary survey outputs and underpin the coverage.
• Independent traffic and referral trackers (StatCounter summaries reported across industry press) place ChatGPT as the dominant chatbot referral source in many datasets, often above 70–80% in comparative snapshots. That public dominance aligns with the user‑facing visibility Wharton and other surveys capture.
• Market intelligence briefings and referral‑traffic trackers (weekly referral reports and industry roundups) show Copilot growing strongly and translating deep Microsoft distribution into measurable enterprise usage, consistent with the Wharton pattern.

Where numbers diverge or remain opaque:

• Absolute user counts and vendor‑reported “active user” figures are often inconsistent across press releases and third‑party trackers; for many vendors these counts are not independently auditable without internal metrics. Flag any precise vendor user totals as vendor‑provided unless corroborated by independent telemetry.

---

Why Anthropic’s Claude lags in public telemetry (and why that doesn’t automatically mean enterprise failure)​

Multiple factors help explain Claude’s lower public usage metrics:

• Distribution disadvantage: Claude lacks the same built‑in placement inside a dominant productivity suite that Microsoft has, and it has less of ChatGPT’s broad plugin and community footprint. Adoption often follows distribution, not just capability.
• Enterprise stealth deployments: Some enterprise deployments of safety‑oriented platforms are negotiated as private, behind‑the‑firewall contracts and therefore don’t appear in public traffic snapshots; public referral share can understate private contract adoption.
• Marketing and perception: Expectations about Anthropic’s wholesale enterprise advantage were high in parts of the market; when those expectations meet real‑world procurement realities — existing vendor relationships, security artifacts, and procurement friction — adoption can appear slower than tech observers expected.

Bottom line: Claude’s public traffic shortfall is a real signal about public usage momentum, but it is not definitive proof of enterprise irrelevance. Procurement cycles and private contracts often lag the public narrative.

---

The Windows‑centric IT perspective: what to watch and what to do​

For IT teams in Windows‑centric organizations — where Microsoft 365 and Windows remain the default productivity surface — these findings carry immediate operational implications.
Key operational takeaways:

• Prioritize tools that reduce friction: start with copilots that live in Office (Copilot for Microsoft 365, GitHub Copilot for developers) to maximize immediate productivity gains.
• Insist on contractual protections: require non‑training clauses or explicit data‑use guarantees for any enterprise tier that will process sensitive tenant data. Ask vendors for SOC 2 Type II, ISO 27001, and clear data residency options.
• Measure before you scale: define KPIs for pilot success (time saved per task, error reduction, throughput improvements), instrument them, and only expand when the metrics hold up. Treat AI rollouts as product programs with release cadences, training, and rollback playbooks.

Practical procurement checklist for Windows admins:

• Do you have a contractual non‑training guarantee and clearly documented data flows?
• What audit reports (SOC 2, ISO) and attestations does the vendor provide?
• How is tenant grounding implemented (tenant‑scoped models, private endpoints)?
• What admin controls and role‑based access are available in the management plane?
• What are the metering limits, rate limits, and cost predictability measures for heavy usage?
• Can the vendor provide references for customers in the same industry and at similar scale?

---

Risks and failure modes enterprises must manage​

Generative AI can deliver real gains, but it also introduces new operational risks:

• Hallucination and factual drift: models can produce confident but wrong outputs; mission‑critical workflows must keep a human in the loop and add verification layers.
• Data leakage and training usage: sending sensitive IP or personal data to third‑party models without contractual protections risks inadvertent training or disclosure. Enterprise contracts and private endpoints mitigate this.
• Concentration risk: heavy dependence on a single provider exposes operations to outages, policy changes, or sudden pricing shifts — maintain multi‑vendor resilience where appropriate.
• Skill atrophy: WHAIR finds leaders concerned about declines in employee proficiency when AI takes on routine tasks — invest in role redesign and training to preserve high‑value human skills.

---

Tactical recommendations for WindowsForum readers (IT managers, sysadmins, procurement)​

• Start with pilot programs tied to measurable KPIs and limited scope (one team, one workflow). Measure thoroughly for 4–8 weeks before scaling.
• Favor copilots and integrations that align with your productivity stack — Microsoft Copilot for heavy Office shops, Gemini where Google Workspace is dominant, ChatGPT where broad capability and integration matters. Don’t buy a product because it’s “best tech”; buy what integrates with your environment.
• Require contractual non‑training guarantees, describe the exact data flows you will send to the vendor, and insist on audit artifacts. If regulation or IP concerns are material, prefer private endpoints or on‑premise/air‑gapped architectures.
• Build multi‑vendor fallbacks for critical workflows to preserve continuity during a vendor outage or if a vendor’s terms change. Maintain scripts and lightweight adapters to switch endpoints if needed.
• Invest in governance and training: update acceptable use policies, add monitoring and anomaly detection on AI outputs, and train users in prompt design and verification. Make human review part of the quality control loop.

---

What we still don’t know — claims to treat with caution​

• Precise active‑user counts from vendors: these are often inconsistent across press statements and are rarely independently audited. Treat vendor‑supplied “monthly active user” numbers as directional unless confirmed by independent telemetry.
• The exact distribution of behind‑the‑firewall deployments for smaller vendors: private enterprise contracts can hide substantial adoption from public referral metrics, so low public traffic is not definitive proof of failure. Anthropic’s Claude is a case in point.

---

The competitive outlook — what the market structure implies​

• Distribution as moat: embedding AI inside existing productivity surfaces (Office, Workspace, IDEs) accelerates adoption faster than incremental model improvements alone. Expect hyperscalers and large SaaS vendors to weaponize distribution as a competitive advantage.
• Specialization will win niches: citation‑forward research tools, privacy‑first deployments, and industry‑specific copilots will continue to attract targeted enterprise budgets. Buyers will combine a generalist assistant (broad tasks) with specialists for high‑assurance or research tasks.
• Open vs. hosted tradeoffs: where sovereignty and IP control are critical, private or open‑weight deployments will gain traction; hosted hyperscaler solutions will remain dominant for rapid time‑to‑value.

---

Conclusion​

The Wharton snapshot captures a decisive moment: generative AI has moved from novelty to operational capability for many enterprises, and most organizations that track ROI are already seeing benefits. But the vendor marketplace is not a pure quality contest — distribution, governance, and procurement mechanics shape who wins inside companies.
OpenAI’s ChatGPT leads in visible usage and ecosystem reach, Microsoft’s Copilot reaps the rewards of integration into productivity workflows, Google’s Gemini brings technical strengths but faces conversion friction in mixed environments, and Anthropic’s Claude underperforms in public telemetry even as it maintains meaningful niche traction. IT leaders and Windows‑centric admins should choose tools by use case and governance posture, instrument outcomes with rigor, and treat AI rollouts as product programs that require training, measurement, and fallback planning. The winners will be the teams that measure outcomes, control the risks, and integrate AI into the daily rhythm of work without surrendering accountability.

---

Source: Business Insider Africa The AI tools used most by companies. There's a surprising winner and a shocking laggard.