Source files: Major consulting and vendor refresh raw ledger; McKinsey AI-economics ledger
Source status: Official pages and PDFs reviewed 2026-08-16; existing local notes reverified where cited.
Confidence: MEDIUM for the cross-firm themes; LOW-MEDIUM for headline ROI comparisons because populations, definitions, and commercial incentives differ.
Executive summary
The major consulting firms now agree on the broad direction: AI is moving from assistant use toward agentic workflows, and value depends more on operating-model redesign, data/integration, governance, and workforce change than on selecting a slightly better model.
They disagree on the intervention and on how quickly to move:
- McKinsey is the clearest on AI FinOps, demand management, cost attribution, and cost per task or outcome.
- BCG is the most transformation-forward: build an agentic transformation factory, a platform backbone, and end-to-end autonomous operations.
- Accenture frames the issue as enterprise reinvention: technology, work redesign, and workforce transformation must move together.
- Deloitte emphasizes the readiness gap, sovereign context, workforce fluency, and the fact that agents are scaling faster than governance.
- EY is the sharpest on token-cost scrutiny and the shift from packaged SaaS toward in-house AI-built software.
- KPMG emphasizes centralization, portfolio discipline, and the gap between perceived value and repeatable ROI.
- PwC focuses on industrial architecture and audit-ready control across cross-platform handoffs.
- IBM argues that infrastructure portability, governance by design, and portfolio discipline determine what the business can safely pursue.
- Capgemini combines microtransformations with a reusable control plane, FinOps, governance-as-code, and vendor independence.
- Roland Berger is the most blunt that the operating model—not the model—is the root cause of weak enterprise results.
Comparison by firm
| Firm | Primary diagnosis | Preferred move | Commercial posture to recognize |
|---|---|---|---|
| McKinsey | AI demand and agent cost are becoming unmanaged; value is lost when spend is not tied to workflows. | Permanent AI FinOps, routing, instrumentation, high-value workflow portfolio, mission owners. | Strongest fit for a CFO/CIO cost-and-value conversation; McKinsey also sells the transformation. |
| BCG | The enterprise needs an agentic operating system and transformation engine. | Choose a platform, create graduation paths, redesign end-to-end processes, run an agentic transformation office. | Most likely to sell a broad operating-model transformation and platform program. |
| Accenture | AI investment fails when technology is separated from work and workforce reinvention. | Rebuild processes and roles; align leadership, tech, and talent. | Strong delivery and workforce/reinvention orientation. |
| Deloitte | Adoption is outrunning readiness, especially governance, infrastructure, talent, and sovereignty. | Move from ambition to activation with readiness, guardrails, fluency, and location/sovereignty choices. | Survey/research plus large implementation, risk, and managed-services footprint. |
| EY | The financial case is being challenged by token costs and bespoke software economics. | Monitor tokens, add guardrails, prioritize different work, and build selected internal software. | Strong fit for CFO, risk, tax, and enterprise-software disruption conversations. |
| KPMG | Organizations confuse use-case value with portfolio ROI; fragmentation and tech debt block compounding. | Centralize investment, supplier selection, and talent; redesign the enterprise around workflows. | Governance, finance, and transformation lens; survey headline claims remain self-reported. |
| PwC | Platform controls do not see the full transaction chain across agents and systems. | Shared state, orchestration, end-to-end traceability, continuous monitoring, and cross-platform controls. | Strongest audit/control framing; architecture guidance is advisory rather than outcome proof. |
| IBM IBV | AI scale is constrained by inflexible architecture, manual governance, and poor financial visibility. | Portable infrastructure, governance by design, digital-twin/orchestration layer, portfolio capital allocation. | IBM can connect research to hybrid cloud, software, infrastructure, and consulting. |
| Capgemini | AI is spreading organically, creating cost, tool, and security sprawl. | Microtransformations first, then platformized scale, governance-as-code, FinOps, and modularity. | Implementation-heavy; the RAISE PDF is a strategic POV, not an independent ROI audit. |
| Roland Berger | Operating models are left untouched, so AI power does not translate to enterprise performance. | Start with the result, redesign roles/processes, and close the people/structure gap. | Strategy-led diagnosis; full study is gated and claims need careful source separation. |
Where the evidence genuinely conflicts
1. ROI: positive everywhere, repeatable almost nowhere
EY reports 98% positive ROI among its surveyed AI-investing leaders. NVIDIA’s financial-services survey reports 83% seeing ROI and 89% reporting both revenue and cost improvement. Capgemini’s RAISE PDF claims 1.7× average ROI on first use cases. Those numbers sit beside McKinsey’s finding that most organizations report no meaningful bottom-line impact, KPMG’s 74% who see value versus only 24% who achieve ROI across multiple use cases, and Roland Berger’s conclusion that measurable results fail to materialize in most organizations.
This is not one empirical contradiction. It is a measurement contradiction:
| Positive camp measures | Skeptical camp measures |
|---|---|
| Self-reported positive impact, selected AI users, first use cases, or vendor/client cases | Repeatable portfolio ROI, bottom-line attribution, operating-model change, and cross-functional scale |
The practical conclusion is not that either side is false. It is that “ROI” must be qualified as perceived benefit, workflow benefit, first-use-case return, or audited portfolio outcome.
2. Scale now versus establish value first
BCG and IBM describe end-to-end, cross-domain agentic operations as the destination and urge leaders to build the platform and transformation machinery now. Capgemini and McKinsey are more incremental in execution: start with high-value workflows, microtransformations, model routing, and cost attribution, then scale reusable patterns.
The tension is sequencing. A platform-first program risks infrastructure before practice. A use-case-only program risks a disconnected estate. The most defensible sequence is: one bounded workflow, instrumented and governed; a reusable pattern; then a platform and cross-domain expansion.
3. Platform bet versus portability
BCG says enterprises should pick one agentic platform and begin, arguing that lock-in will matter less as interoperability improves. OpenAI positions Frontier as a unified operating layer across enterprise systems. IBM emphasizes optionality and workload portability; Capgemini explicitly recommends modularity and vendor-independence; NVIDIA highlights hybrid architecture and open-source importance.
This is a real buyer conflict. A single platform may accelerate deployment and governance. Portability preserves bargaining power and model choice. The answer is not “multi-cloud everywhere.” It is to centralize policy, identity, telemetry, and workflow contracts while keeping model routing, data access, and state export portable enough to change providers.
4. Governance before autonomy versus autonomy with humans as validators
Deloitte, PwC, Anthropic, IBM, and Capgemini emphasize identity, least privilege, traceability, continuous monitoring, and human control. BCG and Capgemini describe a more autonomous future in which agents execute end to end and humans handle exceptions or direct the system.
These are not mutually exclusive technical positions. They differ on the maturity threshold for moving humans out of the critical path. In regulated finance, the burden should remain on the autonomy claim: increase authority only after the agent, tools, data, recovery path, and cross-system handoffs have been tested together.
5. Train people versus redesign roles
Deloitte reports that education is the most common talent response, while Roland Berger says people, skills, and organizational structure are the leading barriers. McKinsey and Accenture increasingly frame new roles—workflow owners, platform leads, governance liaisons, and AI-native operators—rather than training alone. OpenAI’s internal Codex evidence shows nontechnical workers crossing into technical work, but that is a vendor self-observation, not a workforce base rate.
The contradiction is diagnostic: training is easy to count; redesigned decision rights, incentives, and roles are harder. Training without redesigned work produces more AI-literate employees inside the same bottlenecks.
What to ask a consulting firm before accepting the recommendation
- “Which of your numbers measure audited financial outcomes versus perceived value?”
- “What is the smallest workflow you would instrument before proposing an enterprise platform?”
- “How do you preserve state, policy, telemetry, and evaluation portability if the model vendor changes?”
- “Which agent actions remain approval-gated in a regulated investment process?”
- “What evidence would make you stop funding a workflow?”
- “What part of your recommendation is a research finding, and what part is a sellable implementation pattern?”
Evidence Boundaries
This synthesis compares official consulting and vendor publications with the local research corpus. It does not treat firms’ headline percentages as directly comparable or as independent audits. Several claims are commercial case evidence, survey self-report, or executive opinion. The contradiction analysis is an inference from differences in populations, denominators, measurement definitions, and recommended sequencing.