Where organisations actually stand in 2026
The narrative around AI has shifted decisively from experimentation to execution. 88% of organisations now deploy AI in at least one function — yet the gap between wide adoption and genuine value creation has never been wider. Two landmark research programmes — McKinsey's State of AI 2025 and BCG's AI-First Companies Win the Future — converge on a striking paradox: AI is everywhere as a tool, but still rare as true enterprise infrastructure.
"AI doesn't lack capabilities — organisations lack the structure to absorb them."
BCG / McKinsey cross-analysis, December 2025The evidence is now clear: the decisive variable is not AI access — it is organisational plasticity. The willingness to redesign workflows, restructure teams, and rebuild the underlying data architecture determines whether AI becomes a strategic moat or just another IT line item.
How the big firms are actually transforming
The most instructive benchmarks come from consulting's own frontline. These firms are not just advising clients on AI transformation — they are executing it internally at radical scale.
Project Magnolia & the Lilli Platform
McKinsey launched Lilli, a proprietary GenAI platform giving every consultant on-demand access to the firm's full knowledge base, research archives, and analytical tools. Internally, the "Project Magnolia" restructure cut thousands of back-office and support roles — initially targeting 2,000 positions in 2023, with further reductions reducing overall headcount from ~45,000 to under 40,000. Junior teams of 2–3 now routinely replace what previously required 14-person engagements.
Simultaneously, McKinsey deployed over 12,000 AI agents across its operations. AI and tech advisory now contributes an estimated 40% of the firm's $16bn revenue.
Reinvention Services & the $865M Overhaul
In the biggest restructure in Accenture's history, CEO Julie Sweet collapsed five formerly discrete business units into a single integrated entity — "Reinvention Services" — to embed AI and data more fluidly across all service lines. Sweet announced on video (not memo) to her 770,000+ workforce: employees who cannot be reskilled for AI-driven workflows will be "exited on a compressed timeline."
The programme: $865M restructuring, 11,000 exits, 550,000 employees reskilled in GenAI fundamentals, and headcount of AI/data professionals nearly doubled to 77,000 within two years. Target: deploy 100 industry-specific AI tools by end-2025.
The AI-First Operating Model
BCG's internal and external thesis is explicit: "The AI-first operating model rewires how organisations work." Hierarchies flatten as AI agents — overseen by small elite teams — run back-office processes. BCG research identifies the 10/20/70 rule: 10% of AI value comes from algorithms, 20% from technology, and 70% from people and process redesign.
BCG describes a structural P&L shift: technology spending rising to 35–45% of operating cost in some sectors (from 20–30%) as labour is replaced by compute. Business units lead AI adoption; IT provides the scalable platform underneath.
Agentic Infrastructure as Strategy
Bain's Technology Report 2025 frames agentic AI as "a structural shift in enterprise technology, not just another wave of automation." Their architectural framework identifies three non-negotiable layers: orchestration (the command centre), observability (real-time audit of every agent action), and governed data access.
Critically, Bain argues governance and trust are the foundation of agentic AI — not a compliance bolt-on. Security, traceability, and anomaly detection must be embedded from day one, or scale is impossible.
What the AI-first operating model actually looks like
Across all leading firms, a consistent pattern emerges. The AI-first operating model is not a technology upgrade — it is a fundamental redesign of how work is organised, who does what, and how value is created. Below is the maturity arc from experimentation to true AI-first operation.
- Point solutions & pilots
- Individual productivity tools (Copilot, ChatGPT)
- No redesigned workflows
- IT-led, not business-led
- ~60% of enterprises here
- Function-level deployment
- First workflow redesigns
- AI CoE / transformation office
- Role-based training at scale
- Senior leader sponsorship
- End-to-end process "agentification"
- Org restructure around AI capabilities
- Outcome-based commercial models
- Proprietary data moats built
- Human–agent team design
- AI agents run core operations
- Lean elite teams (2–3 per workflow)
- P&L restructured: high tech, low headcount
- Continuous model evaluation in prod
- ~5–6% of enterprises globally
The six organisational moves that separate leaders
Workflow redesign first
High performers redesign processes before deploying AI — 55% vs. 19% of others. Plugging AI into broken processes produces broken AI.
New talent architecture
Junior roles evolve from "first draft" producers to "AI directors." 29% of heavy adopters expect fewer traditional entry-level roles; 43% want more generalists who can manage human–agent teams.
CEO-level sponsorship
Executive engagement is the strongest predictor of AI maturity. Employees in leading firms don't see AI as mandated from HQ — they see their direct managers using it daily.
Dedicated transformation office
A dedicated AI adoption and scaling team — separate from IT — is consistently present in high-maturity firms. It holds commercial accountability for AI-driven value, not just deployment.
Reskill or exit — no middle path
Accenture's explicit "reskill or exit" programme is increasingly the industry norm. A "strategic skills mismatch" — not performance — is the trigger. The message: AI capability is non-optional.
Big bets, not sprinkled pilots
BCG's "reshape and invent" model demands focusing on a few transformational bets rather than hundreds of marginal pilots. EBIT impact only materialises when AI is applied end-to-end, not function by function.
The architecture of an AI-first organisation
Legacy enterprise architecture was designed for discrete, batch-driven workflows. Agentic AI demands a fundamentally different stack — one built for continuous, autonomous, multi-model operation. Eight in ten enterprises cite data architecture fragmentation as the primary blocker to scaling AI beyond pilots. The following layers define the target state.
| Layer | What it is | What "AI-first" looks like |
|---|---|---|
| Data Foundation | Unified data platform across structured, unstructured, vector, and graph stores | Medallion architecture (raw → curated → agent-ready). Real-time ingestion. Every data asset is a product. Lineage and audit trails built in. Fine-tuned domain models on proprietary data reduce inference costs. |
| AI Gateway / LLM Router | Middleware between applications and foundation model providers | Centrally governed access to multiple LLMs (Claude, GPT-4o, Gemini, Llama, SAP-RPT-1). Routes workloads by cost, latency, and compliance need. Enforces usage policies. Prevents vendor lock-in. MCP standard increasingly preferred. |
| Orchestration | The command centre directing multi-step agentic workflows | Manages control flow, retries, timeouts, and parallel agent execution. Shared platform services across all agents. Central registry of approved tools and entitlements. MCP connections govern what each agent can and cannot touch. |
| Observability | Full real-time visibility into agent execution, decisions, and costs | Full reasoning-path traceability: every prompt → tool invocation → output. Integrated token cost management. Anomaly detection, hallucination monitoring, and bias signals. Live dashboards. Audit-ready for regulators. |
| Compute & Cloud | Elastic infrastructure for probabilistic, variable AI workloads | GPU-optimised, multi-cloud (AWS, Azure, GCP). Infrastructure-as-Code with CI/CD for agent deployment. Data residency compliance baked in. AI workloads treated as mission-critical, not experimental. |
| Governance & Security | Access control, accountability, and risk management for AI systems | RBAC at the agent level. Deterministic execution boundaries (agents can only call approved tools). AI governance policies approved at board level — currently only in <25% of firms. AI as important as capital allocation decisions. |
| Applications & Agents | The business-facing layer: AI-native apps and autonomous agents | Domain-specific agents per function (procurement, customer ops, software engineering, knowledge management). Human-in-the-loop only where required. Agents surface within existing SAP / Salesforce / CRM toolchains via API/MCP. |
"The choice of foundation model vendor and agent framework are not independent decisions. If agents run on a vendor's proprietary orchestration layer, lock-in compounds at every layer of the stack."
Kai Waehner — Enterprise Agentic AI Landscape 2026The SAP dimension
For E2E transformation consultancies, the SAP layer is now a critical AI battleground. In late 2025, SAP released SAP-RPT-1, its first enterprise relational foundation model purpose-built for structured business data, and SAP-ABAP-1, trained on 250M+ lines of ABAP code. These are not general-purpose LLMs — they are domain models for the workflows SAP clients actually run. AI-first SAP transformations combine these native models with agentic orchestration to automate code migration, compliance checks, and finance operations at a fraction of previous effort.
What this means for valantic's AI-first story
valantic's E2E positioning — from Strategy through SAP transformation, Data & AI, Customer Experience, and Custom Engineering — sits at precisely the intersection where AI-first value is being created and captured. The benchmarks above reveal both the urgency and the specific moves that matter.
Define our AI posture — publicly
BCG's research shows the highest risk is drifting from pilot to pilot without a named strategic posture. valantic needs a declared archetype: are we "Internal Transformer," "Business Pioneer," or both? The story must be specific, not aspirational.
Build proprietary AI assets
McKinsey's Lilli and Accenture's 100 industry-specific tools show the direction: proprietary, reusable AI assets compound in value. valantic's SAP expertise, industry data, and accumulated client knowledge are the raw material for differentiated models and agent libraries.
Rewire the talent model
A reskilling programme is not optional — it is the programme. Every consultant needs a baseline AI capability (prompt engineering, agent design, output validation). New specialist roles — AI architects, agent designers, AI assurance leads — need to be grown or hired explicitly.
Sell the new operating model, not the technology
Clients are not buying AI tools — they are buying a new way of operating. valantic's pitch should lead with workflow redesign, governance architecture, and measurable EBIT impact — not the underlying model names. The 10/20/70 rule applies to our clients too.
Make data readiness the entry point
80% of enterprises cannot scale agents because their data architecture is fragmented. Data readiness assessments, medallion architecture design, and AI gateway setup are the natural entry-point engagement for valantic — converting data clients into full transformation clients.
Lead on AI governance
Fewer than 25% of enterprises have board-approved AI governance policies. This is a white space valantic can own — particularly in regulated industries where SAP and compliance expertise already opens the door. Governance-as-a-service is a defensible, recurring offer.
McKinsey State of AI 2025 (Nov 2025) · BCG AI-First Companies Win the Future (Jun / Oct 2025) · BCG AI Transformation is a Workforce Transformation (Feb 2026) · BCG Agents Accelerate the Next Wave of AI Value Creation (Dec 2025) · Bain The Three Layers of an Agentic AI Platform (Apr 2026) · Bain Building the Foundation for Agentic AI (Technology Report 2025) · Accenture SEC Form 8-K restructuring announcement (Jun 2025) · McKinsey Scaling Agentic AI with Data Transformations (Apr 2026) · Kai Waehner Enterprise Agentic AI Landscape 2026 (Apr 2026) · Salesforce The Agentic Enterprise IT Architecture (2025)