McKinsey’s State of AI 2025: Findings, Risks & How to Scale

The new State of AI shows a maturing market: more teams use AI daily, but most firms still struggle to scale it and prove enterprise-level impact. Leaders stand out by setting growth objectives, rewiring workflows, and establishing measurable controls—often with AI agents front and centre.

What’s new in 2025 vs. 2024

In 2024, McKinsey reported that 65% of organisations were regularly using generative AI—a rapid jump from 2023’s early adoption. The 2025 materials emphasise a different gap: only around one-third report scaling AI across the organisation. The message: usage is up; value at scale remains elusive.

Five takeaways that matter now

1. Scaling is the bottleneck. Many firms report pilots, few show end-to-end transformation or EBIT impact at enterprise level. Larger companies are more likely to be scaling, but even they cite workflow, data, and operating-model blockers. McKinsey & Company
   
   

2. High performers target growth and innovation—not just cost. Eight in ten cite efficiency aims, but the leaders add revenue and innovation goals, helping secure investment and cross-functional commitment. McKinsey & Company
   
   

3. Workflow redesign is non-negotiable. Winners don’t “bolt on” models; they rebuild processes (e.g., sales playbooks, support runbooks, software delivery lifecycle) and re-platform content/knowledge so AI can act reliably. McKinsey & Company
   
   

4. AI agents move from hype to utility. 2025 materials highlight emerging adoption of agents that can plan, call tools, and perform multi-step work—especially in software engineering and customer ops—when paired with policy guardrails and retrieval. McKinsey & Company
   
   

5. Measurement is immature. Many organisations still lack robust, leading KPIs for gen-AI initiatives; where tracking exists, value realisation rises and risk incidents fall. (Generation Digital view, aligned to McKinsey’s persistent call for governance + metrics.) McKinsey & Company

Numbers the board will ask for

• Adoption level: Regular gen-AI use reached ~65% of organisations in early 2024; 2025 emphasises the scaling hurdle rather than a simple adoption percentage. McKinsey & Company
  
  

• Economic potential: McKinsey’s baseline sizing for gen AI remains $2.6–$4.4 trillion of annual value across 63 use cases, with the bulk in customer operations, marketing & sales, software engineering, and R&D. Treat as directional, not guaranteed. McKinsey & Company
  
  

• Usage vs. leadership perception: Employees often use AI more than leaders think—underscoring the need for enablement and policy rather than blanket bans. McKinsey & Company

Generation Digital’s interpretation: how to move from pilots to proof

1) Start with business goals, not features. Mirror high performers: codify revenue or innovation objectives alongside cost. Translate these into outcome-based KPIs (cycle time, conversion rate, CSAT, defect escape). McKinsey & Company

2) Rewire a whole workflow. Pick one value stream (e.g., L1→L2 support escalation, SDR→AE handoff, incident→post-mortem). Redesign artefacts (playbooks, taxonomies), decision points, and tooling to make AI the default path—not an optional add-on. McKinsey & Company

3) Build an agent-ready stack. Introduce policy-aware retrieval, tool calling, and audit trails. In engineering, connect agents to ticketing, code repositories, CI pipelines; in CX, connect to CRM, knowledge bases, and telephony. McKinsey & Company

4) Govern for safety and speed. Establish an approvals matrix by risk tier; pre-approve tools/datasets; log prompts and outputs; define rollback. This shortens time-to-value while satisfying compliance.

5) Measure leading and lagging indicators. Track adoption (active users, tasks automated), quality (hallucination rate, guardrail blocks), and business results (EBIT impact, revenue lift). Tie use cases to a financial model early.

Where to deploy first (practical bets)

• Software engineering: Code suggestions, PR drafting, test generation, incident summarisation, and root-cause analysis—where McKinsey repeatedly sees value concentration. McKinsey & Company
  
  

• Customer operations: Assisted resolution, knowledge surfacing, next-best-action; fast ROI when paired with well-structured content. McKinsey & Company
  
  

• Sales & marketing: Personalised messaging, proposal assembly, and pipeline hygiene with agentic workflows. McKinsey & Company

Risks & controls

• Data leakage & provenance: Use scoped retrieval and redaction; watermark sensitive outputs.

• Hallucinations: Ground every response; measure answer-acceptance and override rates.

• Change fatigue: Invest in enablement—McKinsey notes disconnects between perceived and actual employee usage; formalise enablement instead of shadow AI. McKinsey & Company

Notes for UK based organisations

• Regulated sectors (FS, Health, Public) can still scale by adopting tiered risk models and human-in-the-loop checkpoints.

• For UK multinationals, align UK GDPR with global AI policies; maintain model cards and DPIAs where appropriate.

The bottom line

The 2025 State of AI tells a clear story: AI won’t create enterprise value by itself. Value arrives when leaders set growth targets, rewire workflows, adopt agent-ready stacks, and measure outcomes. If your pilots aren’t moving EBIT, the problem is likely operating model and measurement—not the model itself. McKinsey & Company+1

FAQ

Q1: What are the biggest changes in McKinsey’s State of AI 2025?
Adoption is broad, but scaling is limited. High performers target growth as well as efficiency and redesign workflows—often with AI agents—to capture value. McKinsey & Company+1

Q2: What percentage of organisations use generative AI?
McKinsey reported ~65% regular gen-AI use in early 2024; 2025 pivots to the scaling challenge rather than headline adoption. McKinsey & Company+1

Q3: Where is the economic value?
Largest pools: customer operations, marketing & sales, software engineering, and R&D—contributing to an estimated $2.6–$4.4T annual impact potential. McKinsey & Company

Q4: How should we measure AI success?
Track adoption and quality leading indicators plus business KPIs (e.g., CSAT, conversion, cycle time, EBIT). Scaling without metrics is where programmes stall.