McKinsey’s 2025 State of AI: agents are rising, impact is uneven, and UK leaders need to redesign work
A recent Reddit post summarised McKinsey’s 2025 State of AI findings and asked the big question: will AI agents complement people or replace them? The snapshot is clear: adoption is broad, scaling is hard, agents are getting traction, and the financial upside is concentrated in a minority of high performers.
“88% of organizations use AI in at least one business function.”
Below I break down what the post says, what’s changing with AI agents, and what this means for UK organisations navigating skills, compliance, and value capture.
What the Reddit post says about McKinsey’s 2025 report
Based on the summary:
- AI is widely adopted, but most organisations are still in pilots or early scaling. Enterprise-wide financial impact is limited to date.
- AI agents are gaining traction, especially in IT and knowledge management.
- High performers use AI for growth and innovation, not just efficiency, with deliberate workflow redesign and executive engagement.
- Workforce impact expectations are mixed: some expect reductions, others increases.
- Risk mitigation is improving but remains difficult, particularly around accuracy and explainability.
For the source, see McKinsey’s page: The State of AI. The Reddit thread is here: McKinsey on 2025 state of AI, plus an AI-enhanced view via Nouswise: link.
What this means for UK organisations: pragmatism over hype
For UK businesses, this maps to what many teams are experiencing: pilots everywhere, a few strong wins, and a lot of groundwork still to do. The shift from “tool” to “agent” is real in specific workflows, but broad productivity impact requires redesigning processes and roles.
Quick jargon primers:
- AI agent: a system that can plan and take actions across tools or data sources toward a goal, with varying degrees of autonomy.
- Explainability: techniques to make a model’s decisions understandable to humans.
Will AI agents complement or replace workers?
Over the next five years, expect mostly complement, with targeted replacement in routine-heavy tasks. The post notes agents are gaining traction in IT and knowledge management – two areas with repeatable patterns, clear guardrails, and measurable outcomes.
- Complementation: agents triage tickets, draft responses, surface relevant knowledge, propose fixes, or keep data in sync. Humans approve, escalate, and handle edge cases.
- Replacement: some first-line tasks that are highly templated and tool-based may be automated end-to-end.
The big lever is workflow redesign. High performers are not just dropping an agent into a broken process; they re-sequence steps, shift handoffs, change KPIs, and retrain teams. That’s where durable value comes from.
Where UK roles are exposed vs augmented
- Most exposed: high-volume, rules-based coordination work (e.g., simple service desk triage, basic report assembly, repetitive data entry). Expect task replacement rather than wholesale job loss.
- Augmented: analysts, marketers, product managers, operations leads – anyone who benefits from faster synthesis, better retrieval, and draft-generation. The human edge remains in judgement, stakeholder context, and accountability.
- New roles: AI product owners, prompt/process engineers, model evaluators, risk and compliance specialists for AI, and workflow designers.
Risks and compliance: accuracy, explainability, and UK data protection
The post highlights accuracy and explainability as friction points. That is consistent with what UK teams face when they move from demo to production. If you can’t trust outputs or explain why a decision was made, you won’t scale beyond pilots.
Practical considerations for UK organisations:
- Data protection: align with UK GDPR and ICO guidance when using personal or sensitive data. Complete Data Protection Impact Assessments for higher-risk use cases.
- Model behaviour: implement guardrails (policy checks, tool-use limits) and set up continuous evaluation for accuracy and bias on your real data.
- Procurement: clarify data retention, training on your data, breach processes, and data residency in vendor contracts.
- Explainability: document model limits and decision points. Even for generative systems, keep auditable logs, rationale summaries, and human-in-the-loop checkpoints.
Costs and value: why the financial upside is still limited
The Reddit summary notes that enterprise-wide financial impact remains limited. That’s not surprising. Value lags adoption when:
- Pilots focus on demos rather than core KPIs.
- There’s no process redesign, just a tool swap.
- Success metrics aren’t tied to time saved, cost avoided, or revenue uplift.
- There’s no plan for change management or skills.
Where the numbers move, it’s often because teams anchor agents to business outcomes (e.g., ticket resolution times, documentation completeness, marketing cycle time) and retire legacy steps. If you’re just adding another step, you won’t see a return.
A practical 12-month plan for UK teams
- Pick 2-3 use cases with measurable outcomes (IT ticket triage; sales content assembly; knowledge retrieval for customer support).
- Prove value with tight loops: define success metrics upfront, run A/B comparisons, and time-bound the pilot.
- Redesign the workflow: reduce handoffs, automate the dull bits, and clarify when humans must approve.
- Upskill: short training for prompt patterns, tool orchestration, and critical oversight. Make it part of job design, not “extra work”.
- Mitigate risk: data minimisation, access controls, red-teaming for hallucinations, and documented exception paths.
- Integrate: wire agents into your existing tools (e.g., Google Workspace, Microsoft 365, ServiceNow) rather than building parallel stacks.
- Governance: light but real – owner, metrics, review cadence, and a stop/go decision.
If you’re looking for an accessible starting point, I’ve written up a simple automation pattern using ChatGPT with Google Sheets: Connect ChatGPT and Google Sheets.
Five-year outlook: not a fad, but not mass obsolescence either
AI agents aren’t a fad, but the transformative impact will be uneven. In five years, expect credible agent adoption across IT operations, customer support, finance back office, and knowledge management. The biggest shifts will be in how work is organised, not just who pushes the buttons.
- Complement dominates: most knowledge workers will supervise, curate, and make decisions with agents handling retrieval, drafting, and routine actions.
- Selective replacement: some first-line tasks in service, content ops, and data hygiene will be automated end-to-end.
- Capability gap widens: high performers who redesign workflows and measure outcomes will outpace those who pilot without changing how work gets done.
The conclusion from the Reddit summary is right: leadership engagement and workflow redesign are the differentiators. If you treat agents as bolt-on tools, you’ll stall. If you redesign around them, you’ll start to see the gains.
Final take for UK leaders
Adoption is high. Impact is still uneven. Agents are real, but success comes from process, not just models. Start small, measure hard, redesign workflows, and keep humans accountable at the points that matter. Do that, and the next five years will look like augmentation-driven productivity – with targeted automation where it genuinely makes sense.
