They add copilots.You redesign.
They save minutes.You move the P&L.
They train a few.You build capability.
The next 90 days belong to those who dare to redesign the work.
A 90-day programme that takes EU midsize companies from scattered AI tools to redesigned, measured workflows — and on to reinvention.
Midsize firms compete with agile digital natives and well-funded incumbents — with lean teams, tight capital and EU-grade regulation.
Global survey: 89% of organizations in McKinsey’s 2026 survey use AI regularly, yet only 37% see any EBIT impact.
Official EU statistics: only 30% of companies with 50–249 employees use any AI, vs 55% of large ones.
Why so little value? Most AI use stops at personal tools — the copilot trap.
SourcesEurostat, The use of AI technologies in the EU (Mar 2026; EU firms with 10+ employees) · McKinsey, The state of AI in 2026 (Aug 2026; 1,719 respondents worldwide).
The illusion of efficiency: personal copilots save time in small fragments that are absorbed, not converted into capacity.
The 3% reality: about 80% of software engineers see a gain of only ~3%; the top 20% see 55%.
Redesign first: in software product teams, redesigning before adding AI made organizations more than twice as likely to report productivity gains above 20% as adding AI onto existing processes.
You cannot chat your way to production-grade organizational performance.
SourcesMcKinsey, The state of AI in 2026 (Aug 2026) · Beyond the copilot (Aug 2026; software teams) · From adoption to impact: Three horizons (Jul 2026; global survey of leaders).
McKinsey describes three horizons of AI transformation; our 90-day project works in Horizon 2.
General-purpose tools assist parts of existing jobs. Micro-gains, little P&L impact.
Existing cross-functional workflows automated end to end.
Our 90-day project
Roles, workflows and operating models redesigned from scratch, with AI at the core.
Shaped in the strategy phase after day 90
SourceMcKinsey, From adoption to impact: Three horizons of AI transformation (Jul 2026); 582 leaders in a global online panel, Feb–Apr 2026.
Example: large-enterprise software delivery, based on McKinsey experience across multiple companies.
Each square = 1 FTE
Stop making individuals slightly faster; redesign the end-to-end product and service life cycle.
Software teams that redesign before adding AI are more than twice as likely to report productivity gains above 20%; redesigned software delivery is 3 to 5 times as productive.
Small teams supervising agent-driven execution; leaders are 5.3× as likely to report enterprise value when workflows are redesigned rather than left unchanged.
SourcesMcKinsey, Rewiring software delivery for the agentic era (May 2026) · Beyond the copilot (Aug 2026) · Three horizons (Jul 2026).
SourcesMcKinsey, How capability building can power transformation (2021; 38 listed companies) · Why transformations stall (2026) · To make a transformation succeed… (2017) · Six steps… (2015).
Adults learn by doing: after three months, people recall 10% of what they were told — but 65% of what they were told, shown and practiced.
Leadership day for the steering committee, then 2 days for the core team on its own workflow; AI basics for affected staff; refreshers around days 30 and 60; hands-on training before go-live.
The breakthrough project: a real workflow redesigned toward hard P&L targets.
Coaching and performance dialogues; the executive showcase; lessons written into standard procedures.
The broad academy to the 20–25% tipping point, co-taught by the change leaders from wave 1.
Transformation stalls when it is restricted to a trusted few. Democratizing change is essential for scale.
At least one third of training time and budget goes to frontline staff and supervisors. One in three companies already spends most on the frontline (33%, up from 22%).
Organizations that act on frontline recommendations are 80% more likely to keep improving how they work — the frontline is the reality check for agentic workflows.
SourcesMcKinsey, Capability for Performance (2013; forum–fieldwork–feedback, four-stage coaching, recall research by IBM via Whitmore) · To make a transformation succeed… (2017) · Building capabilities for performance: Global Survey (2014) · Organizational health is (still) the key to long-term performance (2024). Durations and the one-third rule: our programme design.
Owns the target and decisions; signs off rules and go-live.
Agents, instructions, knowledge and evaluation harness.
Knows the exceptions; writes and checks the test cases.
ERP/CRM connectors, gateway, security, monitoring.
Apprentice builder (your IT team): works alongside our engineers from day 1 and takes over routine changes.
Our engineers build the solution with your team, with full resources and executive attention: prove integration, speed and P&L upside.
Your line leaders run and extend it — another team, site or workflow — under normal conditions. We coach and back you up.
You own what we build together: code, data, prompts, knowledge graph and test cases. Only 9% of companies run double pilots.
SourcesMcKinsey, Beyond the copilot (Aug 2026; definer–builder pods: 8–10-person teams → 4–6) · Capability for Performance (2013) and Building capabilities for performance (2014) on double pilots.
Asynchronous execution: work no longer waits for people. Agents process queued work whenever it suits; people spend their working hours reviewing, deciding and handling exceptions.
Invoice matching, supplier checks, planning runs: reviewed next morning.
Customer-service replies drafted in seconds; nothing goes out before approval.
A new order or quote request starts the agent straight away.
One option, from software delivery: a 16-hour agent “night shift” and an 8-hour human day.
Three rules decide whether faster agents create value — or just more work to check.
If agents produce more overnight than people can review the next morning, you only build a backlog — and review is already the largest running cost of an agent.
In practice: size agent output to review capacity; risk-tier the review.
Outside working hours, agents prepare drafts — they send no offers and take no decisions that significantly affect people. The GDPR restricts solely automated decisions; the AI Act requires human oversight of high-risk uses.
In practice: no customer-facing action without approval.
Escalations wait for the morning, and nobody is put on call for the agents. The rhythm changes for the machines, not for your staff.
In practice: define what agents may do alone — and what waits.
Bonus: batching non-urgent, high-volume requests also lowers model costs. The catch: speed creates volume, and every output an agent produces must be checked — that is where the money goes.
SourcesMcKinsey, Rewiring software delivery (2026) · How to build businesses faster and better with AI (2026) · AI for IT modernization (2024) · MGI, Agents, robots, and us (2025) · Beyond the copilot (2026) · Where AI agents pay off (Aug 2026) · The cost of intelligence (Jul 2026) · Regulation (EU) 2016/679 (GDPR) and 2024/1689 (AI Act). The three rules: our programme design.
About 60% of an agentic task’s cost goes into checking, repairing and re-verifying answers (research on coding agents).
In a banking customer-service example, human oversight is 70–75% of an agent’s variable run cost; model tokens are 20–25%.
Depending on its path, tools and retries, the same task can cost up to 30 times more.
AI spend rises nearly fourfold as use moves from isolated cases to the whole company.
So what: cut the sink at the source — better context means fewer errors to check — and measure what each outcome really costs (FLCCO).
The defaults reproduce the deck’s example — one technical quote. Move the sliders to match your workflow.
More than 10% rework is beyond our ≥90% go-live gate.
The flawed IT metric: cost per token (€0.05) makes an agent look almost free, while the workflow around it is not.
Operations = middleware + maintenance. Running costs only: one-off build and implementation costs belong in the payback calculation.
Cut review effort and errors first. Most of what remains is people’s time — review, rework and maintenance; tokens are almost negligible.
Note on rates: €75 an hour is an assumed fully loaded cost of a senior engineer or IT specialist at a Western-European level — salary, employer contributions and overhead, per hour actually worked (about €10,000 a month). Rates differ between European regions and between industries; Phase 1 of the programme uses the client’s own figures.
Store stable context — instructions, rules, ontologies — and reuse it: up to ~90% lower cost on repeated input.
Small models for routine work, frontier models for hard reasoning. Saving ≈ share routed × (1 − price ratio): 60% to a 10× cheaper model cuts compute by 54%.
Structured prompts, only the relevant context, summarized histories, batching of non-urgent requests — savings measured in the pilot.
Limits on retries, tool calls and loops. After five steps without a verified outcome, the case goes to a person with its full log.
Keep perspective: compute is a small share of the cost per outcome — under 1% in our quote example, 20–25% in a McKinsey banking example where only a sample of answers is reviewed. Even there, a 54% compute saving lowers variable cost by only about 11–14%. Review and rework remain the bigger levers.
SourcesMcKinsey, Is that AI agent worth it? (Jul 2026) · Where AI agents pay off (Aug 2026) · The cost of intelligence (Jul 2026) · Trust in the age of agents (2026). The quote example is a calculation exercise based on the stated assumptions. Routing formula and five-step limit: our design.
Generic models lack your company’s context — and EU rules require data protection, human oversight and auditability by design (GDPR; use-case duties under the AI Act).
Value comes from a secure envelope around your systems: ERP, CRM, documents and the rules that govern them.
The defining question is not how autonomous agents can become, but how much autonomy your company can safely absorb and govern.
tested against golden evals
ERP · CRM · documents · GDPR controls
broad knowledge, no company context: hallucination risk
decision rules, edge cases, rules of thumb
expert shadowing and decision mapping
machine-readable rules and schemas
real cases with expert-validated answers; go-live at ≥90% pass rate
The barrier to scaling AI is capturing the tacit judgment of your top performers and turning it into machine-readable rules.
A proprietary set of 100+ real cases with expert-validated answers: the strict definition of what “good” looks like in your workflows.
Every version of the workflow is tested before and after go-live, so agents behave like your best people — not like generic web crawlers.
“The key to unlocking the value of agentic AI is evaluations.”— McKinsey partner, Technology Trends Outlook 2026
products, customers, suppliers, contracts and rules as connected entities
EU-hosted, on-premises or external
Knowledge graphs organize scattered internal data into connected objects, properties and links that agents can reason over.
Gateways based on the Model Context Protocol (MCP) — an emerging open standard — bridge internal data and models safely.
Audit trails, access control and data residency support GDPR and AI Act duties; they do not replace a lawful basis or impact assessment.
SourcesMcKinsey, Trust in the age of agents (Mar 2026) · Building the foundations for agentic AI at scale (Apr 2026) · Technology Trends Outlook 2026 · MGI, Agents, robots, and us (Nov 2025) · Regulation (EU) 2016/679 (GDPR) and 2024/1689 (AI Act). Thresholds (100+ cases, ≥90% pass rate) and architecture: our programme design.
Measured as: cost and hours per quote; win rate
Measured as: cost per resolved case; resolution time
Measured as: cost per purchase order; savings captured
Measured as: planning cycle time; forecast accuracy
Measured as: cost per invoice or reconciliation; closing time
Measured as: time to answer; expert hours saved
Click the plays that matter to you — they appear in your shortlist at the end of the page.
Seventh option for firms with their own developers: software & IT modernization. Each play ships with connectors, agent and test-case templates and a baseline cost sheet.
Leadership day (before signing): a first pass of steps A–B shortlists the plays; the steering committee then chooses 1–2 workflows plus a reserve. Phase 0, weeks 2–3: all five steps run on the chosen workflows and end with a confirm-or-switch gate.
EvidenceMcKinsey, State of AI 2026: savings most often reported in supply chain, service operations and manufacturing; revenue gains in marketing & sales; agents scaled most in IT and knowledge management.
About 16 weeks from signing to day 90: a three-week Phase 0 (train and diagnose) plus the 90-day project. The paid leadership day comes before signing.
How to use freed capacity is a strategic choice — and it is yours.
Redeploy freed hours to sales coverage, service quality or product development.
Re-skill or hire for new growth directions such as data, automation or AI-enabled services.
Lower cost within a broader cost and sustainability plan.
Governance rule: success means a deliberate, documented decision on freed capacity. If roles change, involve HR, legal counsel and the works council early.
Programme designAI Adoption & Performance method. Targets are set per workflow from the Phase 1 baseline; consultation thresholds for dismissals differ by country. Refresher forums around days 30 and 60.
Your line leaders run the workflow and extend it to another team, site or workflow.
Participation grows to the 20–25% tipping point, co-taught by your change leaders.
An AI strategy aligned with — and part of — your business strategy, including Horizon 3 options.
An optional support agreement and ongoing advice for your next AI initiatives.
Conclusion: AI is not an IT expense; it is the fundamental redesign of midsize European operational capacity.
From the first redesigned workflow to an AI strategy that is part of the business strategy.A half-day executive workshop to decide where AI belongs in your processes — and which workflow to redesign first.
Figures labeled illustrative are worked examples. Programme targets are our design parameters, set per client from the Phase 1 baseline.
