75% of Boards Admit Their AI Strategy Is More for Show Than Real Steering

They're usually built on an unchanged operating model, the technology gets laid over old structures, and nothing structural actually moves, which means the pilot only ever works as well as the broken process it was quietly layered on top of. A pilot that succeeds in a controlled test but never scales is rarely a technology failure; it's almost always a sign the underlying decision rights and workflows were never redesigned to actually use what the technology can do.

Redesign the operating model and decision rights first, then choose pilots specifically because they prove something the next step can build on, rather than picking whichever pilot looked most impressive in a demo. A pilot chosen for demo value tends to prove very little about production readiness; a pilot chosen because it tests a specific operating-model assumption tells you exactly what needs to be true before scaling further.

Most organisations only use the surface layer, text generation and summaries, while the higher-value applications like agents, end-to-end automation, and decision support go untried, largely because nobody has redesigned the workflows those applications would actually need to plug into. The technology capable of producing real return has usually been available for a while; what's missing is the operating-model work required to put it to use.

Allocate every task across four capacity types, fixed employees, flexible talent, machines, agents, based on what needs human judgement versus what can be standardised or automated, rather than making a blanket decision about which roles AI should or shouldn't touch. This task-by-task allocation is more granular than most organisations attempt, but it's what actually reveals where automation adds value versus where it would just formalise a bad process faster.

Multi-step agents that complete sequences of tasks, end-to-end process automation, scenario-based decision support, monitoring that surfaces unread signals already sitting in existing data, and cross-department coordination are all realistic today, not experimental, yet most organisations never get past a chat window for text and summaries. The gap between what's technically available and what's actually deployed is usually an operating-model gap, not a capability gap.

Build one integrated view combining workforce planning and AI strategy, instead of running them in separate meetings where HR plans headcount for work that a parallel AI initiative is quietly planning to change or eliminate. Keeping these two conversations apart means the organisation ends up planning for two different futures at once, and only discovers the mismatch once both plans are already underway.

An operating model that explicitly allocates work across employees, flexible talent, machines, and agents, redesigned before automation is applied rather than automated within whatever structure already existed. The redesign step matters because simply adding AI to an unchanged structure tends to just insert a new tool into old workflows, without capturing the efficiency the tool was actually capable of delivering.

Map every task and decide what requires human judgement, what can be standardised, and what AI can fully handle, automating a badly designed process just makes it fail faster, and at greater scale, than it did when humans were still catching the errors manually. This mapping exercise usually reveals that a meaningful share of tasks in any given workflow don't actually need to happen the way they currently do, automated or not.

Measured via a 7-dimension, 5-level maturity assessment covering organisation, IT infrastructure, data management, ethics, privacy & security, people & skills, and performance management, rather than a general impression based on how many AI tools are already in use somewhere in the business. Tool adoption and genuine organisational readiness are frequently quite different things, and the assessment is specifically designed to separate the two.

Score each dimension in a working session with the client, validated against detailed level descriptions so the score has real meaning, rather than a self-assessment survey completed independently by whoever happens to be available. Detailed level descriptions matter because a maturity score without a clear, shared definition of what each level actually looks like tends to drift toward whatever number feels most comfortable to report.

Make decision ownership explicit as part of the operating model redesign, accountability without authority is the top cause of stalled AI rollouts, where someone is expected to deliver results from a system they have no actual power to change or govern. Naming a specific owner with real decision rights over a given AI application tends to be the single change that unsticks a rollout that's been stalled for unrelated-seeming reasons.

Define what an agent may decide, which sources it may use, how it substantiates answers, when it escalates to a human, and how usage is logged and monitored, treating this as a governance design exercise rather than a technical configuration afterthought. An agent given real decision-making latitude without these boundaries explicitly defined tends to behave unpredictably in exactly the edge cases where predictability matters most.

Make execution capacity visible via the maturity assessment instead of debating pace in the abstract, it turns the conversation concrete, replacing a subjective disagreement about ambition with a specific, evidence-based picture of what the organisation can realistically deliver and by when. Boards tend to respond better to a credible capacity constraint than to a general sense that management is moving too slowly.

Sequence foundational maturity dimensions first, then layer pilots chosen specifically to prove what the next stage needs, rather than starting with the most visible or exciting use case regardless of whether the foundational capability to support it actually exists yet. A roadmap built this way tends to look less impressive in year one and considerably more durable by year two.

Use a development platform where learning outcomes belong to the individual and travel with them, with privacy by design, rather than a corporate training programme whose records disappear the moment someone changes roles or leaves the company. Individually-owned learning records also tend to produce more honest engagement, since employees aren't building a skills profile that only ever benefits their current employer.