Enterprise AI is at an inflection point. Investment in models, compute, and experimentation keeps rising, yet confidence at the board and operating levels is wavering. Leaders no longer ask whether AI is powerful. They ask where it produces lasting business value.
The evidence explains the doubt. MIT's 2025 research found that only about 5% of generative AI pilots produce measurable business impact. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. McKinsey's 2025 State of AI survey found most organisations still experimenting or piloting, with only a minority scaling.
Every technology wave follows this arc
Gartner's hype cycle describes a familiar path: a peak of inflated expectations, a trough of disillusionment, then a slow climb to productive use. Everett Rogers' research on the diffusion of innovations explains why. Early adopters tolerate rough edges and work around gaps. The majority need technology that fits how they already work, with proof that it delivers.
This moment is not a slowdown in AI adoption. It is a necessary correction: a shift from AI as an impressive capability to AI as a dependable system.
The problem is misalignment, not models
Models have become very capable, yet most enterprise implementations remain shallow. AI is layered onto existing workflows rather than built into how work is done. The pattern is familiar:
dozens of pilots, few systems in production;
local productivity gains, limited organisational impact;
AI tools working in isolation from core operations;
unclear ownership between IT, business and strategy teams.
The result is disappointment, not because AI fails at reasoning, but because it is asked to perform without the structure it needs. It fails at execution.
Where value comes from
Summarising documents and drafting content help individuals, but the gains are incremental. They do not change how decisions are made, how processes flow or how organisations learn. Enterprise value appears when AI is used to:
redesign workflows end to end;
coordinate decisions across functions;
reduce dependence on a few individuals' judgement;
shorten whole operational cycles, not just individual tasks;
keep quality consistent as complexity grows.
McKinsey's survey supports this: the organisations seeing the most value were nearly 3 times as likely as others to have redesigned their workflows around AI.
Fewer pilots, not more
Many organisations try to reduce risk by running many pilots in parallel. It often increases risk. Effort fragments, accountability blurs, and none of the pilots gets the attention it needs to cross into production.
What scales instead is focus: a small number of high-impact workflows, clear success measures tied to operations or revenue, deep integration with existing systems, and disciplined improvement rather than constant reinvention. AI maturity is not measured by how many experiments exist, but by how many systems the organisation relies on.
John Kotter's research on why change efforts fail offers a parallel. Transformations stall when organisations declare victory too early, fail to anchor changes in how work is done, or never build a guiding coalition with real ownership. AI programmes fail in the same ways.
Architecture before intelligence
A recurring mistake is treating intelligence as the starting point. Architecture comes first. Before AI can create value, an organisation must answer basic questions. Where does decision authority sit? Which data is trusted, and when? Which actions can be automated safely? Where must people stay involved? How are errors detected and contained? Without those answers, AI amplifies confusion rather than resolving it.
The hidden risk of shallow automation
There is a quieter risk. Some organisations use AI to remove junior roles without redesigning how knowledge flows. Research from Stanford's Digital Economy Lab in 2025 found that early-career employment in the jobs most exposed to AI had already started to fall relative to other groups.
For the organisation, the risk is not only social. Junior roles build expertise. When they disappear without a new path for learning, supervision and succession, organisations lose the pipeline that produces tomorrow's senior people. Better design keeps those roles and changes them: systems handle routine execution, while juniors learn by reviewing outputs, handling exceptions, and working alongside experienced colleagues, with decision memory giving them the precedents a mentor once had to supply.
What durable AI systems share
They are built into workflows people already use.
They work across multiple steps, not single tasks.
They reduce handovers rather than adding new ones.
They make outcomes more predictable, not less.
They have one clear owner for each outcome.
They keep people learning, not only people approving.
Questions for leadership
Which few workflows, if transformed, would change our results?
Who owns each one end to end, across functions?
What will we stop piloting so those few get the focus they need?
How will we measure value in operational or revenue terms?
How will junior colleagues build expertise as routine work moves to systems?
How Mitochondria's agentic AI systems scale
The approach. Each engagement focuses on a small number of workflows that matter, mapped as they run and configured to the organisation's own rules. Deployment moves through four stages, with one owner per outcome and autonomy widening as reliability is shown.
What it enables. Systems built into the work rather than beside it, measured against business outcomes, with decision memory that helps new colleagues learn from the organisation's own precedent.
The impact. Fewer pilots, more systems the organisation relies on, and people whose expertise keeps growing as the routine work moves to systems.
Read more about Mitochondria's approach.
Sources
Rogers, E. M. (2003). Diffusion of Innovations (5th edition).
Kotter, J. P. (1995). Leading change: why transformation efforts fail. Harvard Business Review.
Brynjolfsson, E., Chandar, B. and Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab.
—
Mitochondria builds agentic AI for evidence-led core operations. Research-led and model-agnostic. ISO/IEC 27001:2022 certified, with controls aligned to the GDPR, UK GDPR, India's DPDP Act and Australia's Privacy Act. Amsterdam and Pune. Request a working session →