A framework from Elevation Capital has been widely shared in conversations about how AI changes work. It names five human skills that grow in value as AI adoption deepens, each defined by a clear division of labour. AI generates content; the person decides whether it fits the brand. AI handles discrete tasks; the person connects them across functions. AI produces options; the person picks the right one. AI optimises a solution; the person asks whether it is the right problem. AI drafts messages; the person builds the relationship.
The framework is useful. It describes where human value concentrates as AI takes on execution. A companion piece explores each of those capabilities. But the framework describes the destination, not the route. It says what the relationship between people and AI should look like. It does not say how to build the organisation, the workflow or the system that produces it, or how people develop the skills it assumes. That gap is where most deployments sit.
Each skill is harder to build than it looks
Taste is accumulated, not announced. In an organisation, taste isn’t one person's preference. It is an accumulated understanding of what the organisation values, what its customers respond to and what its position requires. It rarely exists in writing. It lives with the people who have been there long enough to absorb it. When AI generates, and a person applies taste to evaluate, the quality of that evaluation depends on how much of this context the person carries. When context sits with a few senior people, the bottleneck simply shifts from producing to reviewing. Building taste across a team means making that context visible: past decisions, the reasons, and examples of what good looks like.
Context synthesis needs structured information. In most organisations, context is split by design. Sales sees customers, operations sees delivery, finance sees cost, and compliance sees rules. AI deployed within one function reproduces that split: a sales system can optimise for promises operations cannot keep. Synthesis becomes a skill people can exercise only when systems provide a structured view across functions. Training cannot substitute for that structure. It can teach people to use it.
Judgement improves with better options, and suffers from automation bias. The quality of a person's judgement depends on the quality of the options in front of them. Three well-framed options, each with trade-offs, evidence and stated uncertainty, support better decisions than ten loosely framed ones. The opposite risk is also well documented. Research on automation bias, summarised by Parasuraman and Manzey, shows that people tend to defer to automated recommendations, especially under time pressure. High agreement with a system can mean it is good, or that people have stopped thinking. A randomised study presented by J-PAL at the India AI Impact Summit 2026 found exactly this: screeners given AI advice were slowed down without making better decisions. Building judgement means designing review to activate reasoning, and measuring whether it does.
Strategic instinct depends on visible objectives. AI optimises well. The risk is that it optimises well for the wrong thing. A service system tuned for resolution time can learn to close conversations without solving problems. This is Goodhart's law: when a measure becomes a target, it stops being a good measure. Questioning the objective is a human skill, but people can only exercise it if the objective is visible. Systems that state what they optimise for, and report how their behaviour changes, make strategic questioning possible.
Trust governs autonomy. Trust between organisations is built through consistent behaviour, kept commitments and honest handling of problems. AI can draft the message. It cannot decide to absorb a cost because the relationship matters more. In practice, trust becomes the limit on how much autonomy a system is given: routine exchanges can be handled by the system, and anything involving ambiguity, conflict or a significant commitment stays with a person. Building this skill means teaching people to recognise those moments, and designing systems that hand them over cleanly.
The legal duty has changed shape, not disappeared
Since 2 February 2025, the EU AI Act has required providers and deployers of AI systems to address their staff's AI literacy. The Digital Omnibus rewrote that duty on 27 July 2026. Article 4 now requires organisations to take measures that support the AI literacy of the people who operate AI systems on their behalf, without mandating a specific level. The Commission and Member States must support organisations, especially smaller ones, with guidance and practical examples.
There is no certificate to obtain. People working with AI should reasonably understand what it does, where it can go wrong, and what they are responsible for, and the organisation should be able to show the measures it took.
Skills build in a sequence
The Dreyfus model of skill acquisition describes how people move from novice to expert: from following rules, through recognising situations, to acting from experience. Working with agentic AI follows a similar path, in five steps:
Understanding: what the system does, what it does not, and where its limits are.
Reviewing: telling a good output from a plausible but wrong one, and explaining why.
Handling exceptions: resolving the cases the system passes on, and spotting patterns among them.
Governing: setting and adjusting boundaries, watching performance and deciding when autonomy can widen.
Improving: feeding corrections back, and redesigning the workflow as the system matures.
Most AI training stops at the first step. Most of the value sits in the third, fourth and fifth.
Different roles need different depths
Frontline teams need understanding and reviewing. They work with outputs every day and are the first to spot a wrong one.
Specialists need exception handling. They handle the hard cases, and their resolutions teach the system and their colleagues.
Managers need governing. They decide what the system may do and when its scope changes.
Leadership needs enough understanding to set direction and risk appetite, and to ask the right questions of both the system and its provider.
Learn in the flow of work
Kolb's experiential learning cycle describes how adults learn: by doing something, reflecting on it, drawing a conclusion and trying again. Working with AI fits that cycle naturally, if the work is designed for it. The 70-20-10 rule of thumb, which holds that most professional learning happens on the job, is not a measured law, but it points the same way. Classroom courses about AI do little on their own.
The most effective programmes put learning inside the work: reviewing real outputs with an experienced colleague, discussing hard exceptions as a team, and seeing how a correction changed the next result. The World Economic Forum's Future of Jobs Report 2025 expects 39% of workers' core skills to change by 2030. A team that learns continuously through its work keeps pace. One that trains once does not.
Measure whether learning changes the work
Kirkpatrick's model evaluates training at four levels: how people reacted, what they learned, whether their behaviour changed and whether results improved. For AI skills, the third and fourth levels matter most. Useful measures include how often reviewers catch errors, how consistent decisions are across a team, how quickly new colleagues handle exceptions independently and whether override rates fall as configuration improves.
A 12-month plan
Months 1 to 3: brief everyone who will work with the system on what it does and where its limits are. Pair reviewers with an experienced colleague. Record the briefing as an Article 4 measure.
Months 4 to 6: introduce structured review of exceptions. Record what was decided and why, and discuss the hardest cases as a team each week.
Months 7 to 9: give managers ownership of boundaries and monitoring, with a clear process for changing them. Start tracking decision consistency.
Months 10 to 12: use the record of decisions to improve the workflow, and build training for new colleagues from real cases.
How Mitochondria's agentic AI systems help teams build skills
The approach. Each unit shows its reasoning, flags what is uncertain, and explains why it passed a case to a person. Options are presented with their trade-offs, so review activates judgement rather than replacing it. Each system's objective is set with the organisation and stays visible.
What it enables. Teams learn through the work itself. Decision memory turns each resolved exception into an example the next colleague can learn from, which spreads taste beyond a few senior people. Managers see where people override the system, which shows where training or configuration needs attention.
The impact. Teams grow into the review, exception and governing roles that matter most, with a record of learning that supports the AI literacy duty.
Read more about Mitochondria's approach.
Sources
Mondaq, the Digital Omnibus on AI enters into force (Article 4)
Alice Labs, AI literacy under Article 4, before and after the Digital Omnibus
Parasuraman, R. and Manzey, D. H. (2010). Complacency and bias in human use of automation. Human Factors, 52(3).
Dreyfus, S. E. and Dreyfus, H. L. (1980). A five-stage model of the mental activities involved in directed skill acquisition.
Kolb, D. A. (1984). Experiential Learning: Experience as the Source of Learning and Development.
Kirkpatrick, D. L. and Kirkpatrick, J. D. (2006). Evaluating Training Programs: The Four Levels.
J-PAL AI Evidence Initiative, session at the India AI Impact Summit 2026, as presented
—
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 →