A growing tendency describes AI as creative. Models write, design, suggest and simulate. Their outputs feel novel, sometimes surprising.
That framing hides something more useful. AI does not originate purpose. It expresses the choices built into it. Every meaningful AI outcome sits downstream of human decisions: what data is used, what objectives are set, what constraints are enforced, which trade-offs are allowed and which errors are tolerated. These are not only technical choices. They are judgements, made by people, embedded in systems.
What creativity research says
Teresa Amabile's componential model of creativity names three ingredients: domain expertise, creative thinking skills, and motivation driven by the task itself. A fourth, the social environment, shapes all three. Generative models bring something new to the second ingredient: they can recombine and generate options at enormous scale. They do not bring the domain judgement to know which options matter, or the motivation that comes from caring about a particular outcome.
That is why AI outputs alone rarely change an organisation. The value appears when generation is joined to expertise, purpose and judgement, and those come from people.
Design is choosing what should be
Herbert Simon, in The Sciences of the Artificial, described design as devising courses of action aimed at changing existing situations into preferred ones. By that definition, every AI deployment is a design act. Someone decides what the preferred situation is, what the system may do to reach it, and what it must never do.
The quality of those decisions largely sets the quality of the outcome. A capable model inside a poorly designed workflow produces confident errors. A modest model inside a well-designed one produces dependable results.
Technology without work redesign disappoints
The lesson is older than AI. In the 1950s, Eric Trist and Ken Bamforth of the Tavistock Institute studied British coal mines that had introduced new mechanised methods. Productivity often failed to rise, while absenteeism and stress increased. The technology worked. The organisation of work around it did not. Where teams were given back autonomy, and the work was redesigned to fit both people and machines, results improved. Trist called this a socio-technical system: technology and the way people work have to be designed together.
AI repeats the pattern. Organisations with strong models, promising pilots and impressive demonstrations often find that work does not change. The capability never reached operations, because workflows, roles and decision paths were not redesigned around it.
Co-intelligence, by design
The most effective AI systems are designed to work with human judgement rather than override it. Ethan Mollick calls this co-intelligence. In practice:
People define goals, values and success criteria.
Systems handle scale, complexity and repetition.
People step in where ambiguity, ethics or significant consequences arise.
Systems learn from outcomes, not only inputs.
This partnership drives good outcomes: deciding what to build, how to apply it, and when to trust it.
Values belong in the design, not the brochure
Bart J. Friedman's value-sensitive design argues that human values, such as privacy, fairness, accountability and dignity, should be identified and built into technology from the start, not discussed after it ships. For AI, that means explicit decisions: what the system will not say, which customers must always reach a person, which data it must never use, and how it explains itself.
Don Norman's work on human-centred design adds the user's view: systems should fit how people think and work, make their state visible and make errors easy to recover from. An AI system that hides its reasoning or makes mistakes hard to correct fails that test, no matter how capable its model.
Structure lets good design scale
Unstructured experimentation does not scale. In organisations, it turns into inconsistency and risk. For good design to hold at scale, it needs clear boundaries, defined responsibilities, feedback loops and governance.
Agentic system design provides that structure. Intelligence is broken into specialised components, each with a role, limits and clear handovers. The human decisions behind the system, its purpose, its values and its limits, are then applied consistently, thousands of times a day, rather than depending on who happens to be working.
Questions that shape an AI system's outcomes
What outcome is the system for, and how will we know it is achieving it?
What must it never do, no matter what it is asked?
Which values must it protect, and how will we test that?
Where must a person decide, and why?
How will people see its reasoning and correct its mistakes?
How will the workflow around it change?
How Mitochondria's agentic AI systems carry human design
The approach. Every deployment starts with the human decisions: what the organisation wants, what it will not accept, where its people must decide and how the work around the system should change. Those decisions are configured into each unit as boundaries, checkpoints and escalation paths.
What it enables. Systems that apply the organisation's judgement consistently, show their reasoning and bring the right decisions to the right people. The workflow is redesigned alongside the system, so capability reaches operations.
The impact. AI whose outcomes reflect the organisation's purpose and values, and work that changes because people and systems were designed together.
Read more about Mitochondria's approach.
Sources
Amabile, T. M. (1983). The social psychology of creativity: a componential conceptualisation. Journal of Personality and Social Psychology, 45(2).
Simon, H. A. (1969). The Sciences of the Artificial.
Trist, E. L. and Bamforth, K. W. (1951). Some social and psychological consequences of the longwall method of coal-getting. Human Relations, 4(1).
Friedman, B. (1996). Value-sensitive design. Interactions, 3(6).
Norman, D. A. (2013). The Design of Everyday Things (revised edition).
Mollick, E. (2024). Co-Intelligence: Living and Working with AI.
—
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 →