AI programmes fail as often from a lack of evaluative discipline as from technical missteps. Action research offers a rigorous, iterative structure that turns a change initiative into something you can defend to a board — or an auditor.
A conventional project plan assumes the destination is known. AI adoption rarely works that way. The tools change under you, the workforce reaction is not predicted by the business case, the promised benefits materialise in unexpected places and fail to materialise in obvious ones. A method that assumes the destination is fixed treats every surprise as a project management failure.
Action research assumes the opposite. It is a disciplined way of learning by doing, in cycles of plan, act, observe and reflect (Lewin 1946; Reason and Bradbury 2008). It is the natural method for enterprise AI because AI adoption is genuinely exploratory: you cannot know what your users will actually do with Copilot until you give it to them, and you cannot know whether your governance is fit for purpose until it meets real cases.
1. Plan. Establish a baseline. In an AI programme this means a pre-implementation survey of attitudes, confidence, existing shadow AI use and concerns; a mapped view of the data estate the tool will touch; and a documented set of hypotheses about what the intervention is expected to change.
2. Act. Introduce the intervention — training, tool access, governance forums, a community of practice — with clear scope and named accountability. The intervention is not just the technology; it is the whole arrangement of technology, training, communication and support.
3. Observe. Collect evidence during and after the intervention. Survey the same population that gave the baseline. Interview users. Capture the artefacts produced. Log the incidents, the near-misses and the workarounds.
4. Reflect. Interpret the evidence, ideally through a theoretical frame that lets you explain patterns rather than just describe them. Socio-technical systems theory and actor–network theory both do useful work here because they treat outcomes as emerging from the interaction of people, tools, data, policies and governance artefacts, not from any single actor (Trist and Bamforth 1951; Latour 2005).
Most organisational AI research is done by insiders — the person leading the programme is also the person studying it. This creates access advantages and discipline risks in equal measure. The advantages are obvious: proximity to decisions, access to artefacts, ability to observe what actually happens rather than what people report happening.
The risks are less obvious and more consequential. Bias, over-familiarity, an unconscious drift toward celebratory accounts of the programme's own success, and power imbalances that can suppress dissenting evidence from participants (Costley, Elliott and Gibbs 2010). The remedies are procedural, not attitudinal.
ISO/IEC 42001 requires evidence of stakeholder engagement, documented decision rationale, evaluated outcomes and continual improvement. The EU AI Act requires post-market monitoring for high-risk systems. Both are, in effect, asking for the artefacts an action-research cycle produces as a matter of course (ISO/IEC 2023; European Parliament 2024).
Running your AI programme as action research from the start means the evidence a reviewer wants is generated by the programme itself, not manufactured afterwards. The baseline survey becomes your stakeholder engagement record. The intervention documentation becomes your decision rationale. The post-implementation evaluation becomes your effectiveness evidence.
ISO-STANDARD.app gives you the register, the engagement log and the evaluation records action research produces — ready for ISO 42001, the EU AI Act and your own board.
ISO-STANDARD.app ships a ready-to-adopt ISO 42001 workspace with the risk register, controls catalogue, policies and audit-ready exports already wired together — no spreadsheet sprawl, no consultant lock-in.
Prefer a conversation? Email hello@iso-standard.app — a real human responds within one business day.
