Organisational AI adoption
Share of surveyed organisations reporting AI use in 2025, as summarised in the 2026 AI Index; the 2025 edition reported 78% for 2024. Self-reported survey data, not a Swiss law-firm census or a productivity measure.
Stanford HAI, AI Index 2026, Economy chapter.
Consider an illustrative pilot: a team tests a tool, sees promising results and then struggles to integrate it into everyday work. The reasons need investigation; a demonstration does not establish the conditions for sustained use.
One possible obstacle is a gap between the pilot and its operating arrangements. Define the owner, permitted data, review, controls and evidence before extending access.
A 2024 meta-analysis of human and AI collaboration found that, on average, human and AI combinations performed worse than the better of humans or AI alone, with substantial variation between tasks. A successful demonstration therefore cannot establish the value of the whole workflow. Test the combined process, including review, and avoid applying an average from other tasks as your own expected saving.
Why pilots stall
1. The pilot tests the tool, not the workflow
A tool can perform well in isolation but fail inside a real process. The handoff, review burden, exceptions, permissions, and documentation may erase the time saved.
2. Success criteria are too vague
"Better", "faster", and "more innovative" are not enough. Define success before the pilot: time saved, error reduction, quality threshold, review time, user adoption, risk reduction, and decision speed.
3. Ownership is unclear
IT may own the platform. Legal may own risk. Compliance may own policy. Business teams own the work. If nobody owns the combined workflow, the pilot remains a demonstration.
4. Training arrives too late
When training follows tool selection, habits may already have formed. Training should be part of pilot design: what users may do, what they must not do, how to review output, and when to escalate.
The pilot-to-practice bridge
Ada Studio proposes six records for a controlled pilot; these are planning aids, not a validated acceptance test:
- use-case statement;
- workflow map;
- data rules;
- review standard;
- owner and escalation model;
- go / pause / redesign decision.
Pilot-to-practice checklist
0/6Think in operating rhythms
Possible routines to test include: weekly reviews, monthly governance meetings, onboarding, vendor review, policy refresh, training cycles, incident review, and workflow improvement.
This is where ISO/IEC 42001's management-system logic matters. Responsible AI is not a single approval. It is a maintained system for risks and opportunities.
NIST's AI RMF, a voluntary framework, also helps because it organises AI risk management into four functions: govern, map, measure, and manage. A pilot without measurement cannot become practice. A pilot without governance cannot scale responsibly.
A pilot asks: can the tool work? Practice asks: can the organisation work differently and still remain accountable?
What to do next
Choose one promising pilot and write a one-page operating model before expanding access. Name the owner, permitted data, review standard, training need, success metrics, and stop conditions. If you cannot define them yet, keep the tool at pilot scale until you can, however impressive the demo was.