top of page

Three Lessons from AI Strategy & Practice 2026

Writer: Jovanca Garnadi
Jovanca Garnadi
Oct 27
2 min read

What does it take to put AI into practice—and keep it working?


That question brought engineering, business and governance practitioners together at Ham Yard Hotel in London for the first AI Strategy & Practice Conference.


Across the fireside conversations, panels and discussions between sessions, three themes stood out for organisations moving AI systems from experimentation into production.


1. Evals help teams define success


Evaluation should not be treated solely as a final check before deployment.


Teams can use evals throughout development to define acceptable behaviour, measure whether an AI system is creating the intended value and identify where it needs to improve.


Testing remains important after deployment, too. Models, tools, prompts and operating environments change. A system that behaved correctly at launch may not continue to meet the same requirements indefinitely.


Continuous evaluation helps teams detect when performance starts moving outside the required boundaries.


2. Internal expertise is essential


Successful AI deployment requires more than technical capability.


AI systems need to fit the workflows, responsibilities and practical constraints of the people who use them. This makes operational and domain expertise a critical part of designing, evaluating and improving an AI application.


Internal experts understand where existing processes work, where they fail and what a useful replacement would need to achieve. They can also help define the edge cases, policies and outcomes that technical teams need to test.


AI may introduce new capabilities, but the people who understand the organisation’s real work remain essential to determining whether those capabilities are valuable.


3. Deployment and governance need to prepare for scale


Current technology-governance processes often assume that an organisation will deploy a manageable number of systems each year.


AI and agentic systems could change that significantly. Organisations may need to manage many more deployments, while individual systems are updated more frequently after launch.


Testing, approval and monitoring processes will therefore need to scale without becoming superficial or creating an unmanageable volume of manual reviews.


This means thinking beyond whether one system can be governed successfully. Organisations also need to ask whether their processes will continue working across hundreds or thousands of changing AI systems.


What attendees said


“Excellent, practical & enjoyable. Great venue and great crowd.”

— Zach, Waggel

“What a superb event—I had some brilliant conversations. Would absolutely recommend.”

— Kyra

“The speakers and sessions covered highly relevant, timely topics and offered practical, actionable insights.”

— Conference attendee


From discussion to practice


These lessons connect closely with our work on Spec27.


Putting an AI agent into production requires teams to define how it should behave, translate those expectations into repeatable tests and maintain evidence as the underlying system changes.


That includes expected-use testing, edge cases, policy requirements and adversarial evaluation—not simply checking whether an agent produces a convincing answer to a small number of prompts.


Spec27 helps teams create reusable behavioural specifications and use them to test AI agents systematically across development and deployment.



For more lessons, photographs and reflections from the conference, read the complete Safe Intelligence recap:



Did you attend?


If you joined us at Ham Yard Hotel, we would still love to hear what worked for you and what we could improve.


bottom of page