Why most Copilot pilots never become operating systems
The technical demo is rarely the hard part. Production adoption depends on permissions, content ownership and an operating model that survives the launch meeting.
Articles
Things I have learned building enterprise AI systems, cloud platforms and software products.
AI Observability Is More Than Token Counts
Useful observability connects requests, evidence, decisions, actions and outcomes without retaining data indiscriminately.
Microsoft Copilot Isn't Magic. It's Architecture
Copilot experiences are shaped by identity, permissions, retrieval and integration architecture long before a prompt reaches the model.
Why Prompt Engineering Won’t Save a Broken Workflow
Prompts can shape model behaviour, but they cannot repair missing evidence, unclear ownership or wasteful handoffs.
Why Most Enterprise AI Projects Fail (And How to Avoid It)
Enterprise AI projects fail when compelling demos are mistaken for operational systems. The remedy is a narrow outcome, owned data and production evidence.
AI Governance Doesn't Have to Kill Innovation
The best AI governance creates clear, fast paths for experimentation while reserving heavier controls for decisions that can cause real harm.
Choosing Between SharePoint, Dataverse and SQL for Power Apps
Choose storage from required behaviour, security and ownership, not familiarity alone.
What Production-Ready AI Actually Looks Like
Production readiness is not a model benchmark. It is the ability to operate, explain, recover and improve a complete AI system.
Stop Measuring AI Adoption by Licence Activation
Activation and prompt counts measure access and curiosity. Leaders need evidence that AI changes valuable work.
AI governance needs a control plane, not another committee
Policies become useful only when they are translated into enforceable decisions across identity, data, models and runtime behaviour.
Your AI Strategy Is Probably Just a Vendor List
A sequence of product purchases is not a strategy. Real AI strategy makes choices about advantage, operating capability and where not to invest.
The Hidden Architecture Behind Every AI Assistant
The conversational interface is the visible edge of a system spanning identity, retrieval, policy, tools, telemetry and human escalation.
Power Platform Business Continuity Starts Before the Maker Leaves
Continuity comes from designed ownership, deployability and recovery, not emergency transfer during offboarding.
The Real Cost of Enterprise AI Is Operational
Model tokens are visible, but integration, verification, support, data ownership and change management dominate many enterprise AI costs.
AI Agents Need Boundaries, Not Better Prompts
Reliable agents come from permissions, validation and controlled execution, not increasingly elaborate instructions.
Why Every AI System Needs a Refusal Strategy
A useful refusal is a designed product outcome that protects users and moves uncertain work somewhere responsible.
Why Your Copilot Rollout Has Become a Content Audit
Copilot exposes years of unresolved ownership, permissions and lifecycle decisions because its answers can only be as trustworthy as its sources.
Enterprise AI Fails at the Handoffs
The model may perform its task well while the wider system fails between people, tools, decisions and exception queues.
Retrieval-Augmented Generation Is a Data Product
RAG succeeds through owned sources, measurable retrieval and disciplined lifecycle management.
The Architecture of a Trustworthy AI Workflow
Trust emerges from evidence, bounded actions, explicit decisions and recoverable execution across the whole workflow.
The Case Against Building an Enterprise Chatbot
A universal chat interface often shifts the work of understanding systems onto users. Focused workflow products usually create more value.
Your Data Is Not Ready for AI, and That’s Fine
Data readiness is not a binary gate. Start with the sources and quality level required by one valuable decision.
The Power Platform Centre of Excellence Starter Kit Is Not a Governance Strategy
The CoE Starter Kit provides useful telemetry and workflows, but it cannot decide what your organisation should govern or why.
Your Power Platform Estate Is Already a Software Portfolio
Power Platform workloads deserve portfolio decisions about investment, ownership, duplication, risk and retirement.
AI Agents Should Earn Their Permissions
Agent autonomy should expand through evidence, with permissions tied to demonstrated reliability and consequence.
What Good Power Platform Governance Looks Like in Practice
Good governance combines clear boundaries, proportionate controls, usable delivery paths and accountable service ownership.
Human-in-the-Loop Is Not a Complete Safety Strategy
Human review only reduces risk when reviewers have authority, evidence, time and a realistic chance of detecting failure.
The Power Platform Environment Strategy Most Organisations Actually Need
A practical environment model separates personal productivity, team experimentation and lifecycle-managed production.
Stop Asking Whether the Model Is Accurate
Accuracy without a defined task, distribution and consequence is not a useful production measure.
Copilot Can’t Fix Your Information Architecture
Copilot can improve access to knowledge, but it cannot decide which source is authoritative or who should own it.
The Graduation Path Every Power Platform App Needs
Workloads need a predictable route from personal experiment to supported business service.
Ownership Is the Hardest Power Platform Governance Problem
Technical ownership, business accountability and connection identity are different responsibilities that inventories often collapse.
A Risk-Based Governance Model for Power Platform
Risk-based governance links observable workload characteristics to proportionate platform and operating controls.
Power Platform Solutions Should Be the Unit of Ownership
Solutions align ownership with deployment, dependencies and the business change being delivered.
A Cloud Flow Is Not a Business Process
A flow implements part of a process; it does not contain the ownership, exceptions and outcomes that make the process operable.
Delegation Is an Architecture Constraint, Not a Warning to Ignore
Non-delegable formulas can silently produce incomplete results, making data-source choice and query design correctness issues.
Why Long-Running Power Automate Flows Become Operational Risks
Extended flows accumulate state, dependency and recovery problems that run history alone cannot manage.
Power Automate Needs Idempotency
Retries and duplicate triggers are normal; flows must prevent the same business action from happening twice.
A Successful Flow Run Does Not Mean the Business Process Succeeded
Technical completion and business outcome require separate states and measures.
AI Governance Should Be an Engineering System
Governance becomes effective when policies compile into controls that can be tested, observed and evidenced in running systems.
The useful constraint in edge-first SaaS architecture
Cloudflare’s runtime rewards explicit boundaries. That constraint can produce simpler systems when state and background work are designed deliberately.
The Dangerous Gap Between Connector Access and Data Authorisation
A permitted connector does not mean the user or flow should access every operation and record behind it.
The Default Environment Is Not a Development Environment
The default environment is a shared productivity space, not a safe foundation for lifecycle-managed business applications.
Data Loss Prevention Is Not the Same as Data Governance
DLP limits connector combinations; data governance defines meaning, ownership, access, quality and lifecycle.
Power Pages Is a Security Boundary, Not Just a Website Builder
External identity, table permissions and web roles require deliberate design because the audience sits outside the tenant.