AI adoption, agents, and the architecture underneath them — written from production systems, not demo stages.
Twenty-five years of automating the same kind of work across three technology generations — SharePoint workflows, the Power Platform, and now agents. The tools changed three times. The hard part never did.
AI adoption is a pyramid, and it's steeper than the hype suggests: a commodity base everyone can replicate, a middle of one-task-at-a-time usage, and a thin orchestration tier where the economic story actually lives.
Blank-slate discovery feels thorough and is the least efficient way to learn what someone needs. People are bad at describing what they want — and instantly, ruthlessly good at 'no, not that.'
We've watched teams build brilliant orchestrators on top of terrible databases — the agent worked, and that was the problem.
"If I have to check everything the agent does, what's the point?" — the most honest objection in enterprise AI resolves at design time, not run time.
A clinical research lead asked whether a cloud was following her around collecting everything she did — the honest answer was yes, and the fix is three architecture moves, not a model upgrade.
The gap between impressive agent demos and production systems is not a model problem — it's the architecture around the agent, and it's why we named the firm Fixed Point.
Using Claude Code and the Power Automate MCP server to debug flow failures as data instead of squinting through the designer — and why the authoring platform and the operating platform are decoupling.
The Model Context Protocol just got its biggest overhaul since launch. Here's what it means for the people who sign off on agents — not the people who build them.
The Army figured out delegation long before the AI industry did. What a career Colonel's mission planning teaches about working with AI: state the outcome, the constraints, and why it matters — then let execution survive first contact.
Almost every business system overwrites its own memory — and agents learn from history, not current state. Why the audit for agent-readiness starts with finding the fields that overwrite.
A security incident that took 17,000 agent actions to reconstruct is not a model problem. What the Open Secure AI Alliance launch actually teaches about autonomy in regulated environments.
The honest workflow behind the disclosure line: what AI does in my writing, what it never does, and why the judgment layer stays human.
The most valuable knowledge in your organization is why things get rejected — and almost nobody writes it down. How a reviewer's rejection taxonomy became the spec for a clinical intake agent.
The most cautious regulator in the country put AI agents inside its own review process — and the lesson isn't that they did it. It's the five architecture decisions they made first.
One running file of everything that ever broke — what broke, what everyone believed until it did, and what would have caught it. And the second job AI just gave it.
Almost every real process follows a power law — a handful of cases drive the volume, and the long tail is noise dressed up as requirements. Build for the vital few, route the tail to a human, and ship.
Twenty-five years of automating the same kind of work across three technology generations — SharePoint workflows, the Power Platform, and now agents. The tools changed three times. The hard part never did.
The Forward Deployed Engineer optimizes for make-it-work-here-now. The architect optimizes for make-it-hold-everywhere-later. Most org charts have a slot for the first and nobody accountable for the second.
AI adoption is a pyramid, and it's steeper than the hype suggests: a commodity base everyone can replicate, a middle of one-task-at-a-time usage, and a thin orchestration tier where the economic story actually lives.
Architects want the mechanism. Executives want the outcome. Same idea, completely different proof — and a two-minute AI workflow for rendering it twice without rewriting it from scratch.
Every organization runs two processes: the documented one and the actual one. The gap between them is where the truth lives — and where AI earns its place in discovery.
Most of the Fortune 500 is already running agents nobody mapped. Agent sprawl is the identity-sprawl story replayed at machine speed — and the fix starts with one row per agent.
Before a decision that matters, don't ask AI to help make your case — ask it to tear your case apart. The habit that finds the weak spot while there's still time to fix it.
The most expensive agent is the one you didn't need to build. The question was never whether a task can be an agent — it's whether a real decision is being made, or steps are just being executed.
Companies bolted agents onto an org chart built for people, and the chart can't hold them. The entry point of every workflow just moved from a person to software — and the experience pipeline has to be rebuilt on purpose.
The quality of what you get back almost never comes down to how cleverly you phrased the request — it comes down to how much the model knew before you asked. The shift from prompt engineering to context engineering.
Blank-slate discovery feels thorough and is the least efficient way to learn what someone needs. People are bad at describing what they want — and instantly, ruthlessly good at 'no, not that.'
Skills activate by description match — which makes overlapping descriptions nondeterministic routing, not redundant documentation. Skill sprawl is invisible until two playbooks disagree mid-conversation.
The biggest jumps in output didn't come from new tricks — they came from restraint. Five subtractions that beat every prompt hack.
JPMorgan is deploying agents that run for hours with no human in the loop. But runtime multiplies early mistakes rather than earning trust — and coherence is what makes the wrong answer convincing.
Walmart paid its AI executive more than its CEO, then reported shoppers spending 35% more through its agent. The tell isn't the checkbook — it's the structure: four agents organized by who they serve, orchestration underneath.
The agentic conversation in clinical research aims at the wrong target. The agent that earns its keep isn't the one that runs the study — it's the one that notices the missing signature before the inspector does.
The Novo Nordisk–OpenAI partnership is being read as a model story. It's an infrastructure story — messy clinical data, GxP traceability, and an integration layer nobody designed for an AI to orchestrate across.
One in twenty production AI requests fails silently — the output looks right and nothing trips an alarm. Why the eval and observability layer is the part every ROI deck skips.
Most skills get tested by their builder, on curated examples, in the week they were built — with the builder quietly doing half the skill's invisible job. The real test is surviving the handoff.
The first workflow most teams automate is the visible one — and it's the wrong one. The right first candidate happens constantly, describes itself in fifteen minutes, and someone quietly resents doing it.
Five things that break first when an AI agent moves from demo to real work — and the five questions to ask before you sign anything.
The most misleading word in agentic AI: a skill isn't a capability the AI has — it's a workflow you've documented so something else can run it. The bottleneck was never model access. It's articulation.
Most people use AI like a really smart search bar. The value is in tools that work while you're doing something else — like a project health dashboard that's already current when you open it Monday morning.
AI adoption, agents, and enterprise data — founder-led, fixed-scope, built to hand off.