Erik Shear

There is a quiet part of enterprise AI that does not get enough credit:

Services.

Not services as an add-on, but services as the place where strategy has to survive contact with reality.

That is where the questions get specific:

- What data can we actually use?

- Which teams need to change how they work?

- What risks are acceptable?

- What breaks when this leaves the demo room?

- What does success look like 90 days after deployment?

The model can produce an answer.

The enterprise still needs someone to turn that answer into architecture, process, controls, training, and measurable value.

The services layer is where AI stops being a promise and starts becoming an actual capability.

The AI trend I care about most is not a single model release.

It is the widening gap between what the technology can do and what organizations are ready to absorb.

The capability curve is moving fast.

The operating model curve is moving slower.

Teams have access to AI, but no clear rules for where it belongs.

They have pilots, but no path to production.

They have governance principles, but no day-to-day operating model.

The question isn't "Can the model do more?" It can.

The real question is: "Can the organization change fast enough to use it well?"

That is a leadership problem, and an implementation problem.

Your first internal AI tool doesn't need to be a platform.

Platforms attract roadmaps, committees, and invitations to meetings whose only output is another meeting.

Start much smaller:

Pick one person who already feels the problem.

Pick one part of their workflow they would happily stop doing.

Make that single step easier.

The early version can be ugly and manual behind the curtain. That's a feature. Doing steps by hand reveals where the process depends on human judgment.

Automate too early, and the confusion just runs faster.

Fake work is the most productive-feeling procrastination ever invented.

It uses all the same muscles as the real thing. It produces artifacts. It feels identical to progress from the inside. You can lose years in there.

The tooling I kept building was supposed to solve a discipline problem by making the work easier. But the work was never the hard part. Tool acquisition was just my escape hatch.

The bar has to be B+ and out the door. Timeboxed in one sitting.

The standard is useful and true, not the best thing you've ever written. If you aim for A+, you'll never ship.

If you are designing agentic AI systems for an enterprise, the loop has to be specific:

• What systems can the agent access?

• What actions can it take autonomously?

• What decisions are strictly out of bounds?

• What does failure look like, and who reviews it?

"Human in the loop" can't just be an afterthought or a comfort phrase. The real work is defining the boundary.

"Human in the loop" gets said a lot in AI conversations.

Sometimes it is useful.

Sometimes it is just a comforting phrase.

The real question is: which human, in which loop, with what authority?

A reviewer who cannot see the source data is not much of a control.

An approver who does not understand the model's failure modes is not much of a safeguard.

A manager who sees the issue three weeks later is not in the loop. They are reading the history.

Good AI design does not just add a person somewhere in the chain.

It makes the person's role clear enough to matter.

When you struggle with consistency, you only need two operating rules:

1. B+ and out the door.

One sitting, strictly timeboxed. The bar is “useful and true,” not the best piece you’ve ever written. Aiming for A+ is just another disguise for never shipping.

2. One metric: the streak.

Not subscribers. Not opens. Not shares. Binary: published-or-not.

The problem was never "can’t create." It was "can’t sustain." Count the streak.

A lot of creative avoidance doesn't come from a fear of failing.

Failing is survivable. We've all done it.

The real fear is the unironic attempt. The full, sincere, no-jokes-attached try.

Because as long as you're over-tooling or keeping things ironic, you have an excuse in your back pocket. The moment you give it an unvarnished attempt, you have to find out what you're actually capable of.

Most companies do not have an AI access problem anymore.

They have an adoption problem.

The tools are available. The demos are impressive. The pilots are everywhere.

But the hard part starts right after someone says, "This is interesting."

Who owns the workflow?

What changes in the process?

How do we know the output is good?

What happens when the model is wrong?

Who trains the team?

Who supports it after the first excited group moves on?

AI does not become useful because someone got access to a tool.

It becomes useful when it is tied to a real workflow, a clear owner, a measurable outcome, and a support model people can actually live with.

The next phase of enterprise AI will not be won by the teams with the most pilots.

It will be won by the teams that can turn a pilot into a daily habit.

A chatbot answers. An agent acts.

The moment a system can take steps across tools, files, emails, or code, output quality is no longer your only problem.

Enterprise AI agents need actual job descriptions:

What is the agent responsible for?

What systems can it access?

What actions can it take?

What decisions are strictly out of bounds?

When does it require human approval?

The better question isn't just "Can it do the task?"

It's "What job are we willing to let it have?"

52 years old and a long way from retirement

3 hard truths I had to accept about retirement.

- Traditional retirement advice was not going to cut it.

- There is no such thing as stability in corporate America.

- I had to create alternative sources of income quickly.

"Interdum non nisi per ignem progredimur"

Sometimes the only way forward is through fire