By Ryan Low, CBAP — independent business analyst, Digital Block FX

Your company is probably running a dozen AI assistants right now. Ask who approved what any of them did last Tuesday. In most companies, nobody can answer.

That gap has a name this year. People are calling it AI control, and it has quietly become a bigger argument than which model is best.

Two very different machines

Jarvis answers to one person. He lives in Tony Stark’s head, he is brilliant, and nothing he does is written down where a second person could read it. Take Tony out of the room and the system goes with him.

That is what most business AI looks like today. Clever helpers, one per person, each with its own instructions, its own settings, its own quiet reach into company data. They work. They just don’t answer to the company.

The Enterprise computer is the opposite, and it is boring on purpose. It answers any of the crew, at the level of authority that person holds. It says what it is doing. It keeps a record. And when the order is big enough, it stops and asks the captain first.

Nobody falls in love with the ship’s computer. That is rather the point.

I’ve drawn that line before, in Two Kinds of AI at Work. What has changed since is that the market stopped treating it as a metaphor and started shipping it as a product category.

The market just gave it a name

You will hear “control plane” and “harness” a lot this year. Both describe one layer sitting between your people, your data and the models, deciding what is allowed to happen.

Every major vendor is shipping one now. So let’s concede the obvious: having a control layer is no longer a difference. Everyone has one.

What is not settled is whether yours holds up when something goes wrong. Gartner put a figure on that in May. By 2027 they expect 40% of large companies to demote or shut down their self-running AI agents over governance gaps — gaps found only after a live incident. Not because the models were weak. Because nobody could prove what happened.

Rank, not a switch

Gartner’s advice is that one rule for every agent is the mistake. I would agree, and put it more plainly. A ship’s computer has ranks. Your AI should too.

Watching. It reads and reports. It changes nothing.

Advising. It drafts. A person decides.

Asking. It can act, once someone with the authority says yes.

Acting. It runs on its own — and this is the one that needs a brake.

Rotational Labs built a vendor-questionnaire review system on exactly that line. Two models score each questionnaire, and the client’s own analysts review only the cases where the two disagree. Rotational’s published figures put it at roughly 20,000 analyst hours a year taken out of the job, with the scoring coming out about 27% more accurate. The control was not a limit on the work. It was the reason the work could be trusted at that volume. Their case study is here.

Where the enterprise line falls

Three questions tell you which kind of system you have.

Can you show, months later, which agent did what, and on whose authority? Does your private data stop at a boundary you chose, before it ever reaches a model? And if the vendor vanished tomorrow, do you still own the work?

A consumer tool answers those with a screen. An enterprise system answers with a record.

---

The control layer I implement on is Endeavor, from Rotational Labs — prototyping, evaluation and runtime for agents, running in their cloud or inside your own environment, with every step, every record and every dollar visible as it happens. I am an independent business analyst and implementation partner, not a Rotational employee; I recommend it because it is what I would want running my own operation.

If you cannot say who approved your AI’s last decision, that is the place to start. Describe one workflow and I will show you what a governed version of it looks like — or book a call.

By Ryan Low, CBAP — independent business analyst, Digital Block FX

A purchasing manager recently raised a question I suspect many operations leaders are weighing right now:

“I have Claude Code on my desktop. It writes skills. If I want to automate vendor questionnaires, I can just ask it. Why would I pay you to break down my workflow?”

It’s a legitimate question, and the honest starting point is a matter of scope. For individual, low-stakes automation, an AI assistant-built skill is often the right tool — not every workflow warrants a formal engagement.

I explore that distinction more fully in a companion post — Two Kinds of AI at Work — which contrasts the personal “Jarvis” assistant with the Starship Enterprise‘s ship-wide computer: a tool built for one person versus a system the whole organization depends on.

But the question points to a real shift in the economics of automation, and that shift is worth understanding before deciding what any given workflow needs. Two years ago, getting an AI agent to competently draft a vendor-questionnaire response was the expensive part. Today an assistant produces a working skill in minutes. When the cost of writing the automation approaches zero, the cost that remains is everything the quick build didn’t address — and for workflows that matter to the business, that remainder is most of the total.

What the quick build doesn’t address

Evidence that it works. A skill that performed well on a handful of test cases has demonstrated an anecdote, not an evaluation. Does it hold at four hundred questionnaires a month? On malformed submissions? On the vendor who buries a liability clause in question 47? In practice, that gap tends to surface within weeks of real volume. Production systems earn trust through test suites and quality gates that run before anything touches a live decision — and continue running afterward.

An answer for the auditor. Procurement workflows touch vendor pricing, payment terms, and contract data. Where AI influences those decisions, three questions eventually arrive: What model saw that data? Who approved the access? Where is the record of what it recommended, and why? A personal skill on a personal machine has no answer to any of them. In a regulated business, that is the definition of shadow AI — less a productivity story than an audit finding in waiting.

Consistency across the team. One buyer’s well-crafted skill is a productivity gain. Twenty buyers, each running a private variant, amounts to twenty subtly different procurement policies operating in parallel — different vendor treatment, different risk tolerance, no shared baseline. The remedy is not to prohibit the skills; it is to promote the best of them into a shared, governed agent the whole team can see, use, and improve.

A lifecycle. Models update, behaviors drift, and skills regress quietly. Someone has to version them, regression-test them when the underlying model changes, and retire them when the process moves on. Software that matters receives lifecycle management. Skills that matter are software.

Independence. A skill written for one assistant lives inside that vendor’s product, on that vendor’s terms. The AI platform market is consolidating quickly, and product surfaces outside your control have a way of changing beneath you. The logic of a business workflow should not be bound to any single model vendor’s roadmap.

The decomposition is the point

There is one further thing the quick build passes over, and from a business-analysis standpoint it is the most consequential. Asking an assistant to “write a skill for X” automates what one person remembers about a process. Decomposing the workflow properly — with the people who own it — produces something categorically different: the tasks, the metrics, the test cases, and a measured baseline. Volume, time, cost, error rates. The before numbers.

That discipline is not bureaucracy. It is the only path to an after number — the figure that demonstrates whether the automation created value, and the one a CFO renews budget against. Industry research continues to find that the large majority of enterprise AI pilots never show measurable P&L return. In my experience the cause is rarely that the AI failed to work; it is that no one measured the before, governed the middle, or could prove the after.

Both tools, each in its place

To that purchasing manager, then: keep the assistant, and use it freely for individual work. When a workflow is shared across a team, touches regulated data, runs at meaningful volume, or feeds decisions someone must later defend, it has outgrown the personal skill. At that point it needs evaluation gates, governed endpoints, access controls, audit trails, and a measured business case. That graduation is what a governed platform is built for — in my practice, that platform is Endeavor by Rotational Labs, which I implement on as an independent partner.

The summary, if you keep only one line: an assistant can write you a skill; the questions that remain are whether it works, whether you can prove it, and whether it survives your vendor changing course.

If you’d like to see what a governed version of one of your workflows looks like, I’m accepting a limited number of design-partner engagements — three months, complimentary, with the first agent built alongside your team. Book a discovery call.