The real reason companies stall on AI

7 分7 min read

A map of what you need in order to say yes

The problem isn't the model

When an AI rollout stalls, the stated reason is usually technical. It rarely is. What's missing is grounds someone can approve on.

Getting a stalled project moving again doesn't take a stronger model or a longer policy document. It takes one thing: can you say what is visible and what isn't — and can you show, afterwards, that the line held?

This piece maps that line. By the end you can name which of the three reasons your organisation is actually stuck on, and what has to be in place before someone can sign off.

Separating audit from surveillance

If "audit" makes you tense, you're right

Conversations about AI governance slide, quickly, into watching employees. Who typed what. What they asked about. That's control, not trust. It also doesn't hold up in practice — tools built on the assumption that you're being watched stop getting used.

We don't sell surveillance. We sell verifiability.

Redefining audit

Audit does not mean looking inside. In this piece it means exactly two things.

Provenance — what was passed, and when, traceable after the fact. (What the AI's answer was based on is not visible to us. What we see is what crossed the boundary.)
A record that an action happened — who did what, when. The fact, not the content.

With those two, sign-off is possible. Nobody has to look inside.

The third path

What should be visible (the fact of an action; the provenance of what was passed) is visible.

What shouldn't be (content, private context) is structurally not visible.

And you can show, afterwards, that the line was there.

"We can't monitor it" doesn't get past risk review. "We can see everything" kills adoption on the floor. Only the third path clears both — draw the line, show the line, prove the line.

Saying the limits first

Our scope is limited to the following. Saying the limits first is the point, not a disclaimer.

The attribution decided at the boundary is never broken afterwards. Who something belongs to is declared by whoever passes it. We don't adjudicate that.
A person can say "that's wrong." Once they do, that observation stops being surfaced. The record itself isn't deleted.
What was passed, and when, remains. What the answer was based on is not visible to us.

We don't call this full-coverage DLP. There are other paths out, and we don't close them. Being able to call a partial answer partial is the only credibility that lasts in this area.

Sorting out what's actually blocking it

With that settled, work out where your organisation is stuck. It's almost always one of three, and they get argued about mixed together, which is why nothing concludes.

#	The reason	What's actually happening	What makes sign-off possible
1	Exposure	What we put in might end up outside	A line between what leaves and what doesn't — and proof of it
2	Provenance	We can't trace what was passed	Provenance — what was passed is traceable
3	Control	We can't draw the line, or can't show it	A record that an action happened (the fact, not the content)

How to use it: is your organisation stuck on 1, on 2, on 3, or all three? Take them one at a time, unmixed. Piling control measures onto a case that's really about exposure will not produce a yes.

Exposure often shows up as anxiety created by the absence of 2 and 3: "we don't know where it goes, so we don't put anything in." What's needed there is not a stricter ban, but a visible line.
Once provenance is in place, many stalled cases move suddenly — what the approver wanted was never reassurance. It was being able to explain.

Why now

Procurement is changing. The question is no longer "do you use AI" but "can you explain how you use it." A deployment nobody can explain now gets stopped from the contract side.
Third-party risk and continuity work the same way. Depending on an outside service assumes you can say what is going to it. A dependency whose contents are unknown can't be assessed.
This is not alarmism. We don't claim to have predicted any particular incident. The moment is context for why the standard is worth settling now — nothing more.

What you need in order to say yes

Sort the blocker into exposure / provenance / control.
Start with provenance — what the approver actually wants is explainability.
Keep the fact that an action happened, in a form that doesn't look inside.
Show the line — what's visible, what's structurally not, and how you'd prove it.

What stops AI isn't the technology. It's a line you can't prove.

FAQ

Isn't this the Semantic Web again?

No. The posture is the opposite.

The Semantic Web assumed that if meaning were defined up front, machines could work with it. That required annotating the world by hand, and the chicken-and-egg cost of that is where it stalled.

Sens doesn't define. It observes the topical structure of words already in use and records it.

Defining vs recording — the first says "this is what it shall mean." The second says "this is how it is in fact used."
Before LLMs vs after — the Semantic Web assumed symbolic reasoning. LLMs got past that assumption by another route. Sens doesn't compete with them; it works outside the model, on the context you hand it.
One correct meaning vs meaning per subject — the Semantic Web aimed at a single global semantics. Sens treats the fact that the same word means different things in different teams as itself worth recording.

This is not overclaiming. Newer isn't better. It's a different problem — the one the Semantic Web set out to solve (formal reasoning) is one Sens doesn't solve.

So can we monitor employees?

No. And not being able to is the specification. What remains is the fact that an action happened, not its content (see "Redefining audit"). Products that say "you can monitor them" are aiming at something else.

What's built, and what's still design?

Built (v0.5, a trial that runs in the browser): check what you are about to pass, before it goes. The record of what you passed stays in the browser and goes when you close it.
Design: export, and a record that survives closing, are not here yet. Sending it to an external AI automatically is design too — for now, you pass it by hand.
Design: extending the storage layer comes after v0.5.
We don't describe design as if it were built. An explanation that blurs that isn't grounds for approval.

How accurate is it?

We don't answer with a single number — claiming we could would be the dishonest answer. Different axes reach different depths, and averaging them loses the meaning. What we publish is provisional, and comes with the date it was measured and how.