The AI Gap Is No Longer Access. It’s Conversion.
The real divide is not who has AI. It is who has turned AI into a repeatable operating system for work.
Frontier firms are not getting different models. They are connecting the same intelligence to more reusable instructions, more context, more tools, and more repeatable work. I see the same thing in my own commercial role. The breakthrough was not writing better prompts. It was replacing prompts with operating procedure: named, reusable skills that check the math, enforce what can and cannot be claimed, stage work from CRM context, and hand complete inputs to the next agent. The gap is no longer access. It is conversion.
The number matters less than what sits behind it
As of June, OpenAI reports that its highest-usage enterprise customers generated 8.3× as many output tokens per active user as typical firms, up from 2.6× in January.
That is not an ROI metric. OpenAI explicitly calls token volume an imperfect proxy for depth of use.
The more interesting numbers sit next to it.
Weekly plugin use was 21% among frontier firms versus 9% among typical firms. Skills, which let teams preserve reusable task instructions, were used by 19% versus 3%.
I care more about 19% versus 3% than 8.3×. One measures more output. The other points toward reusable work.
I stopped writing prompts and started writing operating procedure
This is the part I rarely see in adoption articles.
I work in a regulated commercial environment. The system cannot simply be clever. It has to know what it is allowed to do, what it must verify, what it must refuse, and what still requires me.
So I stopped treating every task like a fresh conversation.
I started turning repeated work into named, reusable instructions.
The difference shows up in cadence. This is not something I do on a good day. The skills run on every draft. The sweep runs every morning. The checks run whether or not I remember them. AI is at the table every day because the procedure puts it there, not because I decided to invite it.
A quote-building skill
When the system helps prepare commercial material, it does not get a generic instruction to “make a quote.” The skill checks arithmetic, preserves the approved structure, separates assumptions from confirmed terms, and flags anything that still needs verification.
An outreach-staging skill
When the system prepares outreach, the job is not simply to write better copy. Every draft goes through the same procedure: check the suppression list, remove anyone who opted out, strip any claim that is not supported, and flag CRM context that looks stale.
The output is a staged set, each draft carrying its verification status, and one line at the bottom that never changes: sent, zero. Waiting for go.
A handoff skill
This one came from failure.
Early multi-agent workflows regularly handed the next agent a reference to upstream work instead of the actual inputs. The result was predictable: “What’s my list?” “Where is the source?” “Which records am I supposed to use?”
So the rule changed.
A handoff is not a pointer. A handoff is a complete input package.
If another agent needs a list, the list travels with the assignment. If it needs the decision criteria, those travel too. If something is unresolved, that is stated explicitly.
That one rule eliminated a surprising amount of agent-to-agent confusion.
The constraint layer is most of the work
Conversion is usually described as connecting AI to more tools.
That is only half of it.
The harder work is encoding what the system must not do.
- Do not include people who have opted out.
- Do not convert a hypothesis into a fact.
- Do not make a competitive or clinical claim that has not been verified.
- Do not treat stale CRM context as current simply because it is present.
- Do not send anything externally without my go.
That last line is the practice. I am deliberately not turning the approval gate into an architecture diagram here. The important point is simpler: the system can prepare aggressively, but nothing consequential leaves without a human decision.
The CRM is useful because it is context, not because it is CRM
I work from CRM context every day, even when the handoff into AI is not fully automated. HubSpot’s agent documentation shows where this is heading: give an agent CRM context, let it prepare the work, then use the result inside the workflow.
Real usage is messier.
Records can be stale. The most important relationship context may live in a note rather than a field. A deal stage can look precise while the actual account is ambiguous. A previous message can be technically accurate and strategically outdated.
So the agent does not get to treat the CRM as truth.
It gets to treat the CRM as evidence.
That distinction changed the workflow.
The system now separates what is verified, what is historical, what needs a fresh check, and what is simply my current hypothesis.
Only then does it recommend the next move.
Work is changing faster than jobs
That is Revelio Labs’ framing, and it is the one worth keeping.
The labor-market picture is more complicated than a single adoption statistic. Revelio’s research has found slower employment growth in occupations with greater AI exposure, with the pressure especially pronounced among younger workers. At the same time, firms that actually adopt AI have shown stronger headcount growth than non-adopters.
Those two directions can coexist.
AI exposure can put pressure on particular tasks and roles while firms using the technology reorganize work and continue growing.
That matches what I see at the level of the job itself.
My title did not change because I started using agents. The work underneath it did.
Research arrives differently. Account preparation arrives differently. Repetitive checks happen on a cadence. Drafts come with verification status. More of my time moves toward judgment, relationships, strategy, and the conversation itself.
The next conversion test
I am not trying to add more agents.
I am testing which repeated pieces of commercial work deserve a durable operating procedure and which are still too contextual to encode cleanly.
That is the conversion question I care about now.
Not “Where can I use AI?”
“Which parts of this job are stable enough to become a system?”
Signal documents this work as it runs. The next issue is about what the system watches, and what it is told not to.
Subscribe to Signal