Stop Measuring Token Usage. Start Measuring Work Velocity

2026-05-26

Have you ever wondered why companies like Ramp seem to be getting so much more leverage from AI than everyone else?

Most companies think their AI bottleneck is adoption. It is not. The real bottleneck is the process.

To solve for adoption, companies give employees access to ChatGPT, Claude, or Codex. Then they track usage, report activity, and often, nothing meaningful changes.

In fast-growing companies, work slows down in the gaps: the handoff nobody owns, the approval missing context, the customer follow-up that waits three days, the campaign stuck because the data is not clean, the commitment made in a meeting and never tracked.

AI changes what companies can do about this. For the first time, they can start to continuously audit how work moves, identify where velocity is leaking, and fix those bottlenecks one by one.

Ramp seems to understand this. Ramp is hiring AI Product Operators whose job is not to evangelize AI, but to change how internal teams work. Box is making a similar bet with AI Business Automation Engineers focused on finance, legal, people ops, and other internal teams.

The new operating loop is simple: audit the process, identify the highest-leverage bottleneck, automate or augment the work, measure the cycle-time improvement, and repeat.

History does not repeat but it sure does rhyme

A close precedent here is Celonis. In the RPA era, process mining helped companies see how structured workflows actually moved through systems like ERP and CRM. Once the process was visible, automation platforms could remove repetitive steps.

But that model was constrained by the shape of the work. It worked best when the process left clean event logs, followed stable paths, and could be scripted.

AI expands the surface area. The next process map will include Slack threads, meeting notes, CRM fields, emails, approvals, customer conversations, handoffs, and commitments. It will not just show where work went. It will recommend where velocity can be created.

This messy middle is where the next wave of automation will happen: not just inside clean systems of record, but across the break points where work actually slows down.

What does an AI operator do?

AI operators are the first version of the team that runs this loop.

Simply put, they map the work, decide where AI belongs, keep humans in the loop where judgment matters, and prove whether the change helped.

This role sits between product, engineering, and operations.

A PM may understand the user problem but live too far from the daily workflow. An engineer may know how to build the system but maybe too much in the weeds of the system. An ops manager may know the process cold but lack the tools to instrument or evaluate AI-assisted work.

The AI operator blends those instincts: operator, product thinker, automation builder, analyst, trainer.

The value compounds

The first win is one process - a support queue gets triaged faster or a customer issue stops bouncing between teams because the context is finally in one place.

Good. But the bigger win comes after that.

As the operator rebuilds a process, the team learns what to delegate, what to review, and how to measure whether AI is actually helping. People see AI inside the work instead of beside it.

What compounds is not just morale. The company builds reusable process maps, integrations, evals, permissions, escalation rules, templates, and measurement habits. The second process gets easier because the first one created infrastructure. The tenth process gets easier because the company now knows how to change work.

How to start

Do not start with a 12-month AI transformation roadmap. Start with one bottleneck where speed already matters.

Pick a process with a visible owner, frequent volume, measurable delay, and enough business impact to matter.

Give an AI-native operator ownership of it. Their mandate is simple: map the current state, identify the highest-leverage bottleneck, ship one intervention, measure the delta, and make the lesson reusable.

Then capture what changed: which data was needed, which handoff broke, which rule mattered, which metric improved, and what should be fixed next.

AI-native companies will not be the ones with the highest token usage. They will be the ones whose work moves faster every month.