Keep the Coworkers. Connect the Work. Measure the Outcome.

I came across this post from Ashwin Gopinath recently and it has been stuck in my head.

His basic argument is that enterprise AI has spent a lot of time trying to make individual people more productive (copilots, agents, assistants, coding tools, research tools, and so on) when a huge amount of organizational time is actually lost between people and systems.

Waiting for an approval.

Chasing someone for information.

Finding the latest version of a document.

Figuring out who owns something.

Reconstructing why a decision was made.

Waiting on another team before you can move.

The analogy is essentially this: we keep giving every runner in the relay faster shoes, while ignoring how much time we lose passing the baton.

That got me thinking about how our own thinking around enterprise AI has evolved.

I've written before about how hard it is to measure good AI use, and whether the business actually got better. This feels like the next question.

Adoption was the obvious first step

A year or two ago, simply getting meaningful adoption was hard enough.

Did people have access?

Did they know what the tools could do?

Were they comfortable using them for real work?

Were they going back after the novelty wore off?

For a lot of organizations, these are still completely legitimate questions.

But adoption has an obvious limitation as a metric:

So what?

If 70% of your company uses AI, so what?

If employees generate a million prompts this month, so what?

If weekly active usage goes up 20%, what actually happened to the business?

Usage tells you that something is happening.

It doesn't necessarily tell you whether anything important changed.

The next question is whether people are actually changing the way they work.

There is a big difference between someone occasionally asking AI to rewrite an email and someone genuinely collaborating with it.

The range of work people bring to it can expand.

AI use often starts with a transaction:

Draft this email.

Summarize this document.

Answer this question.

Useful, absolutely. And sometimes one good prompt is all the job needs. I've written before about getting that distinction wrong.

But people can also start collaborating with it:

Challenge my recommendation.

Find the holes in my analysis.

Help me think through the tradeoffs.

Look across all of this information and tell me what I'm missing.

Eventually, you get toward something closer to co-intelligence, where AI isn't simply completing tasks. It changes the boundaries of what one person can reasonably take on.

That's a very different kind of adoption.

But even then, there is another question.

Suppose someone becomes dramatically better at using AI.

Great.

What changed?

Did they get meaningful capacity back?

Can they manage a larger book of business?

Did client retention improve?

Are they selling more?

Did the quality of the work improve?

Did something that previously took three days now take three hours?

Did the economics of a role change?

Did a team or region actually begin operating differently?

This is where I think enterprise AI measurement needs to move.

Chats, users and tokens are useful diagnostic metrics.

They shouldn't be the end goal.

The real metrics should increasingly look like capacity, revenue, retention, cycle time, quality, client outcomes and operating leverage.

But Ashwin's post made me wonder whether even that is still one level too low.

Instead of only asking:

How much more productive did AI make this person?

Maybe we should also ask:

How much better did AI make the system around that person?

The work between the coworkers

Imagine a commercial insurance workflow.

One AI coworker gathers information about a property.

Another extracts and compares carrier quotes.

Another helps produce the client proposal.

Each one might be excellent.

And each might save a person a meaningful amount of time.

But the actual business outcome isn't:

The quote comparison took 20 minutes instead of two hours.

The actual job is something more like:

Submission → Quotes → Analysis → Recommendation → Proposal → Client decision

Now imagine the property research saves two hours.

Great.

But then the submission sits somewhere for three days waiting for information.

A quote comes in but nobody realizes a key field is missing.

Someone emails the producer for a decision.

The producer doesn't see it until tomorrow.

The proposal waits for another approval.

We may have made several individual tasks dramatically faster while barely changing the elapsed time of the overall process.

We gave the runners faster shoes.

The baton is still sitting on the track.

This is where things get much more interesting.

What if AI didn't just understand the task someone gave it?

What if it understood the state of the work itself?

Which properties are involved?

Which carriers were approached?

Which quotes have arrived?

What information is still outstanding?

What decisions have already been made?

Who owns the next step?

What was promised to the client?

What is blocking the workflow right now?

Then the first AI coworker doesn't really "finish" and hand someone an output.

It contributes to the state of the workflow.

The next coworker picks up that same context.

The proposal process knows when it has everything it needs.

And the system can notice that the entire thing has been stalled for 48 hours because one decision is still waiting on one person.

That's a fundamentally different model.

I don't think this means we should abandon task-specific AI assistants or coworkers.

Quite the opposite.

People understand jobs.

They understand:

Help me analyze this renewal.

Compare these quotes.

Prepare me for this client meeting.

Build the presentation.

That's a much more natural interface than asking someone to interact with an abstract "enterprise knowledge graph."

So I increasingly think about it this way:

The coworker can remain what users experience.

The client or workflow can become the context.

The business outcome can become the value.

That feels like an important distinction.

Today, a good AI coworker is often brilliant when a person brings it the right information.

Tomorrow, the more interesting version may already understand the history of the client, the decisions that have been made, the work currently underway and what needs to happen next.

Then we're not simply making people better at completing individual pieces of work.

We're making the process itself better.

The progression in my head now looks something like this:

1. Adoption

   Are people using AI?

2. Activation

   Are they using it in a way that meaningfully changes how they work?

3. Impact

   Is that translating into capacity, growth, retention, better client outcomes or some other measurable business result?

4. Flow

   Is AI making the organization itself move better, not just the individual?

That fourth stage is the one I hadn't articulated clearly enough before.

And I suspect it is where a lot of the really meaningful enterprise value will eventually come from.

Rent the intelligence. Own the context.

There's another idea from an earlier essay of Ashwin's that I particularly like:

Rent the intelligence. Own the context.

The frontier models are going to keep changing.

OpenAI will release something better. Anthropic will leapfrog them somewhere. Google will push Gemini forward. Some model we aren't talking about today will suddenly become relevant.

That's fine.

The underlying intelligence is increasingly something companies can rent.

What actually compounds is the context.

Your clients.

Your workflows.

Your decisions.

Your documents.

Your communications.

Your history.

Your institutional knowledge.

The relationships between all of those things.

Permissioned correctly, that becomes the layer every AI coworker can reason over.

And unlike model capability, it is uniquely yours.

This also changes the way I think about integrating AI into tools like email, messaging and collaboration platforms.

The opportunity isn't simply:

Great, now I can chat with the AI without opening another application.

Those are the places where work is actually happening.

Requests arrive there.

Decisions get made there.

Work gets handed off there.

Things get forgotten there.

Things get stuck there.

The really interesting future isn't when AI becomes easier to visit.

It's when AI increasingly understands how the work is moving and can help move it forward.

None of this means adoption suddenly stops mattering.

You can't get to organizational transformation if nobody uses the technology.

Flow still depends on people knowing when and how to use AI effectively.

These questions build on one another.

But I do think the north star keeps moving.

Use AI.

Then:

Work differently.

Then:

Create measurable business impact.

And eventually:

Redesign how the work itself moves.

The shorthand I've landed on is:

Keep the coworkers. Connect the work. Measure the outcome.

That feels a lot closer to where enterprise AI is ultimately going than another dashboard celebrating how many prompts people sent last month.