Checking In

I’ve been with OpenAI for roughly four months. I probably should have written this a long time ago.

I’ve been using this blog to crystallize my thinking. There are a few things that have been on my mind since I’ve been here, and I want to start writing them down.

The first is just how fast this space is moving.

I work with a lot of life sciences customers, and I don’t think most organizations understand how fast the world is changing around them. They are using these products and surfaces the same way they did maybe a year ago. They don’t see that, maybe every month, the landscape has shifted underneath their feet.

While they have been hyper-focused on governance and restrictions, I think they have missed the boat on the rate of change itself. I honestly spend a lot of time with customers reviewing features that are obsolete the second we complete the review.

Connectors

The clearest example of this, and the one I struggle with most, is connectors.

A number of my customers have not enabled connectors, data access, or tool access. To me, that is a massive missed opportunity. Any organization today that does not have a clear connector strategy is completely missing the boat on where we are with AI.

The agentic revolution is old at this point. We are well past agents being the focus. Agents are only made powerful by the tools and data you provide them. I have customers that still do not have connector capability enabled, and honestly, I see them falling significantly behind.

The legacy IT function

That takes me to another point. I see a massive disruption coming for the legacy IT function.

Historically, IT has served as a gatekeeper. It has managed vendors and centrally controlled technology. I would encourage anyone starting their career in tech to avoid that function at all costs, because in my opinion the AI revolution is making technology accessible to far more functions.

I see a lot of IT functions trying to centrally control and centrally restrict. Going back to the point on connectors, I think they are over-governing without real logic behind the governance. A lot of business units are frustrated with IT functions that are standing in the way of real value creation.

This has always been the case to some extent. The AI moment is just exacerbating it.

Value creation

I spend so much of my time with customers arguing about how to measure value for these capabilities. It gets discussed tremendously in the industry. I see so many people saying that it is really hard to measure the value of AI. I honestly cannot understand why it is as hard as most organizations make it out to be.

We give clear capabilities to understand who is using the surface, how they are using it, and how much the capability costs. Not just OpenAI. Every provider gives you some mechanism for this.

There are a few layers to the value discussion. One is a central function trying to understand the value of something. This goes back to IT. IT signs the bill for AI products and then has to explain the value. That is one model.

The other, and the one I think is more sustainable, is large-scale chargeback. The business unit understands the value of the tools it uses. The business unit funds its individual consumption. IT does not need to worry about articulating the value because the business unit generating the actual business change and profit is the one saying: yes, I need this, and I am willing to fund it. That is the clear indication of value in itself.

Even where there is a central function funding these capabilities, it is still possible to articulate value. Some customers are more mature than others in having the apparatus to collect user-usage information and attribute it back to a value story.

Software engineering is a good example. We have been tracking it as an industry for so long. We can put the products in developers’ hands, measure their usage, and measure product throughput. We can look at DORA metrics, velocity metrics, and quality outputs. Any organization that says it cannot measure value should be asking itself different questions.

The next question is value capture. Once you know your throughput has gone up, what do you do with that increased velocity or the shorter release cycles?

Do you ship more features? Do you reorganize? Do you start new ventures, product lines, or investment areas that you would not have invested in before? That is not up to product developers. It is not up to OpenAI, Anthropic, or Cursor. It is up to leaders in these functions to develop a cohesive value-capture narrative and present it to leadership.

The idea that this cannot be done is laughable to me.

Leaders have to be close to this

Most leaders have had the luxury of being a little disconnected from technology. I do not think they have that luxury in this particular technological revolution.

You have to be close to these capabilities. You have to be hands on. It blows my mind how many leaders I interact with who clearly do not know the capabilities, and not just the products I am working with customers on. A lot of leaders do not seem to know where the space is at.

I understand that part of my role is to explain the value of the capabilities I have to the leaders I work with. I am not saying they should already know everything, or that I should not have to make the case. What stands out to me is the clear bifurcation in the leaders I interact with.

The first type is not just hands on. They are using these capabilities, they are aware of where the industry is, and because they are using them, they immediately understand the value. They are pressing their teams to start using these capabilities at scale and exploring them internally. They will call me and say, “Carlo, why don’t we have this capability enabled in our environment? I’m using it at home. It’s ridiculous. Help me figure it out. Set the rollout plan.”

The other type spends most of their time reading about these capabilities instead of using them. They overemphasize ambiguous strategy points without knowing what the technology can actually do. They have grand visions about strategy, governance, and control, but it is immediately apparent that they have never really used the capability.

They do not know what exists today. They do not know what governance controls are already available. They raise security or governance fears that have already been addressed. It is very evident when you talk to them.

This is not an OpenAI employee saying this. It is a technologist saying it. When I sit down and look at these capabilities, it is completely obvious to me that everybody across an organization should be using them.

The personal workflow gap

There are a lot of tools and capabilities today for deeply embedding AI into an individual’s workflow. But let me use a practical example.

If I wanted to build a Carlo digital twin today, it would not be enough to connect Gmail, Jira, text messages, WhatsApp, my personal notes, and my local directory. I use all of those things, and connecting them is useful, but the connections are not the context.

I still have to build the complete context layer around them. The system needs to understand how all of those pieces relate to me, how I work, and what matters in a particular moment. That is the missing part.

There are really two missing links in my mind.

The first is a completely integrated context layer. Memory is part of that, and everybody in the industry is working on memory. But memory and deeply connected context are slightly different things. Remembering something about me is not the same as understanding how the information, tools, and activity across my life fit together.

The second is workflow-specific automation that comes out of the box. Today I have to build automations around how I handle email, how I triage items in Jira, or how I work with Quicken on my local machine. I have to explicitly connect the tool, connect it to the context layer, and build the automation.

I picture a world where those workflows are auto-discovered. The digital twin notices that every week Carlo reviews Quicken, updates it, and tags expenses. It recognizes that this is a recurring workflow, adds an AI-driven automation through the harness I am already using, and just does it for me every week.

There have been attempts at this. OpenClaw is one example. But these systems still feel hacky to me. They are not approachable for the roughly one billion ChatGPT users who still have to act like the tall person getting something off the high shelf for AI.

The tools are powerful. The integrations are growing. But the individual is still doing too much work to assemble the context that makes them useful.