By Michael Domanic, Section Head of AI

The workforce is finally gaining traction with everyday AI … and they’re also getting further behind.

I’ve been paying attention two stats recently:

1. Workers are getting better at general AI. In our latest AI Proficiency Report, weekly AI use jumped from 55% to 67% in six months, and the share of workers who can write an effective prompt tripled (from 7% to 21%).
   
   
2. A small group of companies are pulling ahead on agents. OpenAI's enterprise data shows companies in the top 10% of AI usage now generate 8.3x as many output tokens per active user as the typical firm, up from 2.6x in January. 

OpenAI calls that top 10% "frontier firms," and their AI sophistication shows up everywhere. 19% of their weekly active users work with reusable skills, versus 3% at the typical firm. And the growth is coming from outside engineering: since February, weekly agentic usage grew 108x in legal, 41x in sales and recruiting, and 26x in marketing, against 5x in engineering.

Tokens are a crude proxy for value (OpenAI acknowledges this), but as a measure of how much work is being delegated to AI rather than assisted by it, the direction is clear.

If you're leading AI at your organization, the general proficiency gains your workforce is making are necessary but not sufficient. The companies pulling ahead are doing so because they’ve started redesigning how work gets done around agents, when most organizations haven't even started that conversation.

The next big skill: Redesigning workflows

McKinsey's recent State of AI survey found that 80% of workers say AI has improved their individual productivity. But the share of companies attributing any measurable financial impact to AI sits at 37%, exactly where it was a year ago. 

I think there is a lot going on here (something to explore in another newsletter post) but a big part of this is that AI is just being dropped into old workflows.

When you drop AI into an old workflow, people finish their tasks faster, but everything around them still generally runs at the original speed: meetings, approvals, handoffs, etc. The time savings leak away into a system that was designed for slower humans, which is why they often don’t show up in the bigger efficiency calculations.

The companies breaking through do something different. Among McKinsey's AI high performers, nearly three-quarters have fundamentally redesigned workflows with AI agents.

This is what your Head of AI and your AI champions actually need to know how to do. Companies that don't feel equipped to do this, or aren't willing, will keep falling behind the pace of the broader transformation and the people who work at those companies fall behind with them. 

How to redesign a workflow around agents

At Section, everyone is expected to operate at the “AI expert” level, which means engaging with agents and, in most cases, building them. I spend part of nearly every week with functional leaders helping them redesign their teams' workflows around the agents we deploy. Here's our approach:

1. Pick a workflow, not a task. A workflow has a start, an end, handoffs, and a business metric. “Summarize this call” is a task – “discovery through proposal” is a workflow. Map the current state as a starting point. 

2. Zero-base it. Ask: If we were designing this workflow today, with agents available, what would it look like? Many steps in legacy workflows exist only because humans are slow, forgetful, or unavailable: the status meeting, the handoff email, the batch review. Don't automate those steps. Delete them.

3. Sort what remains into three buckets. Delegate to agents: high-frequency, describable work with verifiable outputs. Keep with humans: judgment, relationships, accountability. Hybrid: agent drafts, human approves.

4. Build the cluster. Redesigned workflows run on agents that feed each other. Our sales discovery workflow runs on a cluster: one agent scores every call against a rubric, another reports to the AE on what they captured and missed, a third checks whether gaps got filled in later conversations, and yet another provides reporting on deal movement to our executive team. 

5. Measure the workflow, not the agent. Track outcomes and total inference cost at the workflow level, against the baseline you established in step 1. 

Then hand your champions the enablement to match, in this order: literacy (what agents are and when to trust them), operation (working with agents others built), judgment (spotting which workflows deserve one), creation (building). And train managers first. Workers whose managers expect AI use score 1.5x higher on proficiency, yet fewer than 8% of managers tie AI use to performance. Redesign dies at the manager layer faster than anywhere else. We talk about this a lot at Section but it bears repeating.

Start with one workflow per department: high frequency, low blast radius, 30 days to ship. You're not just chasing that workflow's ROI. You're building the muscle your organization will run on, because the frontier tripled its lead in six months, and it is not waiting for anyone's planning cycle.

See you next week,

Michael
Your fellow Head of AI

‍