Something clarified in May 2026. Not a single announcement, but a pattern across several of them. OpenAI didn't launch a new model. It launched a deployment company, backed by $4 billion and partners including Goldman Sachs, McKinsey, TPG, and Bain. Anthropic didn't release a new Claude. It backed a $1.5 billion AI services firm with Blackstone and Goldman. Two of the most valuable AI labs in the world looked at the enterprise market and concluded: the bottleneck isn't the model. It's everything that happens after you sign the contract.
"The bottleneck isn't the model. It's everything that happens after you sign the contract."
That's worth sitting with. Because if you've been waiting for a better model before committing to deeper AI integration, the labs themselves are now telling you that's the wrong frame.
📢Upcoming Event Announcement: Exclusive Session | AI Deployment in PE Portcos: The 2026 Playbook. Joining forces with enable2grow on 2nd June 2026 at 15:00h CET to share a practical AI implementation roadmap for PE investors & operators. Save your spot https://luma.com/4qkjb9tk |
Where Most Companies Actually Sit
After running AI implementations across more than a dozen mandates, I keep seeing the same five levels of integration in practice. Not as a theoretical framework, but as a description of where organizations actually are.
Level 1: Tool Adoption. Teams use ChatGPT, Copilot, or Claude for isolated tasks. There are real productivity bumps at the individual level. But the knowledge stays in people's heads, and the gains don't compound across the organization.
Level 2: Knowledge Layer. Meetings are recorded by default (with safety guardrails). Decisions are documented. Processes are written in a way AI can actually use. This is the level almost everyone skips.
Level 3: Process Redesign. Specific workflows are rebuilt with AI as a layer: marketing nurture, M&A target research, content production, deal screening.
Level 4: System of Record. AI is where work happens. Data flows in, intelligence flows out, agents act on it.
Level 5: Org Redesign. Headcount, roles, and decision rights are restructured around AI capacity. Very few organizations are here.
Based on what I'm seeing across clients and the data coming out of mid-market research, approximately 60% of companies are at Level 1. Around 10% have built a real knowledge layer. About 20% are doing some process redesign, though often on top of a weak foundation. Roughly 8% are operating at Level 4, and about 2% have reached Level 5.
Most companies are trying to jump from Level 1 to Level 3 in one move. It rarely works. And the reason is almost always the same.
The Level Everyone Skips
Level 2 looks like cost, not value. Recording meetings, documenting decisions, writing processes in an AI-readable way. None of it shows up on a dashboard. None of it impresses a board. So it gets deprioritized in favor of something that looks more like progress.
Then Level 3 stalls. The agent has no context. The workflow can be automated, but the inputs are scattered across Slack, inboxes, and people's heads. Rowan Trollope, CEO of Redis, put it plainly earlier this year:
"Data and knowledge layer is often unstructured. It's sitting in Slack threads, in email chains, in text messages."
— Rowan Trollope, CEO of Redis (The Register, January 2026)
This is a foundation problem.
Meta understood this. Before deploying any production AI workflows in April 2026, they spent real time building what they called a knowledge layer: mapping 4,100+ files of tribal knowledge across three repositories using a swarm of 50+ specialized AI agents. AI context coverage went from 5% to 100%. Complex workflows that previously took two days now complete in 30 minutes. The knowledge layer wasn't the glamorous part of the project. It was the part that made everything else work.
What the Data Is Saying
The numbers from May 2026 are consistent across sources, and they're uncomfortable.
Kaufman Rossin's State of AI in the Mid-Market report, published this month, found that 94% of mid-market companies are using generative AI. Only 2% have operationalized it at scale. That gap between "we use AI" and "AI works across our organization" is where most companies are living right now.
Microsoft has sold 20 million paid M365 Copilot seats in two years. Accenture alone accounts for 743,000 of them. Actual workplace utilization sits at 35.8%. Fewer than four in ten employees with access use it. Licensing and integration are two different decisions, and most organizations have only made the first one.
The research on agents is blunter still. A Microsoft study published May 11 tested frontier models, including Claude, Gemini, and GPT, on long-running delegated tasks. The finding: these models lose an average of 25% of document content over 20 delegated interactions. Catastrophic corruption occurred in more than 80% of domain and model combinations tested. The only area where AI met a reliability threshold was Python coding. Every other professional domain fell short. Agents aren't failing because companies chose the wrong vendor. They're failing because agents need a structured, documented, machine-readable environment to operate in, and most organizations haven't built one.
What This Actually Means for Senior Leaders
The question I keep coming back to with clients is this: are you treating AI as a tool purchase or as infrastructure?
Tool purchases have a procurement logic: evaluate options, pick the best one, deploy it, measure adoption. Infrastructure has a different logic: what needs to be true about our data, our processes, and our institutional knowledge before this can work reliably at scale?
I've worked with two portfolio companies that had the same revenue, the same tech stack, and very different outcomes. One was at Level 1, the other at Level 4.
"The difference wasn't budget. It wasn't the sophistication of their IT team. It came down to who was in the room when the foundational calls were made."
Building the knowledge layer isn't exciting. It doesn't generate a case study in the first quarter. But it's the reason the Level 4 company's agents have context to work with, and the Level 2 company's agents keep producing outputs that nobody trusts.
The fact that OpenAI and Anthropic both launched deployment vehicles this month, both backed by the same PE firms that advise portfolio companies on operations, is not a coincidence. They're looking at the same gap you're looking at. And they've concluded that the value is in solving the deployment problem, not in selling another license.
The Practical Question
Buying is the easy part. Building the organizational infrastructure that makes AI actually work at scale, that's the hard part. And it starts with Level 2: making your company's knowledge, decisions, and processes readable by a machine.
This is exactly the work we do in an AI Diagnostic Blueprint engagement: diagnosing where a company actually sits across these five levels, identifying what's blocking the move up, and building a sequenced plan that doesn't skip the foundations. If that's a conversation worth having, reply to this email and we can set something up.
And if you want to see what this looks like in practice, join me next week. I'll be walking through real case studies at each level — what got companies there, what's kept them stuck, and what would need to be true to move up.
One question I'd love you to sit with before then: which level is your organization actually operating at, and what would it cost you to stay there for another 12 months?
Talk soon,
Pooja
PS: If this was useful, forward it to a portfolio company leader or operating partner who is still trying to figure out where to start. That is exactly who I wrote it for.