What I saw at Ai4: Enterprise AI is earlier than we think

One thing stood out to me at Ai4 in Las Vegas, a conference of roughly 12,000 attendees, many of them from large enterprises: those of us who work closely with AI are living in a bubble.

I do not mean a valuation bubble, but an awareness bubble. When everyone around you follows the latest models, agent frameworks and “company brain” architectures, and joins a new WhatsApp group for every emerging tool and harness, it is easy to forget how unfamiliar much of this still is elsewhere.

Inside the Tel Aviv tech scene, pushing the boundaries of AI feels like an obvious goal. The same is true in San Francisco, London, Bangalore and other technology hubs. If you are reading a newsletter like this one, you are probably inside that bubble too.

It’s easy to forget that we are early adopters, and that crossing the chasm is still far away in the AI world for most companies. This is of course also a huge opportunity, and I’ll come back to what I think that means at the end.

[A note: this was a US-weighted conference, so the sample skews American, though I’d expect a similar picture in most European and APAC enterprises]

Let’s start with a few use cases presented during the conference, for inspiration:

1. Patient-facing care agents at a clinic group. An owner of several medical clinics built agentic flows (automations with a call to an LLM) for customer care. These agents help the patients understand what’s going to happen at the visit, what the predicted costs will be, and handle back-office tasks like uploading files and receiving prescriptions. Most patient questions are handled before the visit instead of during it.

2. COCO, at Pfizer. A knowledge graph-powered insights platform built internally to solve the “speed to insights” problem in commercial analytics. A process that took around 30 days now returns an immediate answer.

3. PHI redaction and document classification. Dental claims arrive as documents needing protected health information stripped and the document typed before downstream processing. They process-mapped the workflow first before applying anything. ~8M claims/yr handled; 11 hires avoided.

4. Transcription at a localization company. They used to deliver human transcriptions of recorded audio and video, priced per minute. → 1-hour video: 4 hrs, done with AI: 1 hr 15 min of human work. The old price was $10/min, so $600 per hour of video. That per-unit price became indefensible, and it forced them to move to credit-based and then outcome-based pricing. The automation didn’t just cut their cost, it broke their price list.

5. Internal data agents (sometimes known as RAG agents). Several companies built these to let employees ask and extract information across HR, finance, legal, sourcing, ethics and research, and sometimes even take actions on behalf of that employee. The most common project in the room, and from my POV it’s a great use case to start with.

6. Government RFI/RFP portal scanning. An analyst was manually checking 140 government procurement portals every day for new solicitations. The agent scans them and delivers a digest. Went from: 3–4 hrs/day → a morning inbox check.

7. Medical record intake and triage. Records arrive with inconsistent clinical terminology. OCR extracts the text, a normalization step reconciles the vocabulary (their example: “unilateral” vs. “bilateral”), agents apply rules that live in institutional knowledge rather than documentation, and a recommendation engine aggregates the outputs.Went from → 4 hrs → minutes. No cost disclosed. The hard part they described wasn’t the model, it was that the rules were undocumented.

Let’s use this last example as a segue into the next topic, what actually matters in the AI journey?

1. Data and documentation. As we see in the medical record intake example, undocumented knowledge is a huge issue in AI projects. I would even venture to say we are living in the era where documentation is becoming more important than the process itself.

Here’s why I think that. An undocumented process can still run, because a human operator fills the gaps without even noticing they are gaps. An agent can’t do that. The context layer is the most important one for agents and humans working together, and anything that isn’t in it is invisible to half your workforce. So the undocumented process isn’t just fragile any more, it’s un-automatable.

2. Most of the value is in workflow design and people, not in the model. The assumption in the room was that around 70% of the AI value organizations capture is about the process and the people, and only 30% is about the model. Process thinking is key, decompose the workflow into clear decisions, inputs, outputs and edge cases before building an agent to automate it. Every case above with a real number did this first. If you don’t you’ll get to a frustrating point where you have 80% of the work done, and probably you will not finish it ever (stuck in the demo area).

3. Treat any AI output as a draft. You’ll usually get to 90% of what you need, and a human still needs to review the output. Worth designing that review step in from the start, as a job someone owns, not a disclaimer you add later.

So what does the gap actually mean?

If most large companies are still at “AI and vibe coding” while we’re comparing agent harnesses in group chats, a few things follow, and I think they hold whether you’re selling into New York, Frankfurt or Singapore.

The bar is much lower than we assume. Case 6 is a scraper and a digest email, and it got stage time at a major conference. We’d dismiss that as not even worth demoing. And you know what? The important thing is not the tech, but moving the business needle.

What’s scarce isn’t the technology, it’s the process work. Every result above with a real number was gated on someone mapping a workflow or writing down rules that only existed in people’s heads.

The window has a length. While categories are still forming, vendors and internal teams can help define what “normal” looks like. Later, requirements harden and procurement takes over.

To conclude, AI is a tool, not the objective. Within two or three years, the strongest enterprise implementations will be discussed less in terms of models and agents and more in terms of cycle time, conversion, cost, quality and risk, just as cloud and mobile eventually became part of the operating environment rather than the headline.

What I saw at Ai4: Enterprise AI is earlier than we think was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.