LISTEN

Listen to this perspective

Written by Oleg Cohen. Companion narration uses the Brian voice.

Part I, The Missing Owner, described companies that handed AI to the wrong owner. This essay is about the companies where it never had one.

The sequence goes like this. Nobody decides anything. A salesperson starts drafting proposals with a personal AI account. A branch manager builds a scheduling helper. One department automates a monthly report. The work gets done faster. Months later, the CEO discovers that AI is already spread through the company, and nobody can say where it is, what data it touches, or what it has changed.

I think this is the more common of the two sequences. It comes in three forms, and most companies have all three at once.

AI that nobody decided on

At the ACG event I described in the first essay, several CEOs said their employees had moved faster than company policy. The surveys agree. In UpGuard’s research, 81 percent of employees report using AI tools their employer hasn’t approved. A Gartner survey found that 69 percent of organizations suspect or know that staff are using prohibited tools. Most of these surveys come from security vendors, so read them with some caution, but they all find the same thing.

The risks are real. Customer data goes into consumer tools. Customer-facing work varies with whoever wrote the prompt. An improvement stays in one branch, one department or one person’s habits, and it leaves when that person does. A buyer’s diligence team will eventually ask what AI is in use, on whose accounts and with what data.

The instinct is to treat this as a control problem and tighten IT policy. Controls are necessary, and companies with weak ones carry the most risk. But control and direction are different things. In the same UpGuard research, 45 percent of workers found ways around blocked tools. A TrustedTech survey found that senior decision-makers use unapproved AI at about twice the rate of the people they manage. Blocking mostly costs the company its visibility.

There is also something valuable in those experiments. The people closest to the work have been showing, for free, where intelligence helps. A company with strict controls and no owner gets safety and little else. A company with no controls and no owner gets energy and risk. Neither gets capability.

The pull of the tools

The second form shows up in almost every company I talk to. The conversation is about RAG, models, vector stores, embeddings, agents and agent swarms. New tools arrive daily, and technologists are fascinated by them. I am too, and I spend plenty of my own time on them.

This week the conversation is about Jev. TypeSafe AI released it on September 15, with $40 million in seed funding and a founder who helped invent the training method behind ChatGPT. It isn’t a language model. It takes in the state of an application and returns typed decisions with probabilities, instead of text, and the company says it is two orders of magnitude faster and cheaper than an LLM for that kind of work. Demand was high enough that the API briefly went down. I’ve spent time on it, and it may turn out to matter. It’s still a good example of the pattern. Nothing about Jev tells a company which decisions are worth making better, and by next month the conversation will have moved to the next release. A second stream of attention goes to chatbots and assistants that make individuals faster.

The same fascination produces a version of not-invented-here. With modern coding tools, a capable team can build a convincing prototype in days, so people assume they can cobble the whole solution together themselves. I build and sell technology, so I have an obvious interest in this point. The argument doesn’t depend on who supplies the technology, though. A prototype isn’t a capability. Making it reliable, governed, connected to real data and used across the company is most of the work. Building can be the right call for something that differentiates the company. It’s the wrong default when nobody has asked whether this is the thing worth owning.

Tools and assistants are both worth the time. The trouble is that they become the main focus, and there is a reason they do. Tool activity is easy to see and count, in licenses, adoption and hours saved. Whether any of it reaches the P&L is much harder to show, as the survey data in the first essay suggested. The capability question is harder to specify and slower to answer. It is also where enterprise value and durable differentiation come from.

Use cases are too small

The third form looks the most disciplined. The company produces a long list of use cases. That sounds business-led, but it can keep the technology-first view intact. Consider the levels involved in one example.

  • Enterprise outcome: higher retention, lower cost to serve, stronger customer lifetime economics.
  • Capability: customer recovery and retention.
  • Workflow: service case management.
  • Use case: help an agent resolve a customer issue.
  • Feature: summarize a service call.

Most programs start at the bottom of that list and hope value appears at the top. “Use AI to summarize service calls” can be bought, deployed and reported as a success, as an earlier essay in this series argued. The more useful question starts at the top. What would a much stronger customer recovery capability look like if every technician, manager and system could draw on what the company already knows? Summarization might turn out to be one small part of the answer.

What to do with the activity

All three forms are activity, and none of it is wasted. The experiments show where intelligence helps. The tooling knowledge will be needed. Some of the use cases belong inside a larger capability. What’s missing is someone deciding which of it points toward something the company should become good at, and then building that properly.

That is the owner Part I described, and in most companies the owner’s first task is an inventory. Find out what is already in use, and give people safe tools for it. Keep the experiments that point toward a capability, and spread what works beyond the branch where it started. Then start from the top of the ladder and ask what the company should become capable of.

The same applies in diligence. Ask what AI is already in use at the target, who adopted it, and what data it touches. The answer shows the exposure. It also shows where the organization’s energy is.

Where I sit: I advise through Kainora and work as a fractional Chief AI Officer, so I have an interest in this gap being recognized. I’m also building technology shaped by these views.

Two questions for you. If you listed every AI tool and experiment running in your company today, how many point toward something the company should become good at? And who would decide?

Sources and further reading

Share your perspectiveConnect with Oleg on LinkedIn (opens in a new tab)