– Regina Berengolts, Director of Data & AI, Vaul Labs · 4 min read
Earlier this summer I was visiting my mother.
We got talking about AI, and we were both genuinely excited. As a bit of background to understand any generational divide, she is in her sixties, although in case she is reading this, she looks more like a spry forty-two-year-old (hi, mom!). I was excited about the agentic processes I set up to run parts of our business: agents connected to our code bases that monitor, check and recommend resolutions with a human in the loop, project management agents that don’t let anything slip, and engagements that move faster because we have AI agents proactively working alongside us. She was excited that AI helps her write better emails and generate test cases for her QA work, which she then copy-pastes into another system.
It took us a surprisingly long time to notice we were not having the same conversation. We were both “using AI”. We were not using the same AI and we definitely weren’t thinking about it in the same way. And when it finally clicked, I watched her face change. Could she do what I was describing?
Chat is the norm, and chat is a fraction of what is there
Her version of AI is not the exception. It is the overwhelming default. OpenAI’s research with NBER, based on 1.5 million conversations, found that around three quarters of ChatGPT use is practical guidance, information seeking and writing. Pew found that about half of US adults now use AI chatbots, and the top uses are searching for information and routine work tasks. That is chat: asking, drafting, summarising. All of it is useful, and none of it is anywhere near the ceiling.
The difference shows up clearly when the same task moves from chat to an agentic setup. Anthropic’s Economic Index found that producing a blog post through chat takes a median of thirteen rounds of back-and-forth, while the median agentic session producing the same output contains a single human prompt. One approach assists you with the work. The other lets you delegate it, review it and move on.
The gap nobody can articulate
Inside organisations, this plays out as a quiet mutual confusion. Leaders believe they have done their part: they bought the tools, rolled them out, ticked the adoption box. Then they are surprised when productivity and growth do not follow. Meanwhile their people are using AI, in good faith every day, and cannot understand why they keep being told to use more of it.
Both sides are doing their best. They are just not speaking the same language, because neither side knows what is actually possible. “Use AI” can mean drafting an email or it can mean an agent that watches a codebase overnight. Until an organisation gets specific about which, adoption metrics will look healthy while the value stays locked up. Nobody raises the problem because nobody can see it. You do not know what you don’t know.
Sometimes the constraint is purposeful: permissions, policies and tool choices decide what an individual is even allowed to attempt, and two people with different setups will have completely different experiences of “AI”. But often the constraint is nothing more than awareness. I regularly meet highly capable people getting real value from AI who, it turns out, have only ever used a web-based chat window, which until *just *recently offered little in the way of agentic capability. They have never scheduled a task to take admin off their plate, let alone delegated a workflow. The possibility was never shown to them.
Using AI is not the same as being AI enabled. There are tiers to this, and a chat assistant, however well used, is the bottom one.
The ingenuity is already in the building
I can’t stop thinking back to that conversation with my mom. The moment she understood what agents could do, she was already articulating how she could map them onto her own test cases, with no training course and no mandate. All it took was exposure to what is possible, landing on thirty years of domain knowledge.
That ingenuity already exists in your organisation, in people who know the customer, the process and the awkward exceptions better than any tool ever will. What they are missing is the knowledge of what is possible and an operating model that lets them act on it safely. So the leadership question is not “are our people using AI?”. They are. The question is whether anyone has shown them what the next tier looks like, and whether your permissions, tooling and governance would let them reach it.
That gap does not close on its own. Working through something similar? Get in touch.