#Which AI are we talking about?

For three years now, executive teams have had one order: do AI. But “AI” is a suitcase word. It packs several nested ideas: machine learning, deep learning, generative AI, and large language models (LLMs). People often mix them up, but when someone says “AI” today, they usually mean LLMs: ChatGPT, Gemini, Claude and their relatives.

Even if we stick to LLMs, there’s still an ambiguity. An organization can use them in two different ways:

• as intelligence built into existing tools and processes, which then draw directly on what the models can do;
• as a way to build faster, making tools quicker and easier to produce, whether or not those tools contain any AI.

One question separates the two: once the tool ships, is the AI still running? If it is, the AI lives in the product. That’s the first lever. If it isn’t, the AI was in the hands of whoever built it. That’s the second.

##Lever 1: AI inside the work

The first lever is the most visible and pays off fastest. You take an existing process and add generative AI to it, especially for routine or rule-bound tasks, the kind you’d “give to a junior.” This is what LLMs do best. They cut the time a task takes, improve the output, or both.

Example one: writing sales proposals. It looks like deep work, but it’s mostly a chain of document chores: summarize the brief, dig up similar proposals, check past deals, justify the recommendation, then grind through the formatting. Choosing the offer and the price, which is the real sales skill, takes less than 10% of the time. The rest is routine. Three years ago, automating it was science fiction.

A well-built AI agent cuts writing time by a factor of five. It reads the brief, picks the right offers from the catalog, writes the rationale in the client’s own language, and exports a PowerPoint deck. The salesperson tweaks it and sends it. On top of the time saved, proposals come out consistent and the errors go away.

Example two: data analysis on demand, no analyst required. In most companies, any numbers question that isn’t trivial kicks off the same routine. You file a ticket with the data team, wait in the queue, go back and forth to pin down what you meant, and get an answer ten days later. Often it misses the point, because the business context got lost along the way.

With AI properly wired in, a marketing or CRM team opens an agent connected to the database and asks its question in plain English. Three minutes later it has an analysis, which it can challenge, rephrase and refine. Data analysts don’t disappear, but they stop fielding trivial everyday questions.

Example three: pulling scattered sources together. A strategy team wants a consolidated picture of a competitor: annual reports, press releases, executives’ LinkedIn posts, job listings, trade press. It used to take a junior two weeks to compile a fifteen-page memo of uneven quality. Now an agent ingests the sources continuously, writes a weekly summary, and flags when something shifts.

##Lever 2: AI in the builder’s hands

The second lever is less direct, but the change runs just as deep. Building even a modest tool used to mean knowing how to code. Think of a script that reconciles two files automatically and saves everyone on a team three hours a week.

AI coding agents remove that requirement. First, they make development much faster. A developer with Claude Code can ship in a few hours what used to take several days.

More importantly, they put that ability in the hands of people who don’t code. The hard part is less knowing how to code than saying clearly what you want. And nobody is closer to the business need than the business teams.

This clears up a common mix-up: tools built with AI don’t necessarily contain any AI. Most are plain automation tools that could have been built three years ago, but only by tying up serious technical resources. A form that pushes sign-ups straight into the CRM. A script that matches supplier invoices to purchase orders every month. A dashboard that merges five Excel exports into a weekly report. None of this is new. What changed is the cost of building it.

##The real issue isn’t technical

Both cases are fairly called AI. But they shake up an organization in very different ways.

The first, the one everyone thinks of, puts intelligence into the product and adds capabilities without disrupting roles much. The tech teams still control what gets built.

The second works below the surface, and it will give managers headaches. Once LLMs absorb the technical know-how, the line blurs between the people who state the need and the people who build the fix. A business team can now build in one morning the tool it has been requesting since Q1.

That makes governance urgent in a new way. Who gets to build tools, and with what guardrails? Everyone who tried Claude Code on a rainy weekend showed up Monday morning as a convert. Clamping down in the name of security is an understandable reflex. A badly designed tool is a data-leak risk, and it leaves technical debt for the tech teams to carry. But lock things down too hard and people build outside any framework (often on their personal accounts), or the company stays frozen in 2023. “Be agile” was the tech teams’ mantra for fifteen years. Now that anyone can build software, that debate belongs to the whole organization.