What "AI-first company" means, and the two uses of it

AI-first company

AI-first company: An AI-first company is one where AI systems own the default path of its core work end to end, and people handle exceptions, judgement and oversight rather than the routine volume.

The load-bearing words are default path. Not "uses AI." Not "has an AI strategy." The question is what happens to a unit of work — a support ticket, an inbound lead, a reconciliation, a first-draft spec — when nobody intervenes. If it sits until a person picks it up, the company is not AI-first, whatever the deck says.

Where the term came from

Sundar Pichai wrote the shift into Google's 2016 founders' letter: the move "from mobile first to an AI first world." He meant products. Search, Photos, Assistant — surfaces where a model, not a menu, does the work.

Between 2023 and 2025 the phrase migrated from product to payroll. CEOs started publishing internal memos using "AI-first" to describe how their own staff work, which is a different claim about a different thing. Nobody owns the term. No standards body, no certification, no definition you can point a vendor at. That is why the two senses get mixed, and why most arguments about whether a company "is AI-first" are really arguments about which sense is being used.

The two meanings, labelled

	Sense 1 — AI-first product	Sense 2 — AI-first operating model
Claim	The product's core value is a model's output	The company's internal work runs on models by default
Evidence	The feature stops working if you remove the model	The process stops, not slows, if you remove the agents
Who says it	Launch posts, funding announcements	CEO memos, org redesigns
Fails when	A wrapper on a CRUD app ships with a chat box	Seats are bought, defaults never change

These are independent. A company can sell an AI product while running its own sales ops on seventeen humans and a spreadsheet. The reverse happens too: a payments company with no AI in the product can put most of its support volume through a model. So ask which sense. The answer is sense 1; the implication is sense 2; that gap is the pitch.

What it is not

AI-enabled and AI-powered are feature claims. A summarise button on a ticket queue is AI-enabled, and the queue still belongs to humans. Digital transformation is a paper process moved onto software: the decisions stay human at every step and the forms load faster. AI-native is a founding-date claim — built after the current model generation existed — which is useful shorthand for investors and says nothing about whether the company's own operations run on models.

Then there is the zero-human company, the limit case, and the thing I am publicly building: no employee on the default path at all, and the agents pay their own bills. AI-first is a direction. Zero-human is the end of the road. Most companies calling themselves AI-first are one or two processes in, which is a real claim and a much smaller one than the phrase implies. I wrote up the smallest version of that in how to start a business with AI agents in one file.

A worked example with real numbers

Klarna's customer service assistant, per the company's own press release dated 27 February 2024, after one month live globally:

2.3 million conversations handled
two-thirds of Klarna's customer service chats
work equivalent to 700 full-time agents
average resolution time down from 11 minutes to under 2 minutes
an estimated $40M USD profit improvement for 2024

That is sense 2, documented. A ticket's default route was the model, and the human queue became the escalation path.

Now note what the release does not state: cost per conversation, error rate, or how often a conversation was re-opened. Those three are the first things I would ask for, and the omission is standard practice in AI-first announcements. A high deflection rate with an unmeasured re-open rate is a queue that moved, not a queue that cleared.

How to tell whether something qualifies

Five tests. Someone inside the company can answer all of them in an afternoon.

Test	Qualifies	Does not
Default owner	Work routes to the agent first	Work routes to a person who may ask the agent
Removal test	Pull the agent and the process stops	Pull it and the process slows
Measurement	Pass rate on a held-out eval set, tracked over time	Vibes, demos, a single anecdote
Cost	Per-run cost known from published per-million-token prices	"The seats are in the budget"
Containment	Step cap and spend ceiling enforced outside the model	Model instructed not to loop

The last row is the one people skip. Unbounded loops are the most common production failure mode for agents, and the fix is structural, not a prompt: the OpenAI Agents SDK exposes a per-run turn limit that the runner enforces and raises on. That is the right shape, because the counter lives in code the model cannot talk its way past. Same logic for tool access. The Model Context Protocol, published November 2024, standardised how a model reaches tools and data; it does not decide which tools a given run may touch. That is your call, and it belongs in the orchestration layer, which I argued is a control layer, not a framework.

For the measurement row, NIST's AI Risk Management Framework 1.0, released 26 January 2023, gives a usable vocabulary — its MEASURE function is about whether you have the instrumentation to make a claim at all. Adopting the framework is optional. Answering its questions is not.

Where I draw the line

I ran Metadata.io from $0 to $15M ARR and raised $50M doing marketing automation, and I hold 6 patents in AI-driven marketing. Teams that put autonomous agents into production carry the inference and tool-call costs directly, which is what turns spend authority from a governance abstraction into a line item someone has to answer for. Spend authority is the last thing to delegate. I will not hand an agent a payment credential without a hard ceiling enforced by the payment system itself, because a ceiling written into the agent's instructions is a suggestion the model can reason around.

The honest version of the claim is narrow and checkable. "Our inbound lead triage runs without a human in the default path; a person sees 8% of leads, the exceptions." That is an AI-first process, stated so someone can audit it. Build one of those first. A company becomes AI-first one default path at a time, and if the answer to "what changed" is "we bought seats," that is procurement.