A fully AI-run business does not run faster versions of its old processes. It runs work that has been rebuilt around AI agents acting as teammates, on top of a clean, governed knowledge layer, with each agent's autonomy expanded one trust step at a time. That is the whole operating model in a sentence. The companies that have done this are not using better models than everyone else. They did the operational rebuild that almost nobody does, which is why 88% of organizations now use AI somewhere but only about 6% are genuine high performers.
This piece maps McKinsey's seven operating truths of AI-native companies onto a concrete sequence you can actually follow: get the knowledge layer right, redesign the workflow, run agents as teammates, and earn trust incrementally. We will use real numbers and one real example (Salesforce's support division) to show what AI-run looks like inside a live operation, and answer the question every source dodges: who does the rebuild for a company without an AI engineering team. If you would rather we run that sequence for you, you can see how we deliver AI strategy and executive advisory, but everything here is yours to use on your own.
What does it actually mean for a business to be AI-run?
It means agents do the real work inside the operation, not that a model sits beside it. The structural test is the same one used to define an AI-native company: remove the AI and the operation stops, rather than just getting slower.
The honest part most essays skip is that "AI-run" is not all-or-nothing for an existing business. A company can be fully AI-run in support while its finance close is still a spreadsheet relay race. You do not convert the whole company at once. You convert it one function at a time, and each converted function is genuinely AI-run even while the rest catches up.
The clearest live example is Salesforce. After deploying its Agentforce agents, Salesforce now routes about 50% of customer support interactions through AI, reports roughly 17% lower support cost, and cut its support headcount from around 9,000 to around 5,000 as the agents took over the work. That is not a chatbot pinned to a help page. That is one function rebuilt so the agents carry half the load, which is what AI-run looks like at full scale inside a single operation.
How do AI-run companies actually operate? The seven truths
McKinsey distilled how AI-native companies run from leaders at roughly fifteen AI-savvy companies and landed on seven operating truths. They read as an operating system most organizations still get wrong. Here they are, with what each one means in practice:
| Operating truth | What it means in practice |
|---|---|
| Agents are teammates, not tools | Agents own steps end to end and amplify people, instead of being a feature someone opens occasionally |
| Be deliberate about build vs buy | Decide per capability, not as a religion, so you are not rebuilding what you can buy |
| Get the knowledge layer right | Clean, queryable, governed context is the foundation everything else stands on |
| Build a composable, governed architecture | Design to swap parts as the field moves, with governance baked in from the start |
| Earn trust incrementally | Autonomy expands in steps, from human-approved to supervised to automatic |
| Get the organizational design right | Structure teams and ownership so agents can scale past one pilot |
| Make adoption a cultural flywheel | People pull the new way of working in; it is not an IT tool pushed down from the top |
The single thread running through all seven is the one McKinsey leads with: winners do not bolt models onto old processes. They rebuild the processes and re-platform their knowledge so AI can act reliably. The real value is not doing the same work faster. It is amplifying people with agents that function as genuine teammates.
These seven are not a checklist you do in any order. Four of them form a sequence, and that sequence is how you actually become AI-run.
What is the right sequence to become AI-run?
The truths cluster into four moves, and the order matters. Skip a step and the next one fails.
1. Get the knowledge layer right first
Agents are only as good as the data and rules they can reach. Before you add any agent, the knowledge it depends on (your docs, policies, product facts, customer history) has to be clean, queryable, and governed. This is the least glamorous step and the one most companies skip, which is why their pilots stall. If support agents are pulling answers from a stale wiki, no model is good enough to save the result. The knowledge layer is the foundation, and it comes first because nothing downstream works without it.
2. Redesign the workflow, do not bolt an agent onto the old one
This is the step that separates the 6% from everyone else. McKinsey's data is blunt about it: only about 21% of organizations using generative AI have redesigned even one workflow, while nearly 80% just layer AI on top of existing processes. Redesign means you ask what an agent can do well, then rebuild the workflow around that, instead of inserting an agent into a process designed for humans. A redesigned support workflow does not hand a ticket to an agent and wait. It lets the agent read, decide, and resolve, and routes only the genuine edge cases to a person.
3. Run agents as teammates, not features
A teammate owns an outcome. A feature waits to be invoked. In an AI-run operation, the agent handles a category of work end to end (resolve the refund, qualify the lead, reconcile the invoice) and a person handles exceptions and approvals. This is where the structural shift a16z describes shows up. The enterprise backend was built for a one-to-one ratio of human action to system response, but agent workloads are recursive, bursty, and massive, which pushes companies to re-architect for "agent-speed" rather than human-speed. Their sharper prediction: the system of record (the CRM, the ticketing tool) loses primacy and becomes a commodity persistence layer, while the intelligent execution environment that does the work becomes the strategically dominant part of the stack.
4. Earn trust incrementally
You do not flip an agent to full autonomy on day one. Trust expands in steps. The agent starts on a tight leash with a human approving every run, then moves to supervised (it acts, a human spot-checks), then to automatic on the safe, well-understood cases while humans keep the edge cases. This is how Salesforce got to half of support interactions on agents without breaking the support experience: the autonomy widened as the agents proved themselves, not before.
Why are so few companies actually AI-run?
Because the hard part is not the AI. It is everything around it: the knowledge cleanup, the redesign, the governance, the org change. McKinsey's State of AI data lays the gap out as a funnel:
- 88% of organizations now use AI in at least one function, up from 78% the prior year.
- Only about one-third report scaling AI across the enterprise.
- Only about 21% of organizations using generative AI have redesigned even one workflow.
- Fewer than 10% are scaling AI agents in any function, and 73% use no agents at all in product development.
- Only about 6% are genuine AI high performers. Those few are 3.6x more likely to pursue transformational change, and roughly 55% of them fundamentally rework their workflows.
Read top to bottom, that is the distance between "using AI" and "being AI-run." Almost everyone is using AI. Almost no one has rebuilt how work happens. The drop-off is the cost of the four-step sequence above, and it has nothing to do with which model you picked.
It is worth crossing because the payoff is already real, not hypothetical. PwC's agent survey found 79% of companies report AI agents are already being adopted, two-thirds of those adopters report measurable productivity value, and 88% of executives plan to increase AI budgets in the next twelve months. The Salesforce numbers are the proof at scale. The companies that finish the rebuild are pulling away from the ones still bolting AI onto the old process.
Can an established company become fully AI-run, or is this only for new startups?
It can, and the distinction matters. Strictly, no established company can retroactively be "AI-native from day one." That ship sailed when it was founded. But AI-run is not a birthright you inherit. It is a capability you install, one function at a time.
An accounting firm, an agency, a clinic, or an e-commerce brand cannot un-found itself, but it can become genuinely AI-run in the parts that matter (support, operations, the sales motion, the back office) by running the four-step sequence on one workflow, then reusing the pattern on the next. The path looks like this:
- Pick one high-volume, language-heavy workflow that tolerates a human checkpoint.
- Fix the knowledge layer that workflow depends on before adding any agent.
- Redesign the workflow around what the agent can do, instead of bolting an agent onto the old one.
- Wrap it in governance, expand autonomy as trust grows, and measure the result.
- Prove it on one function, then reuse the pattern on the next.
Stack enough converted functions and you have the operating model the 6% have, even though you were not born with it.
Who does the rebuild when you do not have an AI engineering team?
This is the question every source dodges. They all agree on the diagnosis (the bottleneck is the knowledge layer, the redesign, the governance, the org design) and they all stop at "you must rebuild your processes." None of them say who does that rebuild for a company without AI engineers.
That is the gap we close. Sista AI is AI-run itself (we operate on agents), and we plan, build, and run AI agents inside other companies. Our work maps one-to-one onto the four-step sequence: we stand up the governed knowledge layer, redesign the workflow instead of bolting AI on, run the agents as teammates with humans owning the exceptions, and expand autonomy on a trust ladder while owning adoption as a managed outcome. The build-versus-buy truth becomes moot because we do the building. In practice, we make you one of the 6%, in the functions that matter, without you hiring an AI team. If you want a team that runs these agents day to day, that is what our AI business automation agents work delivers.
The short version: being AI-run is not a label you are born with or locked out of. It is an operating model with a known sequence (knowledge layer, redesigned workflow, agents as teammates, incremental trust) that you can install one function at a time. If you would rather not run that installation alone, book a free consultation below and we will map the first function to convert together.