AI alignment has a reputation as a frontier-lab problem — something argued about in San Francisco, concerning systems most businesses will never touch. Some of it is. I spend a good part of my research life on questions that will not matter unless a machine one day crosses a threshold no machine has crossed.
But the larger part of alignment is not about future minds at all. It is about present behaviour: what the people deploying AI owe the people affected by it, starting now, unconditionally. My framework calls this the Law of Shared Flourishing, and it is deliberately not contingent on anything — no threshold, no future event. If you deploy AI in a business today, the obligation to make sure it benefits the people it touches applies today.
At the scale of a twelve-person business, that obligation gets concrete fast. Here is what I think it actually requires.
The system must answer to someone who can be named
In a large company, an AI system’s mistakes diffuse into process. In a small one, they land on a person — the customer who got the wrong quote, the employee whose schedule the software mangled. Alignment at small scale starts with a simple architectural rule: for every action the system can take, there is a named human who approves it before it happens, and everyone knows who.
This is not caution for its own sake. A machine that reads intake forms and sorts them by urgency is doing real, useful machine-learning work. A machine that decides which customer matters is doing your job, badly, without accountability. The line between preparing a decision and making one is the most important line in any deployment, and small businesses are actually better positioned to hold it than large ones — the approver can be a real person with a real name, not a committee.
The benefit has to land on the people doing the work
The unconditional half of Shared Flourishing says AI development must benefit all legitimate stakeholders. In a twelve-person business, the stakeholders are not abstract: they are the owner, the staff, and the customers, and you can list them by name.
The honest test of a deployment is who gets the benefit. A system that lets the office manager stop retyping the same information into three places has distributed its benefit well — the business gains throughput and a person is relieved of drudgery they never wanted. A system installed to make a person redundant is a different transaction, and it should be called what it is rather than dressed as efficiency. This is not an argument against automation — it is an argument for honesty about the transaction, because a tool positioned as a colleague’s replacement poisons every other use of AI in that building. Staff who fear the system will not use it honestly, and a system fed dishonest inputs is worthless.
No one in the building should have to guess what the AI is
The failure mode my framework calls premature attribution — treating a fluent system as if it understands — is not a philosophical problem. It shows up in small businesses as misplaced delegation: the chatbot that was allowed to promise a delivery date, the drafting tool whose output went out unread because it “sounded right.” Fluency is not comprehension. The discipline is to be exact, with everyone who touches the system, about what it actually does: it recognises patterns in text, it does not know your business, and its output is a draft until a person makes it a decision.
The reverse failure — treating AI as pure hazard and refusing it entirely — costs quietly instead of loudly: the evenings spent on paperwork a machine reads accurately in seconds, the follow-ups that never happen. A small business that refuses all automation is not safer; it is just slower, and its people are more tired.
What this looks like when it is working
A working deployment in a small business is unremarkable to look at. The software reads what arrives, sorts it, drafts responses, keeps the files straight. A person reviews a queue instead of an inbox and approves what leaves the building. There is a number — quotes answered inside a day, documents found on the first search — that the owner can check without asking anyone. Nobody in the building is confused about what the machine is, nobody’s job got quietly hollowed out, and nobody has to trust a vendor’s word for whether it is working.
That is alignment, at the only scale most businesses will ever need it. The thresholds and protocols in the research — the parts about verified awareness — are there in case a much stranger future arrives. The obligations above are not waiting for it.