An AI agent sent an email on someone's behalf last night. It may have been the right email. It may have gone to the right person, with the right attachment, in the right tone. The person it acted for was asleep. This morning, can they tell you why it was sent? Can they prove it to a client, a regulator, or a contracting officer who asks?
For almost everyone using agents today, the answer is no. That is the problem I started SAIN Industries to solve. We are building the trust infrastructure for industries and government.
The problem: agents act, and nobody can verify
For two years the conversation about AI was about what models could say. That era is ending. Agents now send emails, sign off on drafts, move money, and train people. They have moved from chat to action.
Action changes the question. When a model gives you a bad answer, you ignore it. When an agent takes a bad action, something has already happened in the world, under your name. The email went out. The payment moved. The filing was submitted.
Every other actor we allow to do that leaves a trail. An employee has an inbox, a manager, and a personnel file. A bank transfer has a ledger entry. A lawyer has a file with their name on every page. An agent, in most products shipping today, has a chat history that the vendor controls and that anyone with the right access can change.
Picture a title lawyer with four inboxes. An agent could take half of that load tomorrow. The lawyer knows it. What stops them is the conversation that follows: a client, a carrier, or a bar examiner asks what the software did with the file, and the lawyer has nothing to hand over but the vendor's word.
So the people responsible for risk are asking a fair question. How do you trust a machine's actions when you cannot see inside it?
The industry keeps dodging that question, or answering it with "trust the lab." I think there is a better answer. You do not need to see inside the machine. You need a private, provable, tamper-evident record of what it did.
Each of those three words is doing work. Private, because the record of a law firm's or a lender's day is full of other people's secrets, and a record that exposes them will never be switched on. Provable, because a log that only the vendor can read is a story, and the person asking wants evidence they can check. Tamper-evident, because a record that can be edited quietly after the fact protects nobody. If it changed, everyone must be able to tell.
Three disciplines that almost never share a building
Building that record takes three disciplines.
The first is privacy. The field of privacy-enhancing technologies, including zero-knowledge proofs, is about showing that something happened, or that a rule was followed, without exposing the underlying data. The simplest version is a record that keeps a fingerprint of a document instead of the document.
The second is transparency. A shared ledger, built properly, is a tamper-evident record: an edit shows, so every party can check the same history. It turns "take our word for it" into "check it yourself."
The third is intelligence. AI agents that do real work, built and shipped to customers who pay for them.
Most AI companies have the third. Almost none have the first two. The reason is simple. The three fields grew up apart. Privacy engineering came out of cryptography research. Shared ledgers came out of finance and supply chains. Agents came out of machine learning labs. The people do not go to the same conferences, do not read the same papers, and rarely end up at the same company.
Each one alone falls short. Intelligence without a record is a black box holding your credentials. A record without privacy is a leak, and no law firm will put client matters on a ledger that others can read. Privacy without a shared record is a claim nobody else can check. Put the three together and you get something new: a machine whose actions can be checked by the people who need to check them, without exposing what it saw.
That is what I mean by trust infrastructure. Privacy, transparency, and intelligence in one stack.
Why now
Three things are happening at once.
Agents are crossing into action. Every month they are handed more of the work that used to require a person with a license, a signature, or a login.
The people who sign off on risk are catching up. Regulators, compliance teams, and contracting officers are asking for the audit trail. Today nobody has one to give them, and purchases stall on that question.
And the frontier labs are competing on capability. That is their job, and they are good at it. But a lab grading its own model's behavior is no audit. The accountability layer has to come from outside the labs. It has to be built by people who already know how to make records that show every edit and proofs that do not leak data.
There is also a clock. Government contractors are being asked, on a schedule, to show how they control and account for the tools that touch sensitive work. Legal and finance teams already carry record-keeping duties, and those duties did not go away when the work moved to an agent. The duty stayed. The record disappeared.
The barrier to AI adoption is trust. Consumers do not care about the audit log. Legal, finance, and government cannot buy without it. We build for the second group.
Why us
I have built this system three times.
The first time, it was sensors and cameras in the field, writing what they recorded to a chain so that nobody downstream could dispute it. The second time, it was enterprise access control that decided, for every request, who was allowed to see what, and left a record of each decision. The third time, it is AI agents whose actions go on a tamper-evident record.
Same problem, three generations of technology. Each time the question was the same. Something happened out of your sight, and now you have to prove it to someone who has no reason to trust you.
I spent a decade on privacy and shared records before I built an agent. When agents arrived, I did not have to learn the accountability side. I had to learn the agent side, and we learned it the honest way, by shipping to customers who pay.
Today SAIN has paying customers in title and advisory services, and Sonara is in production for a federal defense program. Our values are transparency, accountability, and efficiency, and the product is how we practice them.
What SAIN OS is
SAIN OS is the vision made into product: an operating system with built-in apps, and the accountability layer for AI in regulated industries. It has two kinds of apps.
Voice apps do the work: Aura, Sonara, IRIS, Oracle, and ForgeX.
- Aura is the agent that does the work. It lives in iMessage, where non-technical professionals already are, and a new agent sets up in under 10 seconds.
- Sonara trains people on real work and keeps a record of what they learned. It is in production for a federal defense program.
Trust tools work across all of it:
- Anchor'd keeps a tamper-evident record of every action, approval, and draft.
- Guardian watches agents and waits for a person's decision when something is out of the norm.
- Sentinel screens files for hidden models, offline, and flags possible tampering for review, so you know what is entering your environment.
Tap through how they fit around one piece of work:
For files that cannot leave the building, SAIN Edge puts Super Intelligence (SI) in a box on your own network. There is no standing remote access, your staff review every answer, and each installation is a custom quote.
Training matters more than it looks. Accountability without reward reads as surveillance. If the only thing a record ever does is catch people, they will resent it and route around it. The same record that shows what went wrong also shows what went right, so good work can be recognized with evidence. There is more on that in AI that trains people and proves it.
What I am asking of you
If you run a small law firm, a title company, a finance practice, or a government contracting shop, you have probably already hit the wall I am describing. You can see what an agent would save you. You cannot yet defend it to the people you answer to.
Start with one workflow. Identify who approves the actions, what data needs protection, and what evidence you need afterward. Those questions help you decide where AI can be useful and what safeguards it needs. The practical guides are here: small law firms, title companies, advisors and accountants, and government contractors.
If you are building agents yourself, I am asking you to take the record seriously before a regulator makes you. And if you are simply skeptical, stay skeptical. Skeptics are who we build for.
I will keep writing here about what we are building, why, why us, and why now. Every piece will point back to this one. If you are weighing where an agent fits in your work, we are glad to listen first.