Why I Built GhostOne: AI Employees That Have to Prove Their Work

An AI employee should have to prove its work. That conviction is why I built GhostOne, after seeing the same thing again and again: impressive AI demonstrations that nobody would trust with real work.

The questions a company asks are not about how fluent the agent sounds. They are the ones I spent years answering in retail software. Who did this? On whose authority? Did it really happen, and can I see what was there before?

A team, not a chatbot

GhostOne gives each company a team of agents with a first name and a job, in purchasing, HR, finance or management. They work in the company's own software, its messaging, its files, its databases. People talk to them in everyday words, in the app, on WhatsApp, by email or by voice. Those channels are only ways in. Behind them, there is one engine with the same rights and the same proof.

One gate for every action

Everything rests on a single principle: each call to a tool passes through the same gate, in the same order, whatever is connected behind it.

• The right. May this person and this agent do this at all?
• The act. Is it reading, creating, sending, paying, deleting or changing a setting?
• The agreement. What the person clearly asked for goes ahead. Paying, deleting and changing a setting always require a person.
• The proof. The agent only says "sent" or "created" if the code has seen the action succeed.
• The log. Who, what and when, with the exact text.

The system is closed by default. What is missing is refused, an unknown person is restricted, and an error is displayed as blocked, never as done.

The code holds the facts, the model holds the understanding

The division of labour is the design decision I care about most. The code knows who is speaking, what is connected and what was actually confirmed. The model understands the request, even one nobody anticipated, and chooses among what it is allowed to do.

The code never tries to guess the meaning of a message from keywords. The model can never do something the code does not permit.

That is also why a system connected tomorrow works without custom development. The code only knows acts. It has never heard of a specific client or piece of software.

The value is the harness

Models improve every few months, so I treat the model as an interchangeable engine, chosen by measurement. Everything that lasts is around it: rights, approvals, proof, memory, processes and evaluation.

And "fixed" has to be earned. A behaviour counts as working when it has been tested repeatedly against the real model, then with a second client in another line of business, then in real use.

The objection

All of this slows an agent down, and that is a fair criticism. Checking rights, waiting for agreement and writing a log cost time compared with an agent left to itself. I accept that price. A company will forgive an agent for being a few seconds slower. It will not forgive one that sent the wrong message and cannot say why.

GhostOne is live in beta, deployed with a franchise network as its first client. I built it from A to Z, and it is the clearest expression of how I think AI should enter a company: as a colleague you can check.