One person should be able to run a business that serves customers around the clock and still enjoy a quiet evening. The hard part is giving an AI agent enough authority to finish a job while keeping the decisions that matter in the owner’s hands.
Consider someone who runs an electrical supply business in Hyderabad. After closing time, a contractor asks for 400 metres of cable for a job the next morning. Some of it is on the shelf. A regular supplier may have the rest. There’s a sale to be made if someone can put the pieces together tonight.
Here’s what I’d want an agent to do. Check the grade of cable the contractor needs and how much is in stock. Find the balance with an approved supplier. Get a price, include transport and work out whether the sale leaves enough margin. Give the contractor a quote and a delivery time the business can stand behind.
If the contractor accepts, take the order. Buy the missing cable on the owner’s behalf, check that the supplier order went through and give the contractor a confirmation based on what was actually secured.
That’s a whole job one person could delegate. A small shop may have enough after-hours business to make that agent worthwhile long before it can afford a night shift. By morning, the owner could have a sale on the books, the stock accounted for and a customer who knows when to expect delivery.
What could go wrong while the owner sleeps?
Suppose the specified cable is unavailable. The agent finds a cheaper substitute at the same supplier. The quantity is right. The margin is better. The delivery will arrive on time. On almost every measure the agent was asked to optimise, it looks like a good purchase.
The grade of cable is the one thing the agent has no authority to change. An unsuitable cable installed under the wrong load can overheat. In the worst case, a purchasing substitution can contribute to a fire and put lives at risk. The owner may have delegated the purchase. The technical judgement about what is safe to install still belongs with the qualified person responsible for that decision.
This is where my expectations of automation get demanding. If the exact cable is unavailable, I want the agent to leave the order open, tell the contractor that availability is being checked and set the exception aside for review. Losing a sale is a decision the owner can live with. Silently accepting a safety risk on someone else’s behalf is far more consequential.
The same principle applies to an ordinary financial mistake. Suppose the owner gives their agents a shared overnight allowance of ₹1 lakh. One agent places an authorised ₹80,000 purchase. The purchasing system creates the order, but the connection drops before the reply gets back. The agent sees a timeout. The supplier may already be cutting the cable.
Pressing retry takes a fraction of a second. A second ₹80,000 order can tie up the working capital the owner needs tomorrow. Waking the owner at two in the morning won’t establish whether the first purchase went through. That answer has to come from the purchasing system.
Why we’re building action safety
At IronTrex, we call this action safety. The owner defines what an agent may do. The system checks those terms when the agent acts. Afterwards, it checks what actually changed. If the proposed action falls outside the owner’s authority, or the outcome is uncertain, there is a clear reason to pause.
Our ambition is for people to be able to rely on that behaviour every time. A promise of “100% reliable AI” would give a shop owner little to go on. A much better question is one they can test: If the agent changes what I approved, will it stop? If the connection fails after it places my order, can it find out what happened before it tries again?
We’ve open-sourced the core of Protected Agent Actions, the first module of Steelmatter, with a self-hostable purchasing example for Odoo 18. It includes the engine that checks the exact purchase against its authority, reserves the amount against a shared allowance and reconciles the result. Developers also get the SDK, failure tests and a tool for checking signed evidence offline.
In the public example, change the supplier after approval and Steelmatter refuses the altered purchase. Let Odoo create the ₹80,000 order, then deliberately interrupt the reply. Steelmatter keeps that amount reserved while it checks for the original order. If it finds a match, it records one commitment. The other ₹20,000 remains available. If the outcome is still uncertain, the reservation stays in place.
That capability has a precise boundary. Steelmatter can check that a proposed purchase matches the approved specification. Determining whether that specification is suitable for a particular installation takes qualified engineering judgement. The value of action safety is that the agent has to respect the decision that person made.
Who gets to build on it?
A student can run the purchasing example on a laptop and learn something a successful demo rarely teaches: a request can fail on your screen after it has already succeeded somewhere else.
An independent developer can take it further. Someone who works with local distributors knows which goods are quoted by the metre but sold by the coil, when freight wipes out the profit, and which substitutions always need a call. They can build that knowledge into an order-taking agent, charge for their work and support the businesses that use it. A startup can do the same for an entire trade.
That’s why we opened the core. I want advanced automation to be within reach of the developer working with their first customer in a small town and the business owner who could never justify building this infrastructure alone. The owner may never read a line of Steelmatter’s code. They can still ask the developer sitting across from them, “How do I know it won’t order twice?” And the developer can answer, “Let’s disconnect it and see.”
For me, the real measure of progress comes the following morning. The owner can see which customer placed an order, what the agent purchased, what it cost and whether anything needs their judgement. They can also see the sale the agent chose to leave open because the right cable wasn’t available. I’d want as much confidence in that second decision as in the first.
That is the comfort I want IronTrex to make possible. One person can take on more business, earn from opportunities that once passed after closing time, and trust that the limits they set at six in the evening will still hold at two in the morning. They can put the phone down because they know what their agent is free to do, where it must stop and how they’ll find out what happened.