Leaf sends the next job to the device in your hand. It can read it out through a headset and listen to the answer, or it can simply show it on the screen and wait for a tap and a scan. Same job, same rules, same record. The only thing that changes is how it is delivered.
Plenty of people do not want to talk to a computer all day, and plenty of tasks do not suit it. A packer at a bench with a screen in front of her is better off tapping. A picker with a pallet truck and both hands full is better off speaking. Leaf does not force a choice on anybody, and it is a per person setting, not a per site one.
When voice is used, the speech runs on the device itself. That makes it quick, it copes with a noisy building, and it keeps working when the warehouse wifi does not. Whichever mode you are in, the jobs already in hand carry on offline and catch up when the signal returns.
This is the honest bit. Most of what a warehouse does is not a judgement call, it's a calculation, and calculations should be exact and instant every single time. Leaf keeps the two jobs separate on purpose.
Shortest walk, best box, cheapest compliant carrier, who's free next. These are solved with established mathematics, the same methods the biggest operations in the world use. Nothing is invented, nothing is approximate, and the answer never varies with the weather.
A truck with no paperwork. A supplier who sent the wrong thing. A cutoff you're about to miss. These are the moments that used to need your most experienced person. Leaf weighs them up, explains the trade-off, and either acts or asks, depending on what you've allowed.
When a picker scans a location, the confirmation is immediate, that path has no clever machinery in it at all. Your throughput doesn't depend on how busy the intelligence is, which is exactly how it should be.
No change request. No consultant on site for a fortnight. No configuration screen with four hundred checkboxes that only one person understands. You say what you want, Leaf shows you what it would have done differently over the last month, and then you switch it on.
Leaf doesn't keep a picture of how the warehouse looks now and update it as things change. It keeps the whole story of what happened, in order, and works out the current picture from that. It sounds like a small distinction. It's the reason for almost everything good about living with it.
The scan, the timestamp, the person, the trailer, the label. Not a summary written afterwards, the actual moment, still there.
Every location it sat in, everything it was stacked with, every order it went out on. Minutes, from a standing start, at two in the morning.
Leaf replays the month and shows what was really on hand each day, with the receipts and picks behind every figure. Disputes stop being archaeology.
Every decision was recorded with what it knew, what else it considered, and how confident it was. The explanation is looked up, not made up afterwards.
Traditional warehouse software is one large program with modules bolted to the side of it. Leaf is shaped like a leaf. There is a stem, which is the small set of services that must never be wrong, and there are veins, which are sixteen AI agents, one for each part of the operation.
Every vein is genuinely an agent. It reads what is happening, reasons about it, calls on proven mathematics when a calculation is needed, and proposes what to do. The stem is what decides whether that proposal is allowed and what actually gets committed.
A vein never phones another vein. Picking simply announces that a line was picked, and whoever cares about that, layout, packing, stock accuracy, billing, hears it. Adding a seventeenth agent means adding a listener. It does not mean opening up the sixteen that already work.
Route optimisation, box packing, wave planning, forecasting and staffing maths run as exact solvers the agent calls like a spreadsheet. The agent chooses which question to ask and why. The solver answers it the same way every time.
A new agent is switched on in watching mode. It sees real work, records what it would have done, and changes nothing. Only once its answers have been scored against reality does it get to act, starting at the suggestion setting.
Speech runs on the device in your hand. Judgement runs in the agents. Deciding what is permitted, and writing it down, is plain deterministic code. Nothing on the scan and confirm path waits on a model, which is why the floor never slows down.
Leaf publishes a connection that AI assistants understand, using the open standard for it. That means you are not limited to asking Leaf questions inside Leaf. Connect the assistant your business already runs on, whether that is Claude, ChatGPT, Copilot or something your own developers built, and it can query the warehouse directly, in your words, with no export and no integration project.
The same door is open to your clients. A brand you store and ship for can connect their own assistant and ask about their own stock, their own orders and their own invoices, at two in the morning, without ringing your account manager or waiting for a report.
Nobody has to be given a login to your system, and nobody sees a row that is not theirs.
So we don't ask you to. Leaf runs quietly alongside what you have now, making its decisions without acting on them, and you compare. When its answers are better than your current system's for long enough, you start letting it drive, one area at a time.
Come and try to catch it out. Bring the exception that breaks your current system and we'll run it live, then replay it and show you why it went that way.