A feature list will not tell you what an AI agent is for. These are the jobs we are actually asked about and proposing against, described by industry and by task. Every one of them arrives in a format no two trading partners share — which is exactly why fixed-format automation fails on them.
Client names and the substance of our proposals are not given here
Some of these are at the discussion or proposal stage. None of them require you to standardise your formats first — the starting point is the paperwork your team already uses.
Does the cargo in the container match the paperwork the shipper sent? With several hundred shippers, case marks are written differently and the same item goes by different names, so any automation that assumes a fixed format falls over. We read what the documents say, check it against what is physically there, and send only the mismatches to a person.
“When is this part coming in?” arrives by email and as fax PDFs, several hundred times a day. Each one is short, but opening the core system, looking it up and replying adds up until somebody is tied to the desk. The agent reads the enquiry, looks the answer up in the core system, and drafts the reply.
Hand it a drawing as a PDF and it surfaces past jobs of a similar shape, what they were quoted, what they actually cost, and the original documents behind those figures — then reprices the materials and produces a rough figure and a spreadsheet draft. The machine does not decide the price; it assembles the grounds, and the person makes the call.
Inbound and outbound instructions, shipping documents and inspection sheets, all formatted differently by shipper — read as they are, interpreted, and passed into the warehouse management system already in use. The binding condition here is that the core systems are not rebuilt and the working environment is not disrupted.
Part numbers, lot codes and production dates that barcodes do not carry, captured as data the moment the goods arrive. Shipments split into small consignments that land at different times, so visibility is lost — this catches the record at the door. Often asked about by teams who would rather not replace their handheld terminals.
Inconsistent formats between trading partners are exactly what is costing the human time. A proposal that starts with “first, let us unify the paperwork” stops the floor, so we do not make one.
Every value read is checked against the original. Only what the system could not stand behind reaches a person, so nothing wrong slips silently into the next step.
Any proposal that involves modifying the core system raises the cost, the timeline and the disruption by a step. We look first at leaving what you run in place and taking on the work in front of it and behind it.
Give us a few dozen of the forms and emails you handle day to day, and we can tell you concretely how much of it is worth handing over — including when the answer is “not much”.
Book a 30-minute call →Yes. What matters is the shape of the task, not the industry. If any of these fit — reading documents that arrive in a different format from every partner, checking what you read against another system, turning what you found into a written reply — the same approach usually transfers.
Yes, and we would recommend starting small in any case — one process, one team. Projects that begin with a company-wide requirements exercise tend to spend the budget and the calendar before anything works. Pick one process, confirm the effect there, then widen.
Yes. Where contractual or audit requirements mean the data cannot leave, we propose a setup where all processing happens inside your network. See the on-premise page. We never use data you entrust to us to train models.
Usually. We can work on documents already held in an existing document management system or file server, and we can write results back into your systems. How far the connection goes depends on what the existing system exposes, so tell us what you are running when we talk.
This page covers only the AI agent work on document-heavy tasks. Camera and image work on the floor is collected on the Tinyboom use-case page, and results from our co-creation trials are in the news.
You do not need to have picked a process yet. We will tell you plainly what is achievable and what is not.
Book a slot →Three weeks of verification fixes the scope, the price and the schedule before you decide on the build.
Custom development →How we build when everything has to stay inside your network.
On-premise →