An assistant that cannot touch the schedule
Most AI features are sold on what they can do to your data. For a production schedule, the more useful promise runs the other way — here is why Orkyo's assistant reads, shows and proposes, but never moves a job on its own.
There is a particular demo that every scheduling tool wants to give right now. Someone types “reschedule the Henderson job to next Tuesday and shuffle whatever it conflicts with,” and the plan rearranges itself. It looks like the future. It is also, for most shop floors, completely unusable — and the reason has nothing to do with how good the model is.
The schedule is a set of promises
A production schedule looks like data. It is not. It is a stack of commitments: a delivery date someone gave a customer, a machine someone booked with maintenance, an operator whose shift was arranged around a job, a subcontractor who was told to expect material on Thursday. The rows in the plan are the visible part. The obligations are the part that hurts when they move.
This is why “the AI rescheduled it for you” is a worse sentence than it sounds. When a planner moves a job, they carry the surrounding knowledge — that this customer will accept a day’s slip and that one will not, that the second shift is already short, that the machine has a habit nobody wrote down. When a tool moves a job, it carries what it can see in the database. The gap between those two is where the angry phone call comes from.
The uncomfortable part is that the model is usually right about the mechanics. It really can find the conflict, and it really can find a slot that resolves it. It just cannot know which of the promises around that slot are load-bearing, because nobody has ever written them down anywhere it can read.
So we drew the line at data
Orkyo’s assistant reads your workspace and answers questions about it. It explains why a conflict happened and what would resolve it. It can propose a specific change — this request, this time, this machine — and show you exactly which values would change.
Then it stops and waits.
Nothing is applied until you select Apply, and when you do, Orkyo makes the change as though you made it yourself: your permissions, your validation, your audit trail. The assistant has no separate authority. It cannot act while you are not looking, because it cannot act at all.
This is a real constraint and not a setting we ship switched on. There is no autonomous mode behind a flag, no “trusted” tier where it starts writing. We would rather ship an assistant that is boring in the demo and safe in a plant with sixty people depending on the plan.
The one thing we did let it move
There is exactly one thing the assistant changes without asking, and it is your screen.
If you ask where the conflicts are, it opens the Conflicts tab. If it mentions a specific request in its answer, it can open that request. We spent a while deciding whether this crossed the line, and concluded it does not: opening a page changes nothing in the workspace, the page still enforces its own permissions, and you can navigate away with one click. It is the difference between someone pointing at a drawing and someone editing it.
It also removed a small, constant annoyance. The old failure mode was an assistant that gave a genuinely good answer ending in “you can see this under Insights, then Conflicts” — correct, useless, and four clicks from where you were standing. Showing beats describing when what you need is the screen.
What it sends, and to whom
The assistant runs on your own Anthropic API key. Your administrator adds it once, your messages and the workspace data needed to answer them go to Anthropic under your own agreement with them, and Anthropic bills the usage to your key.
That is inference, not training. Orkyo does not supply the key, does not resell tokens, and does not put your production data through a model we control. If you remove the key, the assistant stops for everyone in the workspace immediately.
It also reads strictly through your own permissions. Nothing about asking a question in natural language widens what you are allowed to see — if you cannot open a site, neither can the assistant while it is answering you, and if you cannot edit the schedule, it will explain the fix and show no Apply button at all.
Conversations are saved now, so a thread survives a reload and follows you between devices. They are saved per person: a conversation can quote data that only you can see, so nobody else in your workspace can read yours. Orkyo keeps your twenty most recent and deletes the oldest as new ones arrive.
Try it on someone else’s data first
The public demo has the assistant switched on, running against a workspace with a real-looking plan in it. Ask it why something is in conflict. Ask it where a machine is. Ask it to propose a fix, and then refuse the proposal and watch nothing happen.
The demo has a daily limit on interactions, because everyone shares one account there — when it runs out, it will say so and offer you a guided demonstration instead.
The claim we are actually making is a modest one. Scheduling software has spent years getting good at telling you that something is wrong. The assistant is our attempt at the next sentence: telling you why, in the words you would have used, and then keeping its hands off the plan until you decide.
Further reading: AI assistant in the documentation, and what Orkyo is not if you would rather read about the things we decided against.