The approach
“AI planning” usually means: a language model guesses a schedule. Plainova is built differently — because a plan you have to double-check is not a plan.
They sound like one job. They are three — and each one is best answered by a different kind of machine.
Breaking “website relaunch” into phases and tasks takes knowledge of the world and of language. This is what large language models are genuinely good at.
Estimating effort and demand is statistics: learned from history, honest about uncertainty. A guess in confident prose is still a guess.
Distributing tasks across people with skills, capacity, vacations and deadlines is mathematical optimization. Here, “roughly right” is simply wrong.
Plainova does not push your plan through one model and hope. Each layer has exactly one job.
Understands what you want, drafts task breakdowns, translates wishes into rules — “Maria is out in August, the launch must come before the trade fair” — and explains results in plain sentences. Powered by large language models. In Plainova today.
Forecasts demand and effort from your own data with statistical time-series models — reproducible numbers instead of confident guesses. In Plainova today.
Computes the schedule itself: who does what, when — every rule provably respected, the same input always producing the same plan, and a reason behind every date. Built on exact optimization, not on guessing. In development — the step from understanding plans to making them.
We use language models every day — for what they are good at. Scheduling is not on that list, for four reasons that do not go away as models improve.
A language model can produce a schedule that books someone during their vacation or ignores a dependency — and with fifty tasks, somewhere it will. An optimizer cannot: rules are constraints, not suggestions.
Same input, different plan tomorrow. Plans in Plainova are versioned and comparable — that only means something if recomputing a plan yields the same plan.
Real portfolios have thousands of tasks and bookings. That exceeds what fits into a model’s context window — and long before that, its arithmetic stops being reliable.
Ask an optimizer why a task sits in week 40 and the answer is causal: its predecessor ends in week 39 and no one else has the skill. Ask a language model and you get plausible prose.
The most valuable answer a planning system can give is sometimes “not like this”. The planning core does not silently deliver the least-bad schedule — it names the conflict: “Not achievable by Nov 30: after Maria’s vacation, 60 hours of frontend capacity are missing. Options: move the deadline by 8 days, assign task X to Josef, or reduce scope Y.”
And whatever the AI produces — a draft, a schedule, an answer — it arrives as a proposal, as its own plan version next to yours. Nothing is saved silently. You compare, accept, adjust or discard. That principle already holds everywhere in Plainova, and the planning core changes nothing about it.
Three months free, your own data model, and AI that shows its work. Questions about the approach are welcome.