Why Pyplan
AI knows business.
It doesn’t know yours.
Any company in your sector can license the same frontier model you use. None of them can license a model of your business. Pyplan is where you build it: an executable Corporate World Model of how your company works economically, that learns from every period you close.
For supply chain and finance teams running decisions that spreadsheets and rigid platforms can’t model.
01 · The experiment
Ask any frontier model two questions.
It will answer both fluently. Only one of the answers can be checked.
Question about business
A good answer. Alternative suppliers, prioritising customers by profitability, reallocating inventory, air freight versus stockout. Reasonable, and the same for everyone.
Question about your business
Just as fluent. It proposes a plan and, if asked, gives numbers. Neither it nor you can know whether those numbers are true, it doesn’t know your costs, constraints, lead times, or how they interact.
02 · Context isn’t enough
Data answers what is.
A model answers what if.
Connecting AI to your ERP or data lake is worth doing. But access to information about a company is not the same as understanding its mechanics. Today’s stock and last month’s costs don’t tell you what happens if you change a decision tomorrow.
For that you need the relationships between variables, not just their values.
How one decision propagates
Shows the state of the business. You observe it.
Holds the rules, constraints and dependencies. You experiment with it.
03 · The real gap
The knowledge exists. It just lives nowhere.
Before AI can understand your business, someone has to write down how it works. In most companies, each function holds a fragment, and the connections between them, where the important decisions are made, live in the heads of the people in the room.
knows the supply constraints.
knows real capacity, which isn’t nominal capacity.
knows how customers react to a price change.
knows how it all lands on working capital and cash.
Pyplan puts the fragments and the connections into one executable model: demand, inventory, supply, production, costs, investment, working capital, cash and financial statements in a single system.
One place to read “this is how our business works.”
04 · How do you know the model is right?
Every variance hides two errors.
The monthly review asks: why wasn’t the future what we imagined? Pyplan adds a second question: now that we know what happened, could our model have explained it?
Re-run the closed period with actual inputs. What’s left is what you don’t yet understand about your own business.
Model error is systematic, and contaminates every future decision.
Worked example · one month
100 units × 10 price
variable cost 6 · fixed 250
Result 150
80 units sold
Result 60
–90
The usual conversation ends here: “we sold less than planned.” Re-run the period to see what’s inside the number.
–80
Forecast error
We didn’t anticipate demand. It will always exist.
–10
Model error
Below a utilisation threshold, yield degrades: real unit cost was 6.125, not 6. The model assumed it was constant, and is biasing every low-volume scenario.
05 · The learning loop
A model that doesn’t just plan. It corrects itself.
Planning stops being only a way to anticipate the future and becomes a systematic way to understand the business better. Pick a step.
06
Re-run
Replay the closed period inside the model, swapping every assumption for what actually happened, including the decisions actually taken, not the ones the plan proposed.
The requirement that usually breaks: record decisions, not just results. If the system stores that 80 units were sold but not that inventory was reallocated in week 3, the re-run compares against a history that never happened.
06 · Where AI fits
The LLM understands the question.
The Corporate World Model simulates the consequences.
Role 1 · The interface
Interprets the question, triggers the calculation, explains the result.
The six-week supplier outage is asked in plain language, but answered by running the disruption against your capacities, contracts and mix, not against what the LLM read about supply chains in general.
Role 2 · The residual reader
Reads the gaps of every re-run and proposes structural revisions.
Spotting that gaps cluster in low-utilisation months and conjecturing that unit cost isn’t constant is exactly the inductive reasoning LLMs are good at. The decision to change the model stays human.
What Pyplan gives you to build it
Business logic as code: expressive, open, and native to the ecosystem modern AI already speaks.
Influence diagrams show every variable, dependency and calculation, so the model can be explained and challenged.
Supply chain and finance in the same system, from demand to cash flow and financial statements.
Any LLM or agent can query and run the model through MCP. Swap the LLM; keep the model.
07 · What a skeptical reader asks
Fair questions.
“We already have a model: the spreadsheet.”
True, and that’s the point. The spreadsheet already is a model of the business. What it lacks is traceability, integration across functions, the ability to run scenarios in series, and any memory of its own errors. The difference isn’t having a model or not. It’s having one that learns.
“Isn’t this just overfitting the model to noise?”
It would be, if you changed the structure after every single-period gap. The rule is the opposite of traditional variance analysis: an isolated gap is ignored; a gap that recurs with the same sign under the same conditions is a signal.
“Some things can’t be modelled.”
Agreed. A competitor’s move, a regulatory decision, how a key account reacts to a new account manager. The goal isn’t a complete model of the world, but one that explains the structural part of the business, and is explicit about where it ends.
“Is any of this new?”
The method isn’t. Control engineers call it system identification, and industrial digital twins have lived on it for years. What’s new is applying it to the economic layer of the whole enterprise, and coupling it to an interface that can hold a conversation about the results.
The asset that compounds
The question isn’t which LLM you’ll use next year.
It’s how many learning cycles your model will have behind it by then. A licensed model is the same for everyone. Yours gets better every time you close a period.
Veja o Pyplan em ação
