Testing a Plan Before You Commit to It

Runink FACE · Financial scenarios and hypothesis testing

This is a scenario for Runink FACE, the Fulfilment Autonomous Claims Engine. FACE runs on the Runink TIDE platform, but the work on this page is FACE's own.

In Short

The change has to be written down before it can be argued. A hypothesis is stated explicitly, together with the rules it touches — the reorder points, the lead times, the service commitments your business already runs on. Most of the value is in that step, and it is the step normally skipped.
What comes back is reasoning, ranked, with the rule it invoked. Each consequence is tied to the specific rule it follows from, so you can disagree with it on the merits. It is an argument you can check, not a number to accept.
Your live systems stay untouched. The engine reasons over the rules you gave it, where FACE runs, and the scenario goes to no outside AI service.
You can make it rough on purpose. Push a lane a week late. Drop a supplier. Let a load run warm. Plans that only work when everything goes right show it here, not at the quarter end.

Write The Plan Down Before You Argue It.

When a port shuts or a plant stops, you have about a day to pick a new route. The numbers that would settle it sit in four systems, and pulling them together takes longer than the choice can wait.

Where It Goes Wrong

A strike shuts a port on the Monday. By Tuesday somebody has to say whether to fly the parts in, hold the line, or ship through another port. The answer turns on what each one costs, how long each one takes, and which orders are at risk either way.

So somebody builds a spreadsheet. It takes two days. It holds one version of events, it rests on numbers typed by hand, and it is thrown away once the call is made. The next time this happens, the work starts again from nothing.

The plan is not the hard part. Getting the numbers in time is.

And nobody wants to try a new plan in the live system. That is the system the plant, the warehouse and the accounts run on. An experiment in there is not an experiment.

Who This Is For

Three people in the same room on the Tuesday, arguing from three different sets of numbers.

Operations director. What lands today is a port shut on Monday and a call to make by Tuesday, with the figures that would settle it sitting in four systems. What changes is what you walk into the meeting holding: the change written down as a plain statement, the rules it runs into, and the order in which they bite.
Head of planning. What lands today is a plan argued from reorder points, lead times and service commitments that live in three people's heads, which is how two of them spend an hour discovering they were describing different plans. What changes is the order of the work. The assumptions are written down before the argument rather than reconstructed after it, and each consequence names the rule it follows from, so a colleague can disagree with it on the merits.
Chief financial officer. What lands today is a premium freight line in the payables that nobody can attach to the decision that caused it. What changes is that the option somebody chose and the name of the person who chose it are kept together, so the question six months later is answered from the file rather than from the meeting.

What Happens Instead

This happens in Swarm Intelligence, under Simulate in FACE. Several agents work the same hypothesis together, and the scenarios in progress sit side by side, so the room can compare them. Beside it, Causal Matrix sets out what drives what in your operation, and what changes when one of those drivers is moved on purpose — useful when the question is less which rule a change hits than what it sets moving.

You state the change as a hypothesis and hand it the rules that govern the thing you are changing — reorder points, lead times, service commitments, the reserve assumption. It reasons over those rules and returns a ranked read of what follows, with each consequence tied to the rule it came from.

So what comes back is an argument, not an answer. That is the useful thing: a projection presented as a decision is worse than no projection, because it moves the judgement from someone accountable to a piece of software that is not. What this gives the room is the case laid out — which rules the change collides with, in what order they bite, and what somebody would have to believe for the plan to hold. The decision stays where it was.

Two consequences of that are worth having. Stating the hypothesis forces the assumptions into writing, which is the step teams skip and the reason two people can argue for an hour and turn out to have been discussing different plans. And because the reasoning happens where FACE runs, the scenario you are considering — which supplier you might drop, which lane you might cut — goes to no outside AI service.

You can also make it rough on purpose. Run the lane a week late. Take the second source away. Let a chilled load drift. A plan that only holds when the week goes well will fall over here, in front of you, while it still costs nothing to find out.

The run does not pick for you. It puts the options in a row. When somebody does pick one and send it on to be acted on, the option they chose and their name go on the record together, so the question six months later is answered from the file rather than from the meeting.

How You Would Know It Worked

Every figure below is yours, not ours. Write down where you stand today, because the baseline is gone for good the moment anything changes.

Hours from a disruption to a decision. Take your worst few days of last year. From the incident log and the mail trail: when the port shut, and when the new route was booked.
How many options you actually put a price on. Pull the last few big routing calls out of your records. Count the choices that had a cost worked out before the call, against the ones that were argued from the gut.
What you spend on rushed freight. The premium and air lines in your payables, a full quarter, sorted by the decision that caused each one.
Days of stock you sit on. Days of cover per line from your planning system, and how much of it is there because nobody could put a figure on the risk of holding less.

Bring one lane and your last big routing decision.

The port closure above is drawn to show the shape of the work. It is not an account of a customer engagement, and nothing on this page is a measured result. This page carries no ROI figures, percentages or customer names: the figures that matter are in your own records.

Questions To Ask Before The Next Bad Monday

What you hand it, what it hands back, and who is still making the call.

What do we have to give it?

Two things. The change, stated as a plain hypothesis — we ship through Rotterdam instead, we drop this supplier, we let this lane run a week late. And the rules that govern the thing you are changing: the reorder points, the lead times, the service commitments, the reserve assumption your business already runs on.Writing those down is the step that gets skipped, and much of the value is in it. Two people can argue a plan for an hour and turn out to have been describing different plans, which only becomes visible once the assumptions are on paper.

Is this a forecast, or a simulation of our operation?

It is an argument you can check. You hand it the change and the rules, and it reasons over those rules and returns a ranked read of what follows, with each consequence tied to the rule it came from.That is a different kind of output from a projection, and the difference is the whole point. A number presented as a decision moves the judgement from somebody accountable to a piece of software that is not. What this puts in front of the room is the case laid out: which rules the change collides with, in what order they bite, and what somebody would have to believe for the plan to hold.

Does it touch the systems we run the business on?

It reasons over the rules you handed it, and it does that where FACE runs, not at an outside AI service. The plant, the warehouse and the accounts keep running on their systems, untouched by the exercise — an experiment inside the live system is not an experiment, which is why nobody sensible runs one.The other half of that is discretion. The scenario you are considering — which supplier you might drop, which lane you might cut — is exactly the kind of thing you would not want discussed outside the building, and the reasoning happens where you can see it.

What does it do when the rules we gave it do not settle the question?

It names what somebody would have to believe for the plan to hold, and hands that back as the finding. That is the honest output when the rules run out: the belief the plan rests on, stated in a sentence, so the room can argue about the belief rather than about a spreadsheet.It restates your assumptions and follows them through, and every consequence names the rule it was drawn from.

Who picks, and what does the record show afterwards?

The person accountable picks. The run puts the options in a row rather than choosing between them, and the choosing is where the responsibility belongs.When somebody does pick one and send it on to be acted on, the option they chose and their name go on the record together. So the question six months later — why did we fly the parts in instead of waiting — is answered out of the file rather than out of somebody’s memory of a meeting.

Can we push it until the plan breaks?

That is one of the better uses of it. Run the lane a week late. Take the second source away. Let a chilled load drift warm. State the rough version as the hypothesis and see which rules it hits first.A plan that only holds when the week goes well will fall over in front of you, on a morning when finding that out costs nothing. The alternative is finding it out at the quarter end, when it costs whatever it costs.

Book a consultation