Inventory that diverges from reality
A robot moves a pallet and the WMS does not know, or knows late. Within a week the physical count and the system count disagree, and every downstream decision is made on a number that is wrong.
Robotics · Embodied systems · WMS integration
A tool call you regret is an incident. An actuator command you regret is an injury.
Everything else on this site argues that irreversible actions need a gate. Robotics is where that stops being an abstraction: the blast radius has mass, the rollback does not exist, and the system of record still has to agree with what physically happened.
- Physical actions do not have one
- No rollback
- What moved, against what the WMS believes moved
- Must reconcile
- Written before deployment, not after
- Safety case
The cost
Why these projects stall
Robotics pilots rarely fail on the robotics. They fail on everything the robot has to agree with.
Integration treated as the last mile
The cell works in isolation and then has to talk to an ERP built on assumptions nobody documented. That integration is the project, and it gets scoped as an afterthought.
A safety case written to pass, not to hold
Assessment produced to clear a gate instead of to describe the system, which means the first genuinely unexpected situation is outside everything that was considered.
Autonomy without a stop that means stop
An override that queues behind the current task instead of halting it. The distinction is invisible in a demo and decisive in an incident.
The build
What we build
The layer between the cell and the enterprise — which is where these programmes live or die.
WMS and ERP integration
Physical movement reflected in the system of record with a reconciliation that runs continuously, not a nightly job that hides divergence for a shift.
Action authorisation for actuators
The same gate as a tool call, with physical consequences: policy evaluated before the command, and a class of commands that never executes without a person.
Reconciliation of physical against recorded
Cycle counts, exception handling and a defined tolerance — because the interesting number is where the two disagree, not where they match.
Evidence for incident reconstruction
Commands, sensor state and the decision that led to them, retained so an investigation is a reading exercise, not an inference one.
Human oversight that works under stress
A stop that halts, not one that queues, and an operator interface that shows enough to intervene correctly at three in the morning.
Safety case documentation
Written to describe the system honestly, including what is outside the envelope, because that is the part that matters when something unexpected happens.
The economics
Where the money moves
Automation economics get modelled on throughput. The variance that decides the business case is elsewhere.
- 01Inventory accuracyDrives everything downstream — replenishment, picking, promising a delivery date. Divergence between physical and recorded is the tax that quietly eats the case.
- 02Integration scoped as the projectProgrammes overrun on the enterprise connection, not the cell. Scoping it honestly at the start is the single biggest schedule lever.
- 03Uptime and exception handlingA cell that halts on an unexpected pallet and waits for a human is a throughput number nobody modelled.
- 04Safety incidents avoidedAsymmetric, like security. The control cost is bounded and known; the incident cost is neither, and includes the programme itself.
Why this firm
We built this control for ledgers first
This is the end of the argument the rest of the site makes. A journal entry, a tool call and an actuator command are the same class of action: taken in the world, and not retractable. The control plane is the same — policy before, evidence after, a human on anything irreversible. Only the blast radius changes.
Questions
Answered straight
No. We build the layer between the cell and the enterprise — integration with the system of record, action authorisation, reconciliation and the evidence trail. The hardware and the motion planning come from specialists, and we will say so.
Because the thing that stalls these programmes is the ERP side. A robot that moves stock the WMS does not know about creates an inventory problem faster than it creates a throughput gain, and that is our home ground.
One cell, one flow, with reconciliation running from day one. Prove that physical and recorded agree under real conditions before scaling, because divergence compounds.
This is the horizon we are building toward and we say so plainly. The ERP and agent-governance parts are shipping now; the embodied work is early, and we would rather tell you that than sell you a maturity we do not have.
Next
Bring us the pilot that will not scale.
Usually the cell works and the integration does not. That is the conversation we are useful in, and it is the one most automation vendors are not set up to have.