PLN 2.14m of margin recovered in nine months

The plant runs several hundred jobs a year, and every one of them is costed in advance: materials, labour hours, subcontracting. The problem was not the estimates themselves, but that nobody could see when reality started drifting away from them. Costs reached the accounting system late, and the margin report was produced once a month, by hand, from three different sources.

In practice, the decision to adjust a price, change the process or talk to the client came long after the job had stopped adding up. Some of the losses could be clawed back on later batches. Some could not, because the contract had already ended.

The symptoms the plant came to us with fell into a recurring pattern:

• The margin report ran on a monthly cycle, while production decisions were made every day. There was nothing in between.
• Three sources of truth about cost - the production system, accounting and the process engineers' spreadsheets - only agreed after manual reconciliation.
• Overruns came to light after the job was closed, when the only response left was a lesson for next time.
• Knowing which jobs were risky lived in the heads of two people and did not survive either of them going on holiday.

The rollout started by connecting the sources, not by changing processes. Omnira Predict reads the production system and accounting through their existing APIs, with no data migration and no new duties for the planners. The first weeks were observation only: the model learned what a normal job looks like in this particular plant.

Only then did the alerts go in. Not on a fixed threshold, but on deviation from the pattern: a job that is losing margin faster than its stage of completion would explain raises its hand by itself. The alert reaches Omnira Desk as a case with full context - the costing, the cost so far and where the deviation is coming from.

Desk carries that case through to a decision: it assigns it to the right person, keeps an eye on response time and leaves a record of what was decided and why. That closes the loop that was missing before - the signal does not get lost in an inbox, and after a quarter you can see which kinds of deviation keep coming back.

The rollout steps, in order:

• Connecting the production system and accounting via API, read-only.
• An observation period: the model builds cost patterns from job history while the team works as before.
• Switching on deviation alerts in Omnira Predict and calibrating their sensitivity with the process engineers.
• Wiring the alerts into Omnira Desk: case routing, response times, a record of decisions.
• A review after the first full quarter, with thresholds adjusted based on the decisions that were made.

The most important change is structural: the conversation about margin stopped being a month-end ritual and became day-to-day work on live jobs. Overruns still happen - series production is not deterministic - but they are caught while the job is running, when a correction is still possible.

The figure in the headline of this page is the sum of the corrections made before the jobs closed: renegotiations, process changes, decisions to halt unprofitable runs. It is calculated on the client's side, on the client's data, and marked as provisional until the agreed case study is published.

The second effect shows up less in the numbers: knowledge of risky jobs stopped being personal. A new planner sees the same signals and the same decision history in Desk as someone with ten years in the job.