For managing directors, plant managers, quality directors and operations directors of discrete and food manufacturers running Business Central. Scrap prediction on real history, multi-material invoice automation, predictive maintenance and systematic non-conformance analysis. Without hype, with numbers and a deadline.
Spanish SMB manufacturing arrives at 2026 with a high level of digital maturity in companies that have rolled out Business Central together with the dv* shop-floor extensions: dvproduction for work order, routing and production report management; dvquality for quality control and non-conformities; dvscrap for structured scrap recording by root cause; dvmermas for raw-material waste tracking in food and process manufacturing. In this profile of company the data exists, is traceable and accumulates with statistical quality.
The gap, as in construction, sits on top of the data. OEE is calculated and published, but its predictive analysis is still manual. Scrap is recorded correctly, but root cause is assigned in the monthly quality meeting by intuition. Waste is measured, but parametric optimisation to reduce it is not addressed systematically. Maintenance is still mostly calendar-based preventive plus reactive when something breaks, with dvgmao as the foundation if implemented, but without the predictive layer yet. The raw material for AI is there; the AI kitchen is not.
In food manufacturing, a layer of batch traceability and sanitary compliance is added, already well covered by BC plus the relevant batch-control extensions. AI applied in this sub-sector mainly enters waste optimisation, anticipation of quality rejects on organoleptic or microbiological characteristics and consumption forecasting to adjust launches. Traceability stays with the documentation system; AI analyses and proposes without altering the record.
The entry cost for AI cases in manufacturing has dropped sharply with Azure Machine Learning, Azure OpenAI Service and Document Intelligence. The real barrier is no longer technological: it's the quality and age of the data in BC. Without twelve to twenty-four months of clean history with dvproduction and dvquality live, predictive models don't stabilise. That's why the first case is usually document automation, which doesn't need history, and the predictive case follows.
Four gaps where, with BC + the dv* shop-floor extensions and the right AI layer, the ROI is clear and the timeline is reasonable. If you recognise yourself in one of them, there is a use case to start with.
The plant closes work orders with scrap above target and finds out in Monday's dashboard or, worse, at month-end close. Corrective action arrives late, with the batch already shipped and the loss locked in. The knowledge of the veteran operator who can sense when a batch is going wrong stays trapped in one person.
Industrial scrap prediction →A typical manufacturer processes hundreds of invoices per month with multi-material lines, ancillary charges, agreed waste and supplier item codes different from the internal ones. Analytical posting to cost centre, work order and line item still depends on people with context. It is slow, error-prone and doesn't scale.
Supplier invoice automation →The maintenance strategy mixes calendar-based preventive with reactive when something breaks. Real predictive maintenance, based on real machine signals and breakdown history, is not yet applied systematically. Unplanned downtime consumes between 5% and 15% of theoretical plant availability.
Sector opportunity — pilot to be tailored →The quality system records non-conformities one by one but doesn't cross-reference history to detect patterns. The same non-conformity recurs because nobody has the bandwidth to correlate incidents against material, supplier, shift or machine. AI can read the full non-conformance history and suggest the correlations a human doesn't have time to find.
Sector opportunity — pilot to be tailored →The AI layer sits on top of the Davisa shop-floor stack you already have, or that we deploy for you before the pilot. The better the shop-floor extensions are configured, the more the model pays back. The following flow describes the typical architecture of a manufacturer with the AI Studio live.
Three typical manufacturer scenarios. These are indicative examples, not published case studies with audited figures: they show how the AI Studio lands in companies with a profile similar to yours.
Problem: 5.5% scrap rate over manufacturing cost. Detected at month-end close. The plant manager senses correlation between resin type, machine and shift but has no capacity for systematic analysis.
Proposed solution: Scrap prediction pilot on BC + dvproduction + dvquality + dvscrap. Azure Machine Learning model trained on 18 months of history. Per-work-order prediction before launch and proactive alert to the plant manager.
Expected result: Expected 20% to 30% reduction in scrap rate in the first year, which in this case means recovering 1.1 to 1.7 margin points. Indicative figures, to be confirmed against your actual history.
Problem: Finance department overwhelmed by raw-material supplier invoices with multi-material lines, agreed waste and supplier item codes different from internal ones. Four people spending half a day each typing them in.
Proposed solution: Smart posting pilot on dvinvoice-hub with Document Intelligence: document extraction, matching against the live purchase order, posting proposal to manufacturing line and work order.
Expected result: Expected 60% reduction in capture time per invoice. Person-days reassigned to control and supplier reconciliation. Conservative, indicative figures, to be validated against your data.
Problem: Real plant availability around 78%, with 22% lost between planned downtime and breakdowns. Maintenance mostly reactive. No mature CMMS and no IoT sensors deployed yet.
Proposed solution: Two-phase pilot: first implement dvgmao to organise the maintenance data and breakdown history; then, on top of six months of consolidated data, train a predictive breakdown model on the critical machines.
Expected result: In year one, with dvgmao well implemented alone, expected recovery of 2 to 4 OEE points. In year two, with the predictive model in production, another 2 to 3 additional points. Indicative figures, to be refined in the Discovery.
Indicative examples. Figures are confirmed in the Discovery against your actual data before committing to the pilot.
Honesty before hype. Four things AI does not solve in an industrial plant, no matter how much Discovery we run.
Three phases of the Davisa AI Studio method, adapted to the specifics of manufacturing. Each phase with a concrete deliverable and a fixed timeline.
Field work with your plant manager, quality manager, maintenance lead and CFO. Audit of the data in dvproduction, dvquality, dvscrap and, where applicable, dvgmao. Identification of the pilot case with the highest ROI on real history and definition of the KPI target and baseline.
Development of the chosen model or operational case. Training on history, cross- validation, integration with BC via API, production rollout and tuning with feedback from the first shifts. In manufacturing, the plant manager's and operator's adoption curve is the key for the case to stick.
Periodic retraining of the model with post-pilot data, extension to more SKUs or more lines, rollout of the next case in the roadmap (predictive maintenance if we started with scrap, or the reverse) and training of the internal team so they can interpret and act on model outputs.
Azure ML models trained on production, quality and scrap history to anticipate rejects.
AI use caseExtraction, matching and analytical posting of multi-material invoices with Document Intelligence.
BC extensionAdvanced manufacturing management on BC: work order, routing, production report and consumption.
BC extensionQuality management system on BC: inspections, non-conformities and corrective actions.
BC extensionStructured scrap recording by work order, machine, shift and root cause. Predictive data core.
BC extensionRaw-material waste tracking in food and process manufacturing.
BC extensionMaintenance management on BC: preventive, corrective, breakdown history and cost per machine.
SectorThe traditional sector page with the full Davisa proposition for discrete manufacturing.
SectorThe traditional sector page with the full Davisa proposition for food.
CalculatorCalculate the real annual loss from scrap and unrecovered OEE points in your plant.
HubFull catalog of use cases, sectors, method and Discovery of the AI Studio.
As a practical rule, between 12 and 24 months of clean history in Business Central with dvproduction, dvquality, dvscrap and dvmermas live and properly configured. We need enough volume for the model to distinguish real patterns from statistical noise: recurring batches, machine parameters, material-supplier mixes, shifts, operators and quality outcomes. If you have less than 12 months with BC in production, the honest answer is to wait or start with non-predictive cases like invoice automation or the executive close summary.
Not mandatory, but it helps. With BC data alone (work order, routing, material consumption, production reports, scrap and non-conformance records) you can already train useful models, especially to anticipate material-supplier-shift-machine mix problems. Adding real-time IoT signals (temperature, pressure, vibration, speed) gives the model more resolution and enables in-shift parametric adjustment and proper predictive maintenance. Without IoT we start with batch cases; with IoT we move into real-time.
Manufacturing with order repetition (recurring batches, frequent SKUs) and measurable parameters works very well: injection moulding, machining, stamping, extrusion, blowing, lamination, process chemistry, industrial food. Pure project-based one-off manufacturing (e.g. unique boilermaking) has less room for predictive models because each order is statistically unique. In that case the AI Studio focuses on document automation, operational assistants and per-project margin analysis, not scrap prediction.
No. AI in food manufacturing sits on top of the traceability already provided by BC + dvproduction + dvquality. The AI layer reads, analyses and proposes, but it does not modify the batch record or raw-material tracking. Traceability stays in the standard quality management system and, where applicable, the food batch control extensions. What AI adds is agility: early detection of recurring non-conformities, anticipation of quality rejects and parametric adjustment proposals to reduce scrap, all without touching the traceability documentation system.
The 8-week pilot delivers a model in production or an operational case. The real scrap-reduction curve becomes visible between months three and six, once the plant starts acting systematically on alerts and the model has been retrained with post-pilot data. For invoice automation, impact is visible from the first week. For predictive maintenance, the first savings appear in month two or three in a single avoided breakdown. Industrial patience matters and we don't sell this as a two-week case.
We frame the AI layer inside your current relationship with BC and the dv* shop-floor extensions. Your usual advisor coordinates the AI Studio engagement.
Talk to the team →We start with the 5-day Discovery on your plant. Audit of real data, technical assessment and prioritised roadmap so that in one week you know where it makes sense to invest.
Request AI Discovery →