AI for manufacturers
The order arrives forty different ways. The line needs it one way.
The documented process says an order arrives. The real one is forty customers, no two formatted alike, half the payload in PDFs, and a routing rule that lives in one sales administrator’s head. Downstream, quality alerts are triaged by whoever is free, supplier certificates expire unnoticed until the audit, and the ISO evidence is rebuilt every year from scratch. These are queues, and the people who own them are the ones who know the product. The queue is where an agent belongs.
- Built on top of your ERP and MES. Nothing gets replaced.
- In one European manufacturer’s order handling, three of eleven steps needed a model (SUPALABS engagement data)
- Quality and compliance steps stay with a named person; the chase work does not
Six queues, six roles
Put an agent on the queue, not on the org chart.
Each role owns one recurring queue. It drafts, it shows its evidence, and a named person signs before anything consequential moves. The last column is the sector’s AI Boundary Map: whether the step is mostly a rule, mostly a model, or stays with a person.
| Role | What it drafts | Who signs | What it reads | Mostly |
|---|---|---|---|---|
| Order intake | The order in the ERP from whatever arrived, email body, PDF, portal export or spreadsheet, with every line matched to a catalogue code and every unmatched line flagged. | Sales administration, on the flagged lines only once the pass rate has earned it. | The shared inbox, the catalogue, customer price lists. | a model |
| Goods receipt and invoice matching | The three-way match between order, receipt and invoice, with exceptions grouped by cause: price, quantity, missing receipt. | Accounts payable, on the exceptions. | ERP purchase orders, receipts, SDI invoices. | code and rules |
| Quality exception triage | Each SPC alert or non-conformity report drafted with the batch, the machine, the last three occurrences and the supplier lot, ready for a decision. | The quality manager. | MES, the quality system, supplier records. | code and rules |
| Supplier qualification | The certificate set per supplier, ISO, REACH and RoHS declarations, conflict-minerals statements, with expiries sixty days out and a request drafted for each. | Purchasing, before a supplier is approved or renewed. | PEC and email attachments, the supplier register. | a model |
| Quote from product rules | A quote from a request, applying the configuration rules and the customer’s conditions, with any request that falls outside the rules routed to a person. | Sales, above the threshold you set. | The configurator rules, price lists, the CRM. | code and rules |
| ISO 9001 evidence | The evidence pack for each clause from records the company already keeps, with gaps listed as questions rather than filled in. | The quality manager, before the certification visit. | The quality system, training records, the document store. | a model |
The determinism ratio
How much of this is actually AI.
Three of these six queues are deterministic, matching, triage and configured quotes, and a rules engine does them more cheaply than a model ever will. The other three read documents that were written by someone else: customer orders, supplier certificates, the company’s own records. That is where a model earns its place, and it is where the evaluation suite matters most, because the cost of a wrong catalogue code is a wrong shipment. The order-handling workflow SUPALABS mapped in one European manufacturer came out at three of eleven steps needing a model; this page is the same shape.
3 / 6 queues mostly rules · 3 / 6 with a model
ProofAll engagement data, with sources →
No client names. Real patterns.
Every engagement produces the same five documents, whatever the sector. What we can show is anonymised: the pattern, the number, and where it applies.
Three of eleven steps genuinely needed a model in one European manufacturer’s order-handling workflow. The other eight were parsing, lookups, validation and routing (SUPALABS engagement data, 2024–2026).
The order-handling engagement behind these numbers, as redacted documents →Related case studies
Material Decisions Get Locked in Phase 3. The Data Arrives in Phase 5.Life-cycle assessment tells an architect what a building will cost the climate — after the materials have been chosen. We built the decision support that arrives while the choice is still open. Read the case study →
Customs Paperwork That Files Itself, in 31 CountriesEvery cross-border order needs the right customs form, the right Incoterm and the right return route — and the rules differ per destination. We built the system that decides all three and files the paperwork before the parcel moves. Read the case study →
Four Roles That Must Not See Each Other's DataAn intermediated marketplace has four kinds of user who must not see each other’s data. A pre-launch audit found a public signup path that could become an administrator — before a single paying user existed. Read the case study →
How authority is earned
Three tiers, set per task type, and you set the thresholds.
Every automated task runs at one of three levels of authority. The tier is set per task type rather than per system, so the same workflow can draft one step, wait for approval on another, and act alone on a third. A task moves up only when the evaluation suite shows the pass rate you defined as the threshold, and it moves back down the moment the monthly report shows a regression.
1
DraftedThe system prepares the work and a person finishes it. Every draft carries its evidence: the source documents, the rule or model that produced it, and its confidence. This is where every task starts, and where shadow mode runs before anything is shown to the team at all.Gated by the Exception Ledger — the task is not automated at all until its real deviations are written down.
2
ApprovedThe system prepares the complete action and a named person authorises it before anything consequential moves. The approval, the approver and the inputs go into the decision log.Gated by the evaluation suite — promotion from Drafted needs the per-step pass rate you set, measured against the golden dataset.
3
AutonomousThe system acts and the team audits afterwards, on a sample or on the exceptions. Any irreversible step stays at Approved regardless of accuracy, because a reversible mistake is a cost and an irreversible one is a liability.Gated by the monthly accuracy report — a regression alert sends the task back down a tier until the pass rate recovers.
Nothing is promoted on our say-so. The thresholds are yours, the report that tests them is a named line item, and the decision log shows every action that was taken at every tier.
The full method: five artifacts, three tiers, four rungs →How an engagement runs
Four rungs. You can stop after any of them.
0
Qualification call — free, 30 minutesA straight answer on whether there is anything here worth mapping. No deck, no proposal, no follow-up sequence.
1
Mapping Sprint — five days, with the people who do the workPaid discovery. Produces the Exception Ledger, the AI Boundary Map, the evaluation plan and either a fixed-price build quote or a written no. If it shows you nothing you did not already know, you do not pay.
2
Build — six weeks to productionScoped from what the sprint found. Majority deterministic software on top of your existing systems, every task starting at Drafted, shadow mode before anything acts.
3
Run — ongoing, and optionalMonthly accuracy report against the golden dataset, regression alerts, cost-per-run tracking. Stop the service and keep the system.
From the blog
The problems, one at a time.
How to Automate Quotes and Orders for Italian SMBs - 2026 GuideIn-depth guide: How to Automate Quotes and Orders for Italian SMBs - 2026 Guide...
Goods Receipt vs. Invoice Matching: How to Automate 3-Way Matching in Accounts PayableGoods received but the invoice doesn't match? Automated goods-receipt matching cuts AP cost per invoice from $9.84 average to $2.65 for Best-in-Class teams (Ardent Partners 2025).
SPC and Statistical Process Control with AI for Italian Manufacturers in 2026How AI transforms SPC and Six Sigma in Italian manufacturing: real-time process monitoring, automated control chart analysis, predictive drift detection. Reduce defect rates by 40-60% with intelligent statistical process control.
AI Visual Quality Control and Defect Detection for Italian Manufacturing in 2026How AI vision systems transform quality control in Italian manufacturing: automated defect detection at 99.5%+ accuracy, real-time line monitoring, classification of surface defects. Reduce scrap rates by 30-50% with computer vision QC.
AI Supplier Audits and Incoming Quality Inspection for Italian Manufacturers in 2026Automate supplier qualification and incoming material inspection with AI: risk-based audit scoring, automated non-conformity detection, real-time supplier performance dashboards. How Italian manufacturers reduce incoming defects by 50-70%.
REACH and RoHS Compliance with AI for Italian Chemical and Manufacturing Companies in 2026How AI automates REACH and RoHS chemical compliance for Italian manufacturers: substance tracking, SCIP database automation, supplier declarations management. Cut compliance costs by 40-55% and eliminate SVHC documentation gaps.
ISO 9001 Automation with AI for Italian Manufacturing SMEs in 2026Automate ISO 9001 quality management with AI: document control, CAPA tracking, internal audit scheduling, management review dashboards. How Italian manufacturing SMEs cut certification costs by 40-60% and eliminate audit surprises.
Precast Concrete: From Configurator to Automatic Factory OrderBIM-integrated configurators for precast concrete with automatic ordering to factory, delivery scheduling, NTC 2018 verification, and production tracking.
AI Automation in the Brescia Mechanical Engineering District 2026Italian-jurisdiction guide on AI automation for the Brescia mechanical engineering district (~5,000 PMI). Italian only. Read at /it/blog/automazione-ai-distretto-meccanica-brescia-2026/.
AI for Manufacturing SMEs in Italy: Complete Guide 2026How Italian manufacturing SMEs in Lombardy, Veneto, and Emilia-Romagna are using AI to increase efficiency by 40%, reduce costs, and compete globally. Practical strategies and real ROI data.
FAQ
Frequently asked questions
Do we need clean data first?Everybody has data and nobody has clean data. Building the context a role needs, which catalogue is the source of truth, which price list wins, who approves what, is part of the Mapping Sprint, and it is written down in the Exception Ledger rather than assumed. A workflow that waits for clean data waits forever.
Does this touch the ERP or the MES?It reads from them and writes to them through their interfaces. It does not replace either, and we commit to that in writing at the start. The workflow is slow because of the exceptions and the handoffs, not because of the database underneath them.
What happens when the model gets a catalogue code wrong?It starts at the Drafted tier, so a person sees every line before it is confirmed. It moves to Approved only when the evaluation suite, built from your own past orders, shows the pass rate you set. If the monthly accuracy report shows a regression, the task goes back down a tier and you are told. Nothing ships on a model’s say-so.
Is this only for large plants?No. The queues on this page exist at twenty people and at two thousand; what changes is who absorbs them. What matters is one real workflow, systems you intend to keep, and a named owner who can clear an approval. If you are large enough to run your own internal delivery team, hire it instead.
How long before the first queue is live?Five days for the Mapping Sprint, then six weeks from the start of the Build to the first workflow in production, with shadow mode before it acts. The sprint ends in a fixed-price quote or a written no.
What does it cost?The thirty-minute qualification call is free. Everything after it is quoted per engagement, because the scope is a finding rather than an assumption: pricing a build before the sprint means pricing a guess. If the sprint shows you nothing you did not already know, you do not pay for it.
Other sectorsAll sectors →
The same method, other queues.
Wholesale & B2B distributionEvery order is a promise. Most of the work is checking whether you can keep it.
ConstructionThe site moves only when the paperwork does.
Thirty minutes to find out whether there is anything worth mapping.
The call is free, and we will tell you if the answer is no. Bring one queue that costs someone an evening a week.