01The engagement in brief
Northexzone · explosion-protected electrical equipment, Istanbul
Every unit provably inside its certificate, orders on time, and fewer people in harm's way
The client aimed to prove that every enclosure, control station and socket it ships stays within its ATEX and IECEx certificate, to be ready for notified-body audits at any time, to deliver customised orders on time, and to know how often people step into the danger zone of a running machine.
Client issue: Northexzone, an Istanbul manufacturer of Ex d and Ex e equipment for oil & gas, petrochemical and mining plants, engaged AIx2 to support its growth. Certificate conformity was checked by sampling, QAN and QAR evidence took three weeks to prepare, quotes depended on two senior engineers, only 75% of orders shipped on time, and danger-zone entries at six hazardous machines were not measured.
Solution: AIx2 mapped the order-to-delivery workflow and its data flow in parallel across 12 sources, including the ERP with its timesheets, digitised three years of paper forms on-premise, read every nameplate with a vision model and placed on-premise vision at six machines, three lines and the gates. Order, unit, test and certificate were fused into one organization model, tested against the QMS, ISO 9001 and ISO 10007, and ATEX, IECEx and IEC 60079; it forecasts late orders 3–5 weeks ahead, with 81% caught in the backtest.
Impact: On-time delivery rose from 75% to 95%, every unit is checked against its certificate before release, audit evidence fell from three weeks to one day, quote turnaround from 5 days to 1, and danger-zone entries from 4.8 to 0.9 per 1,000 hours. The general manager kept AIx2 running the model as a managed service, and both 2026 surveillance audits closed without major findings.
02Results, measured
Orders ship on time, quotes go out in a day, and machines are safer to work at
Baseline: 12 months to December 2025. After: orders promised May – October 2026, once the release check and delay forecast were live (n = ≈ 550 orders, ≈ 7,000 units, ≈ 900 quotes). Safety and OEE: four-week shadow mode in April 2026 vs June – October 2026.
Late-order forecast, backtested on 1,050 orders of the prior 12 months before use: 81% of late orders caught ≥ 3 weeks ahead, 22% false alarms. Both 2026 surveillance audits closed with no major findings.
03How the results were verified
KPIs taken from Northexzone's own ERP, test bench, HSE and quality records
Safety baseline measured, not remembered: cameras recorded four weeks before any alert was switched on
Same team and product range over the period
04Financial case, modelled from measured KPIs
Paid back in month 13 (January 2027) and returned 3.1 times over three years. Modelled from the measured KPIs, from contract start. Benefits ramp up from month 4 (Apr 2026) to full in month 13 (Jan 2027). Only half of the win-rate, overtime and faster-payment gains are credited to the engagement.
By today (month 9), 57% of the fee has come back. Full table and three-year view: appendix §12.
Conclusion: The engagement pays for itself. By month 9, 57% of the fee is back; it is fully back in month 13 (January 2027). Over three years it returns 3.1 times the fee. Counting recurring cash only, payback moves to month 20 (August 2027) and the return is 2.5 times. Not priced in: losing the production quality notification (≈ €160k of gross margin per month of suspension) and a serious injury at a machine.
05How the work ran: three phases
Three phases inside Northexzone’s network, each signed off by a named owner
Weeks 0–4 · Connect and map: Read-only links to 12 sources, scanned forms, shop-floor cameras and engineering meetings recorded with consent. Atlas mapped how work and data actually flow.
Weeks 4–8 · Check: Compared with Northexzone’s procedures, manufacturing best practice and ATEX, IECEx and IEC 60079 requirements. Each finding carried evidence, a count and an owner.
Week 9 onward · Forecast and act: Forecasts with confidence and drivers. Holds and proposed actions went to a named person to decide.
Human sign-off at every step: results and every model update required a named person’s approval before work continued.
06Data stayed on site
Data stayed on site
Ran inside Northexzone’s network: no cloud
Read-only connectors; personal data masked; meeting audio deleted after transcription
Test benches and plant equipment: one-way data export only, no write access (IEC 62443 zones untouched; service-provider practice per IEC 62443-2-4)
Cameras: vision on site, no video left it; people reported by line and shift, never scored one by one; face matching only for attendance, with written consent under KVKK
07How AIx2 accessed it
How AIx2 accessed it
No inbound connections; work through a jump host controlled by Northexzone, sessions recorded
Model updates delivered as signed packages, approved and installed through Northexzone’s change control
Models ran locally with documented origin; offline mode with on-site service
08Controls on Atlas itself
Controls on Atlas itself
Role-based access, encryption and a tamper-evident audit log
Software bill of materials
Incident notice within 24 h and deletion of all data at exit, set in the service agreement
09What the work required
WHAT THE WORK REQUIRED
Six capabilities the engagement depended on, and how common tool categories cover them
Process-analytics tools and data platforms are strong in their own scope, and Atlas reads from and feeds both. The right-hand column shows what each capability produced at Northexzone.
Analyse system event logs; compare with a target process — Process-analytics tool: Core strength; Data platform: After pipelines are built; AIx2 Atlas: Yes; Evidence at Northexzone: 3 order variants found instead of the 1 official route
Capture what never reaches a system: paper forms, photos, meetings — Process-analytics tool: Task mining; not paper or meetings; Data platform: Custom build; AIx2 Atlas: Built in, processed on-premise; Evidence at Northexzone: 3 years of paper forms digitised; 26% of changes found outside the change register, now under 3%
Check each item against certificates, standards and procedures, with evidence — Process-analytics tool: Limited; Data platform: Custom build; AIx2 Atlas: Built in; Evidence at Northexzone: every unit checked against its certificate before release; audit evidence 3 weeks → 1 day
See the shop floor: danger zones, output counts, attendance — Process-analytics tool: Not covered; Data platform: Custom build; AIx2 Atlas: Built in, vision on-premise; Evidence at Northexzone: danger-zone entries 4.8 → 0.9 per 1,000 h; line counts match the ERP within 0.3%
Human approval of every result and model update — Process-analytics tool: Not enforced; Data platform: Custom build; AIx2 Atlas: Mandatory; Evidence at Northexzone: every hold, forecast and model update signed off by a named owner
Deployment — Process-analytics tool: Typically cloud; Data platform: Typically cloud-first; AIx2 Atlas: Inside the client network, CPU-only; Evidence at Northexzone: no drawing or certificate left the company's server
Comparison by category, not by vendor; capabilities vary by product and edition.
10What this means for an operator
WHAT THIS MEANS FOR AN OPERATOR
What the Northexzone results imply for the Ex equipment of an operating plant
Every unit checked against its certificate before release → Every Ex device in the plant register linked to its certificate, X-conditions and last inspection (IEC 60079-17)
Customer document packages returned 30% → 5% → Supplier documents on capital projects accepted first time
Supplier certificates now checked at every purchase → Supplier and contractor certificates tracked before materials arrive on site
Changes outside the register 26% → under 3% → Management of change captured as it is agreed, including on site
Late orders flagged 3–5 weeks ahead → Long-lead items for projects and turnarounds flagged while expediting still helps
Audit evidence in 1 day instead of 3 weeks → Regulator and certification audit evidence always ready
Danger-zone entries at running machines cut by about four fifths → Exclusion zones around rotating equipment, process-unit hazardous areas and permit-to-work boundaries, with leading safety indicators by unit and shift
Where the transfer is direct: The device ↔ certificate ↔ X-conditions ↔ inspection chain, supplier-document checks, and capture of changes as they are agreed use the same Atlas components and data contracts as at Northexzone.
What an operating site adds: An operator’s Ex register holds installed devices with inspection grades and intervals under IEC 60079-17, maintained in a CMMS such as SAP PM, and many more suppliers and contractors. Northexzone is a supplier, not an operator: these elements were not part of this engagement and need their own validation.
11Timeline: nine months from first connection to a running managed service
Nine months from first connection to a running managed service
Connect · Jan; Map · Feb; Check · Feb – Mar; Forecast · Mar – Apr; Shop-floor vision · Apr – May; Managed service, Apr – Oct 2026
12Where it stands, October 2026
Release check, late-order forecast and configurator in daily use since April 2026
Cameras live at six machines, three lines and the gates since May 2026
Both 2026 notified-body surveillance audits (QAN and IECEx QAR) closed with no major findings
Each model update approved before release; monthly readouts to the General Manager
13Appendix: the engagement report
APPENDIX · ENGAGEMENT REPORT
AIx2 Atlas at Northexzone
The detail behind the CEO deck: scope, how the work ran, measured results and the financial model.
14Executive summary
Engagement: January – October 2026 (9 months) · 12-week discovery, then managed service
Northexzone designs and manufactures explosion-protected electrical equipment for oil & gas, petrochemical, chemical, power and mining plants. Every product it ships must stay within the scope of its ATEX and IECEx certificates, and every production lot must be defensible in front of its notified body. Before the engagement, that evidence lived in spreadsheets, paper inspection forms and e-mail, and the engineering behind each customised order lived with a few experienced engineers.
In the 12-week discovery (January – April 2026), AIx2 linked every order, drawing, test, nameplate and certificate into one model of the organization (the organization model). It then checked the real process against Northexzone's own quality system, manufacturing best practice and certification requirements. The release check and late-order forecast went live in April 2026. From April 2026 (month 4) Atlas has run as a managed service. A named Northexzone owner approves each result and each model update.
From April 2026 vision cameras were added on the shop floor: at six hazardous machines they warn when a person enters a running machine's danger zone, on the three lines they count output into the ERP, and at the gates they record attendance, replacing the old card system. Cameras ran four weeks in shadow mode, recording without alerts, so the safety baseline is measured.
Measured, May – October 2026: on-time delivery 75% → 95%; quote turnaround 5 → 1 day; non-conformity rate 4.0% → 1.5%; audit evidence 3 weeks → 1 day; 100% of units checked against their certificate before release; danger-zone entries 4.8 → 0.9 per 1,000 hours worked; OEE at the machining centres 58% → 71%. Modelled value: about €359k a year of recurring cash, a one-off €445k from faster customer payment (half credited to the engagement), and 7,378 hours of capacity. Full results: §11; how each one was achieved: third tab.
15Client profile
Public facts from northexzone.com; operating figures from Northexzone’s own records.
Company: Northexzone, Istanbul, Turkey. A young manufacturer of explosion-protected electrical equipment for hazardous areas, whose team brings more than 20 years of Ex design and manufacturing experience
Products: Ex d and Ex e junction boxes and enclosures · Ex e local control stations · Ex d sockets and plugs from 16 A to 125 A · panels and lighting
Certification: ISO 9001:2015 · ATEX Directive 2014/34/EU with a production quality assurance notification (QAN) · IECEx quality assessment report (QAR) · product certificates per family
Industries: oil & gas · petrochemical · chemical & pharmaceutical · power · steel · mining
Operating baseline: about 1,100 orders and 14,000 units a year · about 1,800 requests for quotation · about 120 EPC vendor document packages · 2 notified-body audits a year · revenue ≈ €6.5M · about 85 employees, 58 in production, ≈ 150,000 hours worked a year
16Processes mapped
Request for quote → Engineering & quote → Order & configuration → Certified components → Machining & coating → Assembly → Routine tests → Marking → Documents → Shipment
Quote — What happens: spec review (gas group, temperature class, zone, IP, entries, materials), configuration, bill of materials, price; Systems and traces: e-mail, customer specs, quotation spreadsheets
Order engineering — What happens: configuration drawing, entry and terminal layout, check against the certificate schedule; Systems and traces: drawings, certificate files
Procurement — What happens: certified components, castings, coatings; supplier certificate validity; Systems and traces: ERP, supplier certificates
Production and tests — What happens: machining, assembly, wiring; routine tests for each unit; Systems and traces: ERP work orders, test bench logs, paper forms
Marking and release — What happens: nameplate with Ex marking and certificate numbers; X-condition instructions; Systems and traces: photos, label files
Documents — What happens: datasheets, certificates, test reports, manuals and declarations for EPC vendor requirements; Systems and traces: document folders, e-mail
17Pain points diagnosed
1. Proving each unit is inside its certificate (Critical) — Root cause: no link between order configuration, certified drawings and the unit shipped; conformity checked by sampling; Business impact: certification and liability exposure
2. Audit readiness (Critical) — Root cause: records spread over spreadsheets, paper and folders; Business impact: three weeks of preparation per audit
3. Slow quotes (Critical) — Root cause: each quote engineered from scratch, depending on a few senior engineers; Business impact: lost tenders; slow response to EPC buyers
4. Document packages returned (High) — Root cause: packages assembled by hand for each customer; Business impact: delayed acceptance and milestone payment
5. Changes agreed informally (High) — Root cause: no change workflow linked to drawings and orders; Business impact: rework and uncontrolled revisions
6. Late delivery (High) — Root cause: no forecast across engineering, procurement and production; Business impact: penalties and lost repeat business
7. Machine safety (High) — Root cause: no record of how often people entered the danger zone of a running machine; safety measured only by injuries after the fact; Business impact: injury risk, stoppages, insurer and customer scrutiny
8. Production counts typed by hand (High) — Root cause: quantities entered from memory at shift end; Business impact: ERP output, costing and capacity figures off by several percent
18Scope and data access
All connections read-only, each with a data contract and a named owner.
ERP — Data used: orders, bills of materials, work orders, purchases, stock, invoices, timesheets by order and work order; Method: read-only connector; Privacy: finance aggregated; time reported by team
Quotation files — Data used: requests, configurations, prices, deviations; Method: folder export; Privacy: customer names → IDs
Customer specifications — Data used: EPC specs and data sheets; Method: text extraction on-premise; Privacy: access tags
Drawings and certificates — Data used: configuration drawings, certificate schedules, X-conditions, supplier certificates; Method: export; Privacy: controlled access
Test bench — Data used: routine-test results by serial number; Method: one-way data-logger export; Privacy: no write access
Inspection forms — Data used: three years of paper checklists and sign-offs; Method: scanned and read on-premise; Privacy: names → roles
Nameplate photos — Data used: marking on every unit before release; Method: photo station, image reading on-premise; Privacy: no people in frame
E-mail — Data used: customer and supplier communication; Method: metadata plus threads approved by the owner; Privacy: metadata first
Engineering meetings — Data used: order kick-offs and change discussions; Method: recorded with consent, transcribed on-premise; Privacy: transcript only; audio deleted
QMS documents — Data used: quality manual, work instructions, NCR and CAPA records, internal audits; Method: document export; Privacy: —
HSE records — Data used: injuries, near misses, safety inspections; Method: export; Privacy: names → roles
Shop-floor vision — Data used: danger zones at six hazardous machines; output counts on three lines; attendance at the gates; Method: cameras with vision models on Northexzone's server; counts written to ERP work orders; Privacy: no video leaves the site; people reported by line and shift only; face matching for attendance with written consent under KVKK
Example of the single timeline Atlas builds for a unit, its order and the machine that made it (customer and certificate numbers removed):
[2026-05-14 11:02] UNIT LJB2-W-…-0418 · nameplate read · "Ex db IIC T6 Gb · IP66 · −30…+55 °C" · matches certificate scope · X-condition sheet attached · released after quality approver sign-off [2026-05-14 11:20] UNIT LJB2-W-…-0418 · routine test · insulation 520 MΩ · continuity pass · bench 2 [2026-05-15 09:40] ORDER …-0377 · change agreed in kick-off: "add 2 × M25 entries on side B" · not yet in change register → sent to engineering lead [2026-05-15 14:07] MACHINE MC-2 · spindle running · person inside danger zone 6 s · alert to operator and shift supervisor · zone cleared · no stop required
19How the work ran
Each step ended with results reviewed and signed off in Atlas by the named Northexzone owner.
1 · Connect — What we did: 12 sources with data contracts; ERP and test bench read-only; 3 years of forms and certificates scanned; photo station at release; Output signed off: one evidence record per order, unit and certificate
2 · Map — What we did: linked request, order, work order, serial number, test and shipment; found changes agreed outside the system; Output signed off: map of how work really flows and where data comes from
3 · Check: own rules — What we did: quality manual, work instructions, control plan and certificate schedules; Output signed off: unit-level conformity and deviations with counts
4 · Check: best practice — What we did: ISO 9001, ISO 10007 configuration management, engineer-to-order references; Output signed off: distance to best practice
5 · Check: certification — What we did: ATEX production quality assurance, IECEx QAR, IEC 60079 series, EPC client specifications; Output signed off: compliance matrix: met · partial · not met · no evidence
6 · Monitor and forecast — What we did: check of every unit before release, live audit evidence, daily late-order forecast; Output signed off: release check with human approval; every model update approved before use
7 · Shop-floor vision — What we did: cameras at six hazardous machines, three lines and the gates; four weeks of shadow mode to measure the baseline; then alerts, output counts into the ERP and attendance; Output signed off: danger zones and alert rules approved by the HSE lead and Production Manager
20Findings by level
Strategy: growth targets not linked to engineering capacity or certificate coverage
Detailed processes: 3 order variants instead of 1; 26% of changes agreed outside the change register
Outcomes: document-package acceptance not tracked against payment; safety measured only by injuries, with no leading indicator
Data: most quality records on paper; configuration rules in unversioned spreadsheets; non-conformity causes not fed back to design or supplier rating
21What the organization model answers
Which order will ship late? — Signals: engineering load, component lead times, changes, work-order progress; Warning time: 3–5 weeks; Validation: backtested on 1,050 orders of the prior 12 months: 81% caught, 22% false alarms
Is this configuration inside the certificate? — Signals: certificate schedule, drawings, bill of materials; Warning time: at quote time; Validation: rules approved by the engineering manager
Which lot is at risk of non-conformity? — Signals: supplier lot, machining centre, test drift; Warning time: before assembly; Validation: reviewed weekly with quality
Which supplier will be late or non-conforming? — Signals: delivery history, certificate status, lot quality; Warning time: 4 weeks; Validation: reviewed with purchasing
Where is the next injury most likely? — Signals: danger-zone entries by machine, task and shift; machines running with the operator away; overtime; Warning time: weekly; Validation: reviewed with the HSE lead
Is the line on plan this shift? — Signals: vision output count against the plan, stops by cause; Warning time: hourly; Validation: shift supervisor
Where would the next audit find gaps? — Signals: evidence coverage per requirement, open corrective actions; Warning time: continuous; Validation: reviewed monthly by quality manager
Technical note: survival analysis, gradient boosting and anomaly detection on CPU; the vision models run on a local inference server. Everything stays on Northexzone's own premises.
22Check findings (summary)
Details and counts were shared with Northexzone only. Each finding had evidence, an owner and a due date in Atlas.
Own rules — Area of finding: traceability between order configuration and certificate schedule; completeness of test and X-condition records; Owner: Engineering, Production, Quality
Best practice — Area of finding: configuration management; supplier certificate checks at purchase; standard configurator; Owner: Engineering, Purchasing, Sales
Safety — Area of finding: machine safeguarding and leading indicators against ISO 45001 practice; break cover at hazardous machines; Owner: Production, HSE
Certification — Area of finding: digital evidence for production records and routine tests; completeness of customer document packages; Owner: Quality, Production, Document control
23Deliverables in operation
Unit release check — Content: configuration, nameplate, test and X-condition sheet checked; exceptions held for the quality approver; Cadence: every unit
Always-ready audit pack — Content: ATEX, IECEx and ISO 9001 evidence per requirement; Cadence: always ready
Order configurator — Content: rules from certificates and past orders; out-of-scope options flagged to the engineer; Cadence: live
Document package builder — Content: EPC package assembled from the requirement list, checked by the document controller; Cadence: every order
Late-order forecast — Content: probability of late shipment with drivers; Cadence: daily
Machine safety zones — Content: danger-zone alerts at six machines to operator and supervisor; weekly review by machine and shift; Cadence: live
Line count and attendance — Content: output counted by vision and booked to ERP work orders; attendance replacing the card system; Cadence: live
OEE and stops board — Content: availability, speed and quality losses by machine and shift, stops by cause; Cadence: live
Supplier and component watch — Content: certificate validity, delivery and lot quality; Cadence: weekly
Executive readout — Content: risks, progress, next actions; Cadence: monthly
Decision-support agents propose actions (compliance reminders, supply alerts, risk escalations, document drafts); a named person approves each one.
24Measured results
Baseline: 12 months to December 2025. After: May – October 2026, once the release check and forecast were live. Safety and OEE: four-week shadow mode in April 2026 vs June – October 2026. Day-90 readings are not used as results. How each result was achieved: third tab.
SAFETY · six hazardous machines · before = four-week shadow mode, Apr 2026 · after = Jun – Oct 2026
Danger-zone entries — Definition: person inside a running machine's danger zone, per 1,000 hours worked; Before: 4.8; After: 0.9; Sample: ≈ 95,000 hours worked; Source: vision, attendance
Alert to zone cleared — Definition: median, alert to person out or machine stopped; Before: 45 s; After: 8 s; Sample: ≈ 610 alerts; before = first month of alerts; Source: vision
Machine running, operator away — Definition: minutes a week across the six machines; Before: 310; After: 40; Sample: all shifts; Source: vision, machine state
Recordable injuries — Definition: count, lost-time injuries in brackets; Before: 4 (1) in 12 mo; After: 1 (0) in 7 mo; Sample: small sample: counts, no rate claimed; Source: HSE log
PRODUCTION
OEE, machining centres — Definition: availability × speed × quality; Before: 58%; After: 71%; Sample: 4 machining centres; Source: vision count, ERP, test bench
Unplanned downtime — Definition: minutes per shift, machining; Before: 74; After: 31; Sample: all shifts; Source: vision, ERP
Line count vs ERP — Definition: gap between units counted and units booked; Before: 6.5%; After: 0.3%; Sample: all units; Source: vision, ERP
Schedule adherence — Definition: work orders finished in the planned shift; Before: 68%; After: 90%; Sample: all work orders; Source: ERP
First-pass yield at routine test — Definition: units passing first time; Before: 93%; After: 98.5%; Sample: n = ≈ 9,300 units; Source: test bench
QUALITY AND CERTIFICATION
Units checked against certificate — Definition: automated check before release; Before: sampling; After: 100%; Sample: all units; Source: ERP, drawings, test bench, nameplate photos
Non-conformity rate — Definition: units with a non-conformity; Before: 4.0%; After: 1.5%; Sample: same units; Source: QMS, ERP
Cost of poor quality — Definition: rework, returns, penalties and credit notes ÷ revenue; Before: 3.4%; After: 1.6%; Sample: annualised; Source: ERP
Audit evidence preparation — Definition: per surveillance audit; Before: 3 weeks; After: 1 day; Sample: 2 audits; Source: all sources
Major audit findings — Definition: per notified-body audit; Before: 3; After: 0; Sample: 2 audits; Source: audit reports
Changes outside register — Definition: share of engineering changes; Before: 26%; After: < 3%; Sample: all changes; Source: e-mail, meetings, ERP
DELIVERY AND CUSTOMERS
On-time delivery — Definition: orders shipped on or before promise date; Before: 75%; After: 95%; Sample: n = ≈ 730 orders; Source: ERP
First promise date kept — Definition: orders shipped without the promise date being moved; Before: 61%; After: 88%; Sample: same orders; Source: ERP
Late orders announced early — Definition: customer told ≥ 1 week before the promise date; Before: 15%; After: 90%; Sample: late orders only; Source: ERP, e-mail
Customer chase e-mails — Definition: "where is my order" e-mails per order; Before: 2.6; After: 0.6; Sample: same orders; Source: e-mail
Document packages returned — Definition: returned at least once; Before: 30%; After: 5%; Sample: n = ≈ 80 packages; Source: e-mail
Document package acceptance — Definition: days from first submission to EPC approval, median; Before: 34; After: 12; Sample: same packages; Source: e-mail
Days sales outstanding — Definition: receivables ÷ revenue × 365; Before: 95 days; After: 70 days; Sample: month-end average; Source: ERP
SALES AND ENGINEERING
Quote turnaround — Definition: request received to quote sent, median; Before: 5 days; After: 1 day; Sample: n = ≈ 1,200 quotes; Source: e-mail, ERP
Quote win rate — Definition: won ÷ quoted value; Before: 22%; After: 28%; Sample: same quotes; also driven by price and market; Source: ERP
Engineering hours per quote — Definition: all engineering time booked to quotes (≈ 3.8 FTE before); Before: 3.5 h; After: 1 h; Sample: same quotes; Source: ERP timesheets
Technical RFQs only the two seniors could answer — Definition: share of technical RFQs answered by the two senior engineers; Before: 78%; After: 30%; Sample: same quotes; Source: e-mail metadata
SUPPLY
Supplier certificate check at purchase — Definition: certified components checked; Before: 82%; After: 100%; Sample: all purchases; Source: ERP, e-mail
Supplier on time, in full — Definition: purchase lines; Before: 80%; After: 92%; Sample: all purchase lines; Source: ERP
PEOPLE · reported by line and team only
Labour efficiency, by line — Definition: standard hours earned ÷ hours on site; Before: 71%; After: 84%; Sample: 3 lines; Source: vision count, attendance, ERP routings
Hours on site not booked — Definition: gap between attendance and timesheet hours; Before: 11%; After: 2%; Sample: all production staff; Source: attendance, ERP timesheets
Overtime per production worker — Definition: hours a month, team average; Before: 19; After: 9; Sample: 58 workers; Source: attendance
No other changes in team size, product range or pricing policy over the period. Safety results are reported as counts and leading indicators; one year is too short to claim an injury-rate trend.
25Financial model
Modelled from the measured KPIs. One horizon: annualised at the performance measured May – October 2026. Contract start: January 2026. At month 9 (October 2026) year 1 is in progress, so year-1 figures are partly measured and partly projected. Recurring cash, one-off working capital and capacity reported separately.
Extra quotes won — Type: Cash; Calculation: €20M quoted × (28% − 22%) = €1.2M revenue × 30% gross margin = €360k; 50% credited to the engagement; € / year: 180,000
Less rework — Type: Cash; Calculation: 14,000 units × (4.0% − 1.5%) = 350 units × €220; € / year: 77,000
Fewer late-delivery penalties — Type: Cash; Calculation: 1,100 orders × (25% − 5%) = 220 orders × €5,900 average value × 3% penalty; € / year: 38,940
Less overtime — Type: Cash; Calculation: 58 production workers × (19 − 9) h a month × 12 = 6,960 h × €18 overtime cost = €125k; 50% credited, since order intake also drives overtime; € / year: 62,640
Recurring cash — € / year: 358,580
Faster customer payment — Type: One-off cash; Calculation: (95 − 70 days) × €6.5M ÷ 365 = €445k; driven by document packages accepted first time; 50% credited, since payment terms and collections also matter; € / year: 222,603 once
Engineering time — Type: Capacity; Calculation: 1,800 quotes × (3.5 − 1.0) h = 4,500 h × €45 (from ERP timesheets); € / year: 202,500
Document rework — Type: Capacity; Calculation: 120 packages × 25% fewer returns × 16 h = 480 h × €45; € / year: 21,600
Audit preparation — Type: Capacity; Calculation: 2 audits × 14 days × 2 people × €320 a day (448 h); € / year: 17,920
Production counts no longer typed — Type: Capacity; Calculation: supervisors and clerks: 1,950 h a year × €22; € / year: 42,900
Capacity total — Calculation: 7,378 h ≈ 4.5 FTE at 1,650 h; € / year: 284,920
Payback against the fee, from contract start in January 2026 (benefits ramp from month 4, April 2026, to full in month 13, January 2027) | Payback
Recurring cash with the full win-rate gain (€539k / year): month 16 (Apr 2027)
Recurring cash + half the one-off payment gain (€223k): month 13 (Jan 2027)
… + half the freed capacity: month 12 (Dec 2026)
Recurring cash (year 1 at 37.5% of full, from the ramp) — Year 1: 134; Year 2: 359; Year 3: 359; Total: 852
One-off payment gain, 50% credited (released months 4–13) — Year 1: 200; Year 2: 22; Year 3: —; Total: 223
Inputs from client finance: quoted value, gross margin, rework cost per unit, penalty terms, overtime cost per hour. Annual values assume the performance measured May – October 2026 holds over a full year.
Conclusion. By October 2026 (month 9), 57% of the fee had come back in cash. The fee is fully recovered in month 13 (January 2027); counting recurring cash only, in month 20 (August 2027). Over three years the engagement returns 3.1 times the fee (2.5 times on recurring cash only). Not counted: avoided risk of shipping outside certificate scope or losing the production quality notification, which was why the General Manager sponsored the engagement; avoided injuries, which we do not put a price on; and the extra machining capacity from higher OEE.
26Timeline and governance
Jan 2026 · weeks 0–2 — Phase: Connect; Work done: 12 sources, data contracts, scanning of forms and certificates, photo station; Signed off when: every source has an owner and a contract
Feb 2026 · weeks 2–4 — Phase: Map; Work done: process map, data map, order ↔ unit ↔ certificate links; Signed off when: maps approved by General and Quality Managers
Feb – Mar 2026 · weeks 4–8 — Phase: Check; Work done: three checks, evidenced findings, compliance matrix; Signed off when: findings with counts and owners
Mar – Apr 2026 · weeks 8–12 — Phase: Forecast; Work done: release check, late-order forecast, configurator v1, dashboards; Signed off when: backtest on the prior 12 months approved
Apr – May 2026 · months 4–5 — Phase: Shop-floor vision; Work done: cameras at six machines, three lines and the gates; four weeks of shadow mode; alerts, counts and attendance live in May; Signed off when: danger zones and alert rules approved by the HSE lead
Apr – Oct 2026 · months 4–9 — Phase: Managed service; Work done: document package builder, supplier watch, decision-support agents, monthly readouts, both 2026 surveillance audits; Signed off when: each model update approved before release
Sponsor: General Manager: scope, priorities, investment
Quality Manager: certificates, notified-body audits, release check approvals
Engineering Manager: drawings, configuration rules, change control
Production Manager: test bench, photo station, shop-floor records, line counts
HSE lead: danger zones, alert rules, incident and near-miss review
AIx2: connections, modelling, checks, reporting
27What made it work
Started from the certificate: linking every order, drawing, test and nameplate to the certificate it must respect made value visible in week 4.
Paper did not block us: three years of inspection forms were read on-premise servers in Istanbul.
A simple photo station at release gave a reliable view of every nameplate.
Engineering meetings, recorded with consent, turned experienced engineers' know-how into configuration rules that the engineering manager approved.
Cameras recorded for four weeks before a single alert sounded, so the safety baseline is measured, and workers saw how the system works before it flagged anything.
People data was reported by line and shift from day one; nobody was scored individually, which kept the shop floor on side.
Everything ran on Northexzone's own server; no drawing, certificate or video left the company.
28How each result happened
HOW EACH RESULT HAPPENED
The story behind every number
What we did at Northexzone, result by result: what we saw when the systems and the shop floor were joined, what we compared it with, what we proposed, what Northexzone's managers decided, and how we checked the outcome. Atlas found, measured and proposed; Northexzone's people decided and made the changes.
29What combining the data adds
An alert tells someone that a number crossed a line. Most results below came from what Atlas could see only because it joined Northexzone's systems with the shop floor: where an order waits, which certificate a unit must respect, where two records of the same thing disagree, and where people and machines meet.
Work and wait time Each order rebuilt as one timeline from first e-mail to shipment; every stage split into working and waiting.
The certificate chain Order, bill of materials, drawing, test, nameplate and certificate linked for every unit, so conformity is checked, not sampled.
Reconciliation across systems Vision counts against ERP bookings, attendance against timesheets, packages against requirement lists, changes in e-mail against the register.
Comparison with a target As-built work compared with the QMS, ISO 10007, ATEX and IECEx requirements, SMED and ISO 45001 practice; Atlas drafts the redesign, people approve it.
Vision on the floor Danger zones, stops and output seen as they happen and joined to work orders and shifts.
Decisions, then verification Findings reach managers as decisions with options and expected effect; the choice is recorded and the outcome measured against it.
30The loop every result followed
Understand → Identify → Trace data & dependencies → Compare with target → Recommend redesign → Execute → Verify
Atlas proposes at each step; named Northexzone owners decide.
31From alert to action: why we waited four weeks
A camera that sounds an alarm on day one teaches people to ignore it. We ran the cameras in shadow mode first, then used the pattern of entries to change the work itself, and only then switched the alerts on, routed to someone who can act.
Signal People enter the danger zone of a running machine
1 · Is the work laid out for it? Entries cluster at a task or machine → redesign the task: sequence, layout, lock-out
2 · Is the shift short? Entries rise on thin or overtime shifts → fix cover and load, not people
What remains after both is what the alerts are for.
32Two engineers and every quote: capacity or culture?
When work piles up at a few people, we test capacity before we look at culture. Here the two senior engineers were overloaded and the others were not holding back: the rules lived nowhere else. So the fix was to move the knowledge, not to change behaviour.
• rules captured from recorded kick-offs, with consent, and approved by the Engineering Manager
• juniors quote within the rules; seniors review exceptions
• contributing a rule counts in how engineers are assessed
People data is shown by line and team, never scored by individual.
33How each result happened: Safety
Six machines in the hall can hurt someone badly: two machining centres, the press, the overpressure test rig, the coating booth and the saw. Before this engagement, nobody knew how often people stepped inside their danger zones while they were running. The only safety number was the injury count.
Danger-zone entries · 4.8 → 0.9 per 1,000 h · Vision zones · four-week shadow baseline · entries by machine, task and shift
We were tempted to switch the alerts on from day one. We did not. For four weeks in April the cameras only recorded, and nobody on the floor was warned or judged. We wanted a baseline we had measured rather than one people remembered, and we wanted operators to see what the cameras did before the cameras said anything about them. The count came to 4.8 danger-zone entries per 1,000 hours worked.
When we laid those entries over the ERP work orders and the shift plan, the picture was not "careless people". Most entries happened at two machines and during two tasks: clearing chips while the spindle was still turning, and changeovers where the tool trolley stood on the far side of the zone, so the operator reached across it. Overtime shifts had more entries than day shifts.
The Production Manager and the HSE lead redesigned the work from those clusters: a stop-before-clearing step, a changeover sequence that parks the spindle first, and the trolleys moved to the operator's side. Only then did the alerts go live, in May, to catch what the redesign did not. Each week Atlas compares entries by machine and shift against the previous month, so a new pattern is visible within days.
Alert to zone cleared · 45 s → 8 s median · Alert routing · escalation · response timing
In the first month of alerts we timed every one, from the alert to the moment the person left the zone or the machine stopped. The median was 45 seconds. Watching the footage with the supervisors explained why: the alert sounded only at the machine, and in a hall full of compressors and spindles the operator often did not hear it. Nobody else knew it had happened.
We changed where the alert goes, not how loud it is. It now reaches the shift supervisor's handset as well, escalates if the zone is not clear within 15 seconds, and three alerts at one machine in a shift open a review task for the HSE lead. The alert became a conversation between two people who can do something about it, which is why the time fell.
Machine running, operator away · 310 → 40 min a week · Vision × machine state · capacity test by shift
Combining the machine's running state with where its operator was gave a number nobody had asked for before: 310 minutes a week of hazardous machines cycling with nobody at the post. Our first instinct was a discipline problem. The data said otherwise.
Almost all of it happened at break times on the two shifts with the fewest people on site, when one operator covered two machines. The attendance data, which the cameras now record, confirmed that those shifts were short-staffed against the plan. It was a capacity problem. The Production Manager rebuilt the break-cover rota from the attendance data so that no hazardous machine is left mid-cycle, and operators pause the cycle before leaving; the cameras confirm it. Nobody was disciplined, because nobody had done anything wrong.
Recordable injuries · 4 (1 lost-time) in 12 months → 1 (0) in 7 months · HSE log × attendance hours × danger-zone data
When we linked the 2025 injury records to the zone data, three of the four injuries had happened at the two machines with the most danger-zone entries. That was the most useful thing we learned about safety here: the leading indicator pointed at the same places as the injuries, so it could be managed before the next one.
We are careful with this number. One injury in seven months against four in twelve is encouraging, but the sample is far too small to claim a trend, so we report counts and no rate. The evidence that the workplace is safer is in the three stories above.
34How each result happened: Production
OEE at the machining centres · 58% → 71% · Vision counts × ERP routings × test results · losses split by type · comparison with SMED practice
Before the cameras, OEE could not be calculated honestly: output was typed into the ERP at shift end and stops were not recorded at all. Once vision counted every unit and timestamped every stop, we could split the losses at the four machining centres into availability, speed and quality for each shift. Most of the loss was availability: machines standing still.
Following those stops back through the ERP showed what the machines were waiting for: material that had not been released, a program not yet checked, the right tools still in the crib. We compared the as-built changeover with SMED practice and drafted a different sequence: tools and programs kitted before the stop, material released to the machine together with the work order. The Production Manager piloted it on one machining centre for three weeks, we compared the two in Atlas, and it went to all four. The machines did not cut faster; they waited less.
Unplanned downtime · 74 → 31 min per shift · Stops by cause from the vision timeline and the ERP
Every stop got a cause from the timeline: material missing, program missing, tool breakage, breakdown. Material and program waits were the largest two, and neither was a machine fault; both were decisions made upstream, in planning and engineering. The weekly production meeting now starts from the top three causes of the previous week, each with an owner, and the board shows whether the cause came back.
Line count vs ERP · 6.5% → 0.3% · Reconciliation: units counted by vision × units booked in the ERP
In the first weeks we simply compared what the cameras counted with what was booked in the ERP. They differed by 6.5%: scrapped parts not booked, the same batch booked twice, counts typed from memory at the end of a long shift. Every number built on those bookings, unit cost, capacity and the late-order forecast, inherited the error.
Today the vision count is booked to the work order directly and the supervisor only confirms exceptions. Most of the remaining 0.3% is rework moved between work orders, and Atlas lists it for review each week. Fixing this first made every other production result in this report trustworthy.
Schedule adherence · 68% → 90% · Plan vs actual by shift · capacity vs load
Comparing the plan with what each shift actually finished showed two things. The plan assumed machine availability that the measured OEE did not support, and rush orders were inserted without anyone seeing which orders they pushed out. The plan is now built on measured availability, and before a rush order is accepted the planner sees which orders it will delay and by how much. The late-order forecast reads the same plan, so both tell the same story.
First-pass yield at routine test · 93% → 98.5% · Test failures traced to supplier lot, machine and configuration
We linked every failed routine test by serial number to its casting lot, the machining centre that cut its flamepaths, and its order configuration. Failures were not random: they concentrated in lots from one casting supplier and in units whose entries did not match the drawing. Lots from that supplier are now flagged before assembly, purchasing opened a corrective action with the supplier, and configurations are checked against the certificate when the quote is made. Fewer units reach the test bench with a fault already built in.
35How each result happened: Quality and certification
Units checked against certificate · sampling → 100% · Order ↔ BOM ↔ drawing ↔ test ↔ nameplate ↔ certificate graph · release gate
This was the reason the General Manager hired us. Each certificate defines what a product may contain: which entries, glands, terminals and operators, in which combinations. The unit on the truck has to match. We linked, for every unit, the order configuration, the bill of materials, the drawing, the routine test, the nameplate photo and the certificate schedule. When we ran the check over historic units, a small number had configurations that the certificate schedule did not list. The counts stayed with Northexzone.
The release gate now checks every unit: configuration against the certificate, nameplate marking read from the photo, routine test present and passed, X-condition sheet attached. A unit that fails any check is held, and the quality approver decides. Atlas does not release anything; it makes sure nothing is released unseen.
Non-conformity rate · 4.0% → 1.5% · NCR causes traced to design and supplier rating
Non-conformities had been recorded and closed one at a time. Linked to serial numbers, lots, machining operations and configurations, they turned out to have a handful of recurring causes, and none of them ever reached the designers or the supplier ratings. A monthly loop now sends the top causes to engineering as design rules and to purchasing as supplier scores, and Atlas shows whether each cause recurs.
Cost of poor quality · 3.4% → 1.6% of revenue · Reconciliation: NCRs × rework orders × returns × penalties × credit notes
Nobody had added up what poor quality cost in euros. Rework orders, customer returns, late-delivery penalties and credit notes lived in different corners of the ERP. Once we added them up, the total was 3.4% of revenue, more than half of it rework at assembly and penalties for late orders. Seeing it in euros changed the order of priorities. The fall to 1.6% is the financial view of the results on non-conformities, first-pass yield and on-time delivery.
Audit evidence preparation · 3 weeks → 1 day · Requirement-to-evidence map: ATEX Annex IV, IECEx QAR, ISO 9001
Before each surveillance audit, two people spent three weeks pulling evidence from folders, spreadsheets and paper. We mapped every requirement to where its evidence lives and check continuously that it is current. The audit pack is now generated from that map, and the work that used to happen in the three weeks before an audit happens every day in small pieces.
Major audit findings · 3 → 0 · Continuous evidence coverage · open corrective actions tracked
The previous majors were incomplete production records for two batches and routine tests evidenced only on paper. With evidence coverage visible per requirement all year, gaps surfaced months before the auditor arrived, each with an owner and a due date. Both 2026 surveillance audits closed with no major findings. The Quality Manager closed those gaps; we made sure they were visible early enough to close.
Changes outside the register · 26% → under 3% · Decisions found in e-mail and meeting transcripts · matched to ERP revisions
We compared changes that customers and engineers agreed in e-mail and kick-off meetings with the revisions recorded in the ERP. A quarter never reached the change register: a customer asked for two more entries by e-mail, an engineer said yes to keep the order moving, and the drawing on the shop floor stayed the old one.
When Atlas now finds an agreed change in an e-mail or meeting transcript, it drafts a change record already linked to the order, the drawing and the certificate check, and the engineering lead approves or rejects it. Production does not start a revision that is not in the register. Because the change is also checked against the certificate at that moment, an out-of-scope request is caught before anyone cuts metal.
36How each result happened: Delivery and customers
On-time delivery · 75% → 95% · Order timeline: work vs wait per stage · late-order forecast with drivers · decisions with expected effect
We rebuilt every 2025 order as one timeline from the first e-mail to the shipment, and split each stage into time someone worked on it and time it waited. The official order process existed in three variants. Most lateness was not slow work; it was waiting: for an engineer on customised orders, for certified components, for a slot on the test bench.
The late-order forecast reads the same signals every morning. Backtested on 1,050 orders before we used it, it caught 81% of late orders at least three weeks ahead. Each flagged order reaches the planning meeting as a decision, not an alarm: the drivers, the options (expedite a component, resequence, split the shipment, tell the customer) and the expected effect of each, estimated from how similar orders turned out. The planner decides and records the decision, and the outcome feeds back into the forecast.
First promise date kept · 61% → 88% · Promise-date history in the ERP
While checking on-time delivery we noticed something uncomfortable: 39% of orders had their promise date moved at least once, and on-time delivery was measured against the latest date. That flattered the number. Promise dates had been set without looking at engineering load or component lead times. They are now set at order entry from the same forecast, and we report delivery against the first promise as well, so nobody can improve the figure by moving the goalposts.
Late orders announced early · 15% → 90% · Forecast × customer correspondence in e-mail
Reading the e-mail threads of late orders, we found that customers usually learned about a delay on or after the due date, often because they asked. Now, when the forecast flags an order and the planner confirms a new date, sales receives a drafted e-mail with the new date and the reason, edits it and sends it. A late order announced three weeks ahead is a planning problem for the customer; one discovered on the due date is a trust problem.
Customer chase e-mails · 2.6 → 0.6 per order · E-mail classification per order
We counted "where is my order" e-mails per order. They peaked on engineered orders, where weeks passed between the order confirmation and the shipment without a word. Status messages now go out at three milestones, drawing approved, tested and shipped, drafted from the ERP events and sent by sales. Customers stopped chasing because they no longer had to.
Document packages returned · 30% → 5% · Customer requirement list × package contents · checked before submission
For an EPC contractor the document package is part of the product, and payment waits for its approval. We compared each customer's requirement list, extracted from the specification in the e-mail thread, with what was actually sent, and grouped the comments on returned packages. Most returns were for missing items: X-condition sheets, test reports, the current certificate revision.
Packages are now assembled from the requirement list, with every gap flagged before submission, and the document controller checks and sends them. This is the result an operator recognises fastest, because it is the same problem seen from the other side of the desk.
Document package acceptance · 34 → 12 days · Submission-to-approval timeline from e-mail
From the e-mail threads we measured how long a package took from first submission to approval. Each return added about two weeks, and returns often came in pairs. Fewer returns removed most of the delay; tracked reminders to the customer's document controller removed most of the rest.
Days sales outstanding · 95 → 70 days · Package approval × invoice × payment
When we put package approvals, invoices and payments on one timeline, the cause of slow payment was plain: milestone payments waited for document approval. DSO was a documents problem, not a collections problem. Finance now sees an approved package the day it is approved and invoices the same day. We credit half of this gain to the engagement, because payment terms and collections also matter.
37How each result happened: Sales and engineering
Quote turnaround · 5 days → 1 day · RFQ timeline: work vs wait · configurator built from certificates and past orders
We traced each request for quotation from the customer's e-mail to the quote sent. The engineering itself took hours; the days were spent waiting for one of two senior engineers, because every quote was engineered from scratch.
Working with the Engineering Manager, we turned certificate schedules and three years of past orders into configuration rules: which combinations are inside the certificate, which have been built before, at what cost. He reviewed and approved every rule. Sales now quotes standard configurations the same day, and only real exceptions go to an engineer.
Quote win rate · 22% → 28% · Quote outcome against response time, price and specification fit
Across past quotes, those answered within two days won noticeably more often than slower ones at similar prices. EPC buyers compare several bids against a deadline, and speed counts. The improvement follows from the faster quotes. Price and the market also moved during the period, so the financial model credits only half of it to the engagement.
Engineering hours per quote · 3.5 h → 1 h · ERP timesheets booked to RFQs
Because engineers book their time to RFQs in the ERP, we could see what quoting really cost: about 3.8 full-time engineers, most of it repeating designs that already existed. After the configurator, the hours went back to customer orders and new certificates, the work only engineers can do.
Technical RFQs only the two seniors could answer · 78% → 30% · E-mail metadata · knowledge dependency graph · capacity test
The e-mail metadata showed a graph we had seen in other companies: almost every technical question ended at the same two senior engineers. We ran the capacity test before drawing conclusions. The two booked heavy overtime, and the others were not refusing work; the rules simply existed nowhere but in two heads. This was a capacity problem caused by where knowledge lived, not a culture problem.
With consent, we recorded order kick-offs and change discussions, and the rules that came out of them went into the configurator after the Engineering Manager approved them. Junior engineers now quote within the rules and the seniors review exceptions. Contributing a rule became part of how engineers are assessed.
38How each result happened: Supply
Supplier certificate check at purchase · 82% → 100% · Purchase lines × supplier certificates · expiry watch
Every certified component, a gland, a terminal, an operator, needs a valid certificate of its own. We matched each purchase line in the ERP with the certificate the supplier had sent by e-mail and its expiry date. 18% of certified components had been bought without that check. Now a purchase order for a certified component cannot be approved without a matching valid certificate, and Atlas warns purchasing 60 days before a supplier certificate expires.
Supplier on time, in full · 80% → 92% · Purchase order vs receipt · supplier scorecard · late lines forecast
We built a scorecard for each supplier from purchase orders and receipts, combined with lot quality from the test results. Late lines are now forecast from each supplier's history and flagged early enough for purchasing to expedite or switch, and the scorecard is shared with the suppliers every quarter. Two suppliers improved once they could see their own numbers.
39How each result happened: People
Reported by line and team only. The cameras can see individuals; we decided with the General Manager at the start that no one would be scored one by one. That decision kept the shop floor with us, and it is the one we recommend to every client.
Labour efficiency, by line · 71% → 84% · Vision count × attendance × ERP routings, by line and shift
We compared the standard hours each line earned, units counted times routing time, with the hours its people were on site. The gap was not individual effort: it was the same waiting we found behind OEE, for material, programs and test slots, plus rework. As those causes were removed, line efficiency rose without anyone working harder.
Hours on site not booked · 11% → 2% · Reconciliation: attendance × timesheets
Once attendance came from the cameras, we could compare hours on site with hours booked to work orders. 11% of paid hours were booked nowhere: waiting, cleaning, rework nobody wanted to record. Those hours were invisible in unit costs. Unbooked time now gets a reason code, waiting, rework or training, which supervisors confirm weekly. The waiting codes fed straight into the OEE work.
Overtime per production worker · 19 → 9 h a month · Overtime against load by team · capacity test
Overtime looked like a staffing problem until we ran the capacity test: it was concentrated on the lines with the lowest OEE and on the weeks when rush orders landed. The company was paying overtime for time lost to waiting. As availability improved and rush orders became visible before acceptance, overtime halved with the same headcount. The financial model credits half of the saving to the engagement.