The central principle: digitization is not the act of photographing paper. It is the creation of reliable, searchable, time-aware operational records that can be traced to their origin and interpreted in the context of the correct machine, shift, product and process.
A production notebook may contain remarkably valuable information: output counts, waste, reason codes, temperatures, operator observations, job numbers, changeovers and breakdown descriptions. The problem is not that the factory has no data. It is that information is fragmented across books, forms, spreadsheets and individual memories, making it difficult to compare, verify or use for decisions.
A factory can begin improving this situation without installing sensors on every machine, replacing its controllers or purchasing a comprehensive ERP. The first step is to understand how the records are created and what decisions they should support.
1. Start with a decision, not a scanning project
Suppose a plastics plant wants to understand why a production line repeatedly misses its shift target. Management may initially request a dashboard. But before selecting software, ask:
- Which output figure is authoritative: counter reading, finished pieces, packed cartons or dispatched quantity?
- Are rejected units reported separately? Are reworked units counted twice?
- Is the shift boundary at 06:00, 07:00 or a supervisor-defined handover?
- Can each entry be linked to a physical machine and a production order?
- Are production stops recorded with start and end times, or only a description written later?
These answers establish whether useful analysis is possible. A polished dashboard cannot repair missing definitions.
Practical exercise: choose one question that matters financially or operationally. Map every source needed to answer it. If a necessary fact is not captured, record that as a gap rather than substituting an estimate.
2. Inventory the records currently in use
Walk through a complete shift and inspect real—not hypothetical—documents. A useful inventory includes:
Scroll horizontally if needed to read all columns.
| Record | Origin | Typical fields | Frequency | Decision supported |
|---|---|---|---|---|
| Production tally | Line operator | Machine, product, shift, output, rejects | Every shift | Output and quality trends |
| Breakdown log | Technician or supervisor | Asset, reported time, symptom, action | On incident | Failure history |
| Quality inspection | Quality team | Batch, measurement, result, disposition | By inspection | Compliance and rejects |
| Job card | Maintenance team | Work requested, assigned, activities, closure | Per work order | Work management |
| Material issue | Stores or ERP | Stock item, quantity, transaction reference | Per transaction | Material usage |
For each document, record ownership, date range, legibility, whether forms have changed over time, required signatures and retention expectations. Photograph or scan a small authorized sample with sensitive details protected.
3. Choose the capture route per source
Structured digital forms work well for new records when staff can enter information directly on an accessible device. Use dropdowns for known machines, explicit units, sensible defaults and validation. Do not force an operator through a long form to capture four essential values.
Excel or CSV import is appropriate when teams already maintain spreadsheets. Preserve the original file and its version, map columns to canonical fields, validate rows and produce a visible exception report. Never assume that columns with similar names mean the same thing.
OCR-assisted transcription is useful for existing paper records and handwritten forms. OCR should propose values, not declare them unquestionably correct. Store the original image, extracted text, field positions where available, confidence estimates and any subsequent correction. A confidence value from a model is not a measured probability that an entire record is correct.
Read-only connections to existing systems can capture records from databases, ERP software or approved industrial interfaces where those systems exist. External access must be authorized, scoped and isolated from equipment-control writes. Digital entry inside the company’s own record-keeping workflow is different from commanding a machine.
A hybrid approach is often best. The goal is one consistent information model, not one mandatory data-entry method.
4. Create a canonical record with evidence attached
A robust production record should describe what happened, when, where, to which asset, according to which source, and how confidently it is known.
A minimal example might be:
{
"event_type": "production_shift_summary",
"site": "PLANT_A",
"asset": "LINE_03",
"shift_start": "2026-10-12T06:00:00+02:00",
"product_code": "P-104",
"good_units": 4800,
"rejected_units": 120,
"unit": "pieces",
"source_type": "paper_log_transcription",
"source_reference": "scan-2026-10-12-044",
"validation_status": "verified_against_source"
}Illustrative example only; not a Zorithal customer dataset.
In production, add stable IDs, authenticated user or system attribution, capture time, revision history, source version and retention policy. Event time may differ from the moment someone scanned the form. Both matter. If machine identifiers were renamed last year, historical records must be resolved using the configuration valid when the event occurred, not today’s label.
5. Design for ambiguity instead of concealing it
A handwritten “350” may be a quantity, temperature or job reference. An abbreviated machine name might refer to two different assets. A missing time could make a stop duration unknowable.
Use three explicit outcomes: validated, provisional, and unresolved. Rules can safely resolve familiar variations; uncertain mappings should remain marked as uncertain. High-risk ambiguity should trigger targeted review, not approval of every field on every record. The correct threshold varies by use case: a misread staff note and an incorrect quality-release value do not carry equal consequences.
Build deduplication around source identifiers and event semantics. Two separate production entries with the same quantity are not necessarily duplicates. Two scans of one signed paper form may be.
6. Build views that match real operational questions
A machine page should not display a pile of unconnected scans. It should show a timeline of eligible events, production totals with units and time windows, quality results, associated work records and links back to originals.
Where evidence is insufficient, say so. For example:
- A job card open for six hours does not prove six hours of machine downtime.
- A shift with no breakdown entries does not prove uninterrupted operation.
- A recorded production count without scheduled time and ideal rate does not support a reliable overall equipment effectiveness (OEE) calculation.
Trust grows when the system explains what it cannot establish.
7. Start with one contained pilot
A practical first pilot can focus on one production area, two or three form types and one decision question. Define success in observable terms: percentage of records linked to the correct asset, percentage with valid event times, error rate against source documents, time to find a historical record, and number of unresolved exceptions.
Run old and new processes in parallel for a limited period where feasible. Verify that imported totals reconcile, train the people entering records, set a backup procedure and prove the process still works during connectivity interruptions.
Implementation checklist
- Name one decision the data must support.
- Inventory source documents and responsible teams.
- Define asset IDs, shifts, units and product codes.
- Choose manual forms, imports, OCR or integrations per source.
- Preserve originals and field-level corrections.
- Validate time, identity, duplicates and totals.
- Separate facts, calculations and uncertainty.
- Test recovery, permissions and historical retrieval.
- Measure actual pilot usefulness before expanding.
Frequently asked questions
Do we need to connect our PLCs first? No. Many useful production, quality and maintenance records can be digitized independently. Approved read-only telemetry can be added when its benefits and security are justified.
Can handwritten OCR eliminate human review? Not reliably in every circumstance. Good systems automate straightforward fields while escalating uncertain, high-consequence or conflicting information.
Should every historical notebook be scanned? Usually not at first. Prioritize records with decision value, clear definitions, sufficient quality and a lawful retention basis.
Will digitization automatically reduce downtime? No. Digitization makes evidence more accessible. Operational improvement still depends on analysis, decisions and execution.
Where to go next
Use the accompanying Factory Data Readiness Assessment to identify the biggest gaps before buying equipment or commissioning a full system. For the distinction between maintenance records and downtime, read Job Cards, Breakdowns and Downtime.
