AI Intake: the process of uploading unstructured maintenance documents and using AI to extract structured event data (dates, meters, parts, labor) for human review before credential issuance.
Key Takeaways
- 1Upload PDFs, photos, or scans of work orders and invoices
- 2AI extracts event type, date, meter reading, parts, labor, and technician
- 3Parts normalization standardizes part numbers and descriptions across documents
- 4Passport Q&A lets you ask natural language questions about maintenance history
- 5Confidence scores show where AI is certain vs. uncertain
- 6Humans review every field before creating a draft credential
- 7AI never has signing authority — only humans sign credentials
The problem: manual data entry
Maintenance shops produce paper work orders. Fleet managers receive PDF invoices. Technicians snap photos of service stickers. All this valuable maintenance data exists — but it's trapped in unstructured formats.
Entering this data manually into digital systems is slow, tedious, and error-prone. A single work order might take 5-10 minutes to transcribe correctly. Multiply that by hundreds of events per asset over its lifetime, and the cost of accurate records becomes prohibitive.
The data entry bottleneck
Most maintenance management systems have a data entry problem, not a storage problem. The records exist on paper — they just never make it into the system. AI intake removes this bottleneck.
The solution: AI-assisted extraction
UMP's AI intake system handles the tedious work of reading documents and extracting structured data. Here's how it works:
Upload document
Upload a PDF, JPEG, PNG, or other image format. Scanned documents, photos of paper work orders, and digital PDFs are all supported.
OCR processing
If the document is an image or scanned PDF, optical character recognition (OCR) extracts the raw text. AWS Textract handles complex layouts and handwriting.
AI extraction
The AI model analyzes the text and layout to identify maintenance event data: what type of work, when it was done, meter readings, parts used, labor hours, technician name.
Confidence scoring
Each extracted field gets a confidence score. High confidence (green) means the AI is certain. Low confidence (yellow/red) means human attention is needed.
Human review
A human reviews every extracted field, correcting any errors. Low-confidence fields are highlighted for special attention.
Draft creation
Once reviewed, the data becomes a draft maintenance event ready for signing. The original document is attached as evidence.
What gets extracted
The AI model is trained to recognize common maintenance document formats and extract specific data fields:
Event type
Oil change, inspection, repair, overhaul, component replacement, etc.
Event date/time
When the work was performed, extracted from document headers or body.
Meter reading
Hour meter, odometer, or cycle count at time of service.
Parts used
Part numbers, descriptions, and quantities from line items.
Labor hours
Time spent on the work, often from invoice line items.
Technician/provider
Who performed the work, extracted from signatures or headers.
Document quality matters
Clear photos and readable PDFs produce better extraction results. Blurry images, faded text, or unusual layouts may require more manual correction. The confidence scores help identify where AI struggled.
Parts normalization
One of the challenges with maintenance records is inconsistent part descriptions. The same filter might be listed as "OIL FILTER", "Oil Fltr", "57899 OIL FILTER", or "Wix 57899" across different work orders.
AI-assisted parts normalization maps these variations to standardized entries, making maintenance history more searchable and comparable:
Part number matching
AI matches raw part numbers to known OEM and aftermarket catalogs, correcting typos and format variations.
Description standardization
Inconsistent part descriptions are normalized to consistent terminology for better searchability.
Cross-reference lookup
When available, the system identifies equivalent parts across different brands and suppliers.
Human override
Suggested matches are flagged for human confirmation. Users can accept, modify, or reject AI suggestions.
Better data quality over time
Normalized parts data enables better analytics: which components fail most often, which brands perform better, and which parts are used across your fleet. Clean data unlocks insights.
Passport Q&A
Once your asset has a maintenance passport with historical events, you can ask natural language questions about its history:
"When was the oil last changed?"
The last oil change was on October 15, 2024 at 1,247 hours. The next service is due at 1,497 hours.
"What parts have been replaced on the hydraulic system?"
Two hydraulic components have been replaced: the main pump seal on June 3, 2024, and the control valve on March 12, 2024.
"Are there any overdue services?"
Yes, the annual inspection is overdue by 45 hours. It was due at 1,200 hours; current meter is 1,245 hours.
Passport Q&A makes maintenance history accessible without scrolling through event timelines. Useful for pre-buy inspections, compliance checks, and quick lookups during service.
Natural language queries
Ask questions in plain English. No need to learn query syntax or filter through event lists manually.
Source citations
Answers include references to the specific events and credentials that support them, so you can verify the source data.
Context-aware responses
The AI understands maintenance context — it knows what "annual inspection" means, how hour meters work, and service interval patterns.
The trust model: AI assists, humans sign
A critical design principle: AI never has signing authority. The AI intake system is purely an efficiency tool that prepares data for human decision- making.
| Aspect | AI does | Humans do |
|---|---|---|
| Data extraction | Read documents and extract fields | Review and correct extracted data |
| Confidence scoring | Flag uncertain extractions | Decide what values are correct |
| Draft creation | Pre-populate event fields | Approve draft for signing |
| Credential signing | Never | Always (cryptographic signature) |
Full audit trail
Every AI-assisted draft shows that extraction was used. The signed credential includes the original document as evidence, so reviewers can verify extraction accuracy.
Human accountability
The human who signs the credential is accountable for its accuracy. They reviewed the extraction, made corrections, and chose to sign.
No hidden automation
AI intake is optional and visible. Organizations control whether to use it. There's no background AI modifying data without human awareness.
Confidence indicators
Not all extractions are equally certain. The AI provides confidence scores to guide human review:
High confidence
AI is very certain about this extraction. Standard document format, clear text, unambiguous value. Quick human verification is sufficient.
Medium confidence
AI found the field but has some uncertainty. Unusual format, partial match, or multiple candidates. Human should verify carefully.
Low confidence
AI is guessing or couldn't find the field. Human must locate and enter the correct value manually. Often due to poor image quality or unusual document layout.
The review interface highlights low-confidence fields so humans focus attention where it's most needed. High-confidence fields can be verified quickly; low-confidence fields require careful examination of the source document.
Supported document types
AI intake works with common maintenance document formats:
Work orders
Shop work orders with service details, parts lists, and technician signatures.
Invoices
Service invoices with line items for parts and labor.
Inspection reports
Formal inspection documents with checklists and findings.
Service stickers
Photos of door jamb stickers showing last service date and next due.
Logbook pages
Scanned pages from paper maintenance logbooks.
Component tags
Photos of component installation tags with date and meter readings.
Time savings
AI intake dramatically reduces the time required to digitize maintenance history:
| Aspect | Manual entry | AI intake |
|---|---|---|
| Simple work order | 5-8 minutes | 1-2 minutes (review only) |
| Detailed invoice | 10-15 minutes | 2-4 minutes (review only) |
| Batch of 50 documents | 4-6 hours | 1-2 hours |
| Historical backlog | Often abandoned | Achievable in reasonable time |
Unlocking historical records
Many organizations have years of paper records they've never digitized because manual entry was too expensive. AI intake makes it economically viable to build complete maintenance histories for existing fleets.
Getting started
Ready to try AI intake for your maintenance documents?
For service providers
Speed up credential creation by extracting data from your existing work orders and invoices.
Learn moreFor fleet operators
Digitize historical maintenance records to build complete passports for your assets.
Learn more