FieldproxyFieldproxy
Get Demo
FieldproxyAI-native field service software

Work order intake

Can AI read a PDF work order and create the customer and job?

Yes. Current AI models read both the text and the page image of a PDF, so they can pull the client, bill-to company, service site, scope, priority, due date, NTE and PO number from most work orders, including scanned ones, and fill a job from them. The safe setup is to let the AI propose the customer and job, auto-create only when every required field is found and the customer and site match records you already have, and send anything new, unclear or low-confidence to a person to confirm. Always attach the original PDF to the job, because it is the scope, the price ceiling and the PO you will invoice against.

  • AI reads layouts it has never seen, but it still misreads handwriting, faint scans and tables that run across pages.
  • Match before you create: bill-to company first, then the service site, then a new record only if nothing matches.
  • The bill-to company and the job site are often different; extract both and never merge them.
  • Auto-create only when required fields are found; everything else goes to a review queue with the PDF open beside the form.
  • Keep the original PDF on the job, unchanged, next to what the AI extracted.
Fieldproxy

This page is about reading the work order document. Getting the email into the system in the first place is covered in our page on turning emailed work orders into jobs.

Facts checked as of November 2026. Sources at the end of the page.

Step by step

How it works, step by step

The order matters. Most bad records come from creating before matching, or from trusting a field the model was unsure of.

1. Read the whole document, text and image

A digital PDF has a text layer; a scanned one is only pictures of pages. Current models handle both. Anthropic's Claude converts each page into an image and also extracts the page's text, then reads both, up to 600 pages per request (100 when the context window is under 1M tokens), and it cannot open password-protected or encrypted PDFs. OpenAI's API likewise sends both extracted text and page images on vision models such as gpt-4o and later.

2. Extract into a fixed list of fields

Ask for the same fields every time, not a summary: client work order or PO number, bill-to company and billing address, service site name and address, on-site contact, trade, scope text, priority, due date or SLA, NTE (not-to-exceed amount), and any line items with quantities. For each field, keep the value, the page it came from and how sure the model was. A field that is not on the document should come back empty, never guessed.

3. Match the bill-to customer

Look up the paying company by name, domain of the sender and billing address. Work orders from facility management companies are often billed to the management company, not to the retailer whose store needs the repair, so the bill-to is the customer and the store is a site under it.

4. Match the service site

Normalize the address first (street suffix, unit, ZIP), then compare it to the sites already under that customer. The US Postal Service's Publication 28 standard exists for exactly this: uniform address formats so records match. Store number or site ID, when the client prints one, is a better key than the street address.

5. Check for an existing job

Search open jobs for that customer by the client's PO or work order number. A revised PDF for the same PO should update the job and add the new file, not create a second job.

6. Create, or send to review

If every required field is present, the customer and site matched, and no field was flagged as uncertain, create the job in a ready status. If the customer or site is new, or a required field is empty or uncertain, create a draft in a review status. The reviewer sees the PDF and the extracted form side by side and confirms or corrects each flagged field.

7. Attach the original and log what the AI did

Store the PDF on the job as received. Keep the extracted values and who confirmed them, so a later dispute about scope or NTE can be settled from the source document, not from what was typed.

Where it breaks

Why it goes wrong

These are the problems buyers raised on our sales calls: facility maintenance coordinators, electrical contractors, equipment dealers and installation networks that receive work as PDFs.

A new customer for every store

A two-owner facility maintenance coordinator with three main retail clients explained that the work is billed to the facility management company, not to the retail brand on the store sign. An extractor that creates a customer from the most prominent name on the page makes a new customer per store and sends invoices to the wrong company.

Do I have to set up clients first?

The same buyer asked whether clients must exist before a job can be created. If the AI is allowed to create customers freely, near-duplicates pile up ("ABC Facility Svcs" and "ABC Facility Services Inc"). If it may only match, every new client stalls. The answer is in between: match automatically, propose new customers for a person to approve.

Scope and line items, not just header fields

A commercial electrical contractor wanted an emailed PDF contract turned into a job and a quote, with approval before it went anywhere. Header fields (site, date, PO) extract well; line items in tables that wrap across pages, with quantities and units, are where values get dropped or shifted a row. Line items need a person's check before they become a quote or an invoice.

AI that guesses instead of leaving a gap

Buyers worried about the AI misreading intent and asked to be able to switch AI interpretation off. A model asked to fill every field will sometimes fill one that is not on the page, such as a due date inferred from the email date. Empty is better than plausible.

Confidence scores that mean less than they look

Extraction services report confidence, but it is not one number. Microsoft's Document Intelligence returns a confidence between 0 and 1 for many results, notes that not all fields return one, and says the confidence of the text reading and of the key-value extraction are separate and both need checking. Amazon Textract reports confidence from 0 to 100 per block.

The PDF disappears

A standby generator dealer said the load evaluation from the manufacturer's portal goes away once it is sent. If the PDF is read and then discarded, the job loses its evidence. Buyers asked directly whether the original attachment stays on the job.

The options

Ways to do it

ApproachHow the PDF becomes a jobWeak spotGood for
Retype itA coordinator reads the PDF and types the jobSlow; typos in PO numbers and addresses; the PDF often stays in an inboxA few work orders a week
Scanning built into an FSM inbox (ServiceM8)With the Email Inbox add-on, a PDF work order that arrives is scanned, then staff click Convert to Job and the job card is filled with job and contact detailsServiceM8 says it works best with popular real estate management templates and results vary with layout; a person converts each oneTrade teams taking work from property managers
Document AI service plus your own code (Azure Document Intelligence, Amazon Textract)The service returns fields with confidence scores; your code maps them into the FSM through its APIYou build matching, dedupe and review yourself. Amazon's A2I human review is no longer open to new customers, and Google deprecated Document AI Human in the Loop on January 16, 2024Teams with developers and high volume
General AI model (Claude, ChatGPT) by handUpload the PDF, ask for the fields, paste the answer into your systemNothing is matched or saved; someone still copies it inTesting what extraction can do on your own PDFs
AI intake in a configurable FSM (Fieldproxy)An agent reads the PDF, matches the bill-to customer and site, checks the PO for an existing job, then creates the job or a draft for review, with the PDF attachedRequired fields and matching rules are agreed with you before go-liveCompanies with several clients, varied formats and 10+ field staff or subcontractors

As of November 2026, from each vendor's own documentation (sources below). Check the current version before relying on any one feature.

Worked example

A worked example: one week of PDF work orders

Say a facility maintenance coordinator gets 30 PDF work orders in a week from three facility management clients, covering 26 stores. Made-up numbers to show where each PDF ends up.

PDFsWhat the AI foundWhat happened
19All required fields; bill-to and store matched existing recordsJob created ready to dispatch, client PO on the job, PDF attached
4Known client, but a store address not on fileDraft job; reviewer approved the new site under the existing customer (not a new customer)
3A revised PDF for a PO already open (higher NTE, new date)Existing job updated, revised PDF added beside the first one
2Faint scan; NTE and due date flagged as uncertainDraft job; reviewer read the two values from the PDF and confirmed them
1A new facility management companyDraft customer and job; owner approved the customer after checking the billing address
1A quote request with 14 line items across two pagesDraft quote, not a job; estimator checked quantities before it went out

The team reviews 11 drafts, most of them one or two fields, instead of typing 30 jobs. No customer was created without a person approving it, and every job carries the PDF it came from.

Before you buy

What to check before you buy

Good signs

  • Reads scanned PDFs and phone photos, not only PDFs with a text layer
  • Shows the PDF and the extracted fields side by side for review
  • Separates bill-to customer from service site, and matches both before creating anything
  • New customers are proposed for approval, not created silently
  • Leaves a field empty and flags it when the value is not on the page
  • Uses the client's PO number to find an existing job before creating one
  • Keeps the original PDF on the job, plus a record of who confirmed what
  • Lets you turn AI extraction off for a client or a document type

Red flags

  • A demo on the vendor's sample PDF instead of yours
  • One template per client that you have to maintain
  • Every extracted field accepted with no confidence or review step
  • Line items go straight onto an invoice with no check
  • The PDF is processed and then not stored
  • A new customer record for every store or every sender

In Fieldproxy

Setting it up in Fieldproxy

Fieldproxy is set up around your process, so the fields, statuses and rules below are configured with you. Names marked e.g. are examples.

The Document intelligence agent reads the PDF

The Document intelligence agent (/agents/document-intelligence-agent) turns incoming PDFs, POs and work orders into jobs and line items. PDFs arrive by a connected mailbox (/solutions/email-management), by upload on the web, or in the Command Center (/agents/command-center), where you can drop a PDF and ask for the job.

Bill-to and site kept apart

Customers hold the billing company; sites sit under them with their own address and store number (e.g. site_code). The agent matches the bill-to first, then the site, and only proposes a new customer or site when nothing matches.

Draft first when it matters

Jobs with every required field and matched records are created in your ready status. Anything new or unclear lands in a review status (e.g. needs_review) with the PDF attached and the uncertain fields marked. When you work in the Command Center, every change comes as a confirm card with Undo, so a wrong extraction is reversed in one click.

Line items go through approval

When the PDF is a contract or a priced scope, the agent drafts the quote lines; Approval workflows (/solutions/approval-workflows) route it to a project manager or estimator before it is sent.

The original stays on the job

The PDF is stored on the job as received, with the client's PO number (e.g. client_po) as a searchable field for follow-ups and invoicing. The Watcher agent (/agents/watcher-agent) flags drafts that sit in review too long.

One price per company

Office staff, technicians and subcontractors all work in the same system. Fieldproxy is one price per company with unlimited users. See /pricing.

Records

How long to keep the original PDF

In the US, the IRS says to keep records that support an item on a tax return until the period of limitations runs out: generally 3 years, and 7 years if you claim a bad debt deduction. A work order PDF that supports an invoice falls under that, so keep it at least as long as the invoice it backs. Client contracts often require longer; check yours.

If you archive work orders outside your field service software, PDF/A (ISO 19005) is the standard family of PDF formats meant for long-term preservation. Keep the file as received, though, and do not re-save the client's original in a new format inside the job.

How switching works

How switching to Fieldproxy works

We move your data over and set the system up to the way you already work. Fieldproxy replaces the old tool; it does not run alongside it.

  1. Send us 20 real work orders

    A mix from each client, including the worst scans and the longest ones. We list the fields each format carries and which are often missing.

  2. We agree the fields and the matching rules

    Required fields, bill-to versus site, the PO number as the key, and what goes straight to a job versus a review draft.

  3. Run it on last month's PDFs

    We extract your past work orders and compare them to the jobs you typed by hand, field by field, before anything goes live.

  4. Go live with review on

    New PDFs create jobs or drafts. Once your team stops correcting a client's fields, those work orders can skip review.
Document intelligence agentWork order managementPricing: one price per company, unlimited users

FAQ

Questions people ask

Can AI create a customer and job from an uploaded PDF work order?
Yes. It should match an existing customer and site first and create the job under them. When the customer is new, the safer setup is to have the AI propose the customer for a person to approve, so you do not end up with near-duplicate records.
Do I need to set up my clients before the AI can create jobs?
No, but it helps. Importing your existing clients and sites first means most PDFs match on day one; anything new is proposed for approval.
Does the original PDF stay on the job?
It should, unchanged. The PDF is the scope, the NTE and the PO number you will invoice against, and it settles disputes about what was asked for.
Can AI read scanned or photographed work orders?
Yes. Models such as Claude read each page as an image as well as its text, and Microsoft's invoice model handles phone-captured images, scanned documents and digital PDFs. Faint scans and handwriting still produce more fields that need review.
What does AI get wrong most often on work orders?
Line items in tables that run across pages, handwritten notes, and fields that are not on the page but get guessed. Require a review step for line items and for any field the extraction marks as uncertain.
Can it turn a PDF contract into a quote instead of a job?
Yes. The lines become a draft quote, which should go to an estimator or project manager for approval before it is sent.
Can I turn the AI off for some clients?
You should be able to. Some clients send clean, consistent PDFs where extraction is reliable; others send handwritten forms where a person is faster. Ask any vendor to show the switch.
Does PDF extraction use AI credits?
On most AI-enabled platforms, yes, each document read uses some AI usage. Ask how a long PDF is counted against a short one and what happens when you run out. Fieldproxy's plans include AI credits; see /pricing.

Sources

Competitor details on this page come from the vendors' own pages and documentation, checked as of November 2026. Plans and prices change, so check the vendor's current page before you decide.

Related

Keep reading

See it built on your workflow

Tell us how your jobs run today. We build a demo on your stages, forms and pricing before the call.

Get DemoGet Demo