Change Order Field Note Minimum Fields Before AI Draft

Contractor AI Hub · 2026-10-19

A change order goes wrong long before anyone argues about markup. It goes wrong when a field note says "customer wants extra outlet" and an office model expands that fragment into a polished document with invented panel capacity, invented cable path, and invented schedule impact. The safer pattern is the same discipline used for dispatch intake: collect a minimum field set, mark what is still UNKNOWN, name who owns each gap, and only then ask an AI to draft language from facts that are already present.

This guide is office-process education for contractors and home-service teams that want AI help on change-order drafting. It is not legal advice, not a contract template, not a pricing schedule, and not a promise that any insurer, GC, or homeowner will accept the result. For a broader playbook on AI-assisted change-order writing on this site, see AI change order writing for contractors. Privacy expectations for notes that include customer identifiers still matter; the FTC's protecting personal information guide and start with security guide are useful baselines for what not to spray into prompts.

Why change-order drafts need a stricter field set than a job summary

A job summary can tolerate a fuzzy sentence about what the tech saw. A change order is a commercial instrument: it alters scope, price, and often schedule. If the model invents a quantity to sound complete, the company may later argue from a document nobody on site meant to authorize. Minimum fields exist to keep the draft tethered to what the field actually recorded.

Treat the AI as a formatter and checklist assistant, not as a witness. It can turn structured fields into readable paragraphs, suggest missing headings, and flag blank cells. It cannot stand in for a measured length, a photo of the existing condition, or a customer signature. When a field is empty, the honest output is UNKNOWN plus an owner—not a plausible filler sentence.

Separate three documents in your head even if they share a folder: the original scope or estimate pointer, the field change note, and the customer-facing change order draft. The AI should see the first two as inputs and produce the third as a draft marked for human review. Collapsing all three into one chat thread is how baseline scope quietly mutates.

If your office already uses a change-order form in the estimating system, map these minimum fields onto that form rather than inventing a parallel dialect. The point is honesty of inputs, not a new brand of paperwork.

Minimum fields before any model drafts language

Collect or explicitly mark UNKNOWN for each of the following before prompt time:

Job / estimate pointer — the ticket, estimate, or contract ID the change attaches to. Without this, the draft cannot say what is being changed.
Customer-facing location label — unit, floor, or area name the customer recognizes. Avoid dumping full account numbers into the prompt when a job ID already keys the record.
Change trigger — who requested the change and in what channel (on-site verbal, text, GC directive, uncovered condition). Mark the channel even if the quote text is still open.
Baseline vs. proposed scope — one sentence for what the signed or estimated scope already covered, and one sentence for what is newly requested. If either sentence is missing, stop.
Quantity and unit — count, length, area, or "lot" with the unit named. If not measured, write UNKNOWN — measure before price.
Pricing basis — T&M, unit price from the bid sheet, allowance draw, or "office to price." Do not let the model invent a dollar amount from a vibe.
Schedule impact flag — none / possible / definite, with a named owner for the schedule call.
Photo or attachment pointer — filenames or vault links for before-state evidence. If none, mark UNKNOWN and decide whether pricing may proceed without them.
Authorization path — who must approve before work proceeds (homeowner, GC, office manager). AI drafts do not create authorization.
PII minimization note — confirm the prompt excludes payment card data, government IDs, and unnecessary family details.

Illustrative HVAC example with explicitly stated assumptions; not a live job, not a price quote, and not a contract. Assumptions: residential change on an existing install ticket; tech recorded fields on a phone note; office will draft with AI; dollars are office-owned; schedule impact is a flag only.

Field	Reproducible example value
Job pointer	EST-10482 / ticket 7781
Location label	Main floor utility closet
Change trigger	Homeowner verbal on-site after seeing condensate route
Baseline scope	Install as estimated: condenser pad location A, line set path B
Proposed change	Add condensate pump and alternate drain route to laundry standpipe
Quantity / unit	1 pump assembly; drain line length UNKNOWN — measure before price
Pricing basis	Office to price from vendor sheet + labor units; no field dollar guess
Schedule impact	Possible — owner: dispatcher Dana
Attachments	closet-before-01.jpg; standpipe-02.jpg
Authorization path	Written homeowner approval required before parts order
Prompt PII	No card numbers; customer phone withheld from model prompt

Read the table as a gate, not as decoration. If quantity is UNKNOWN, the AI may draft descriptive scope language only if the draft header says pricing is incomplete. It must not emit a total. If authorization path is blank, the draft must not imply that work may start after the email is sent.

Prompt rules that keep the model from inventing scope

Give the model the field note as structured text or a table paste, then constrain the output:

Use only fields present in the note. If a field is UNKNOWN, print UNKNOWN in the draft rather than omitting the heading.
Do not invent lengths, breaker sizes, permit numbers, manufacturer models, or dollar totals.
Separate "description of requested change" from "price" and "schedule" sections so reviewers can accept narrative while rejecting numbers.
End with a human verification checklist: quantity confirmed, price basis confirmed, authorization path confirmed, attachments linked in the system of record.

Keep secrets out of the prompt. Payment data, government IDs, and unnecessary household details belong in the CRM—not in a model context window. Redaction habits from the same-site article on redacted HVAC notes for AI office handoff still apply when the change order note includes customer identifiers. Field-level routing habits from field-level routing for HVAC note summaries help when some lines must escalate to a licensed reviewer before any draft is sent.

Store the approved change order in the system of record after human edit. Do not treat the chat transcript as the contract file. The transcript is a drafting aid; the signed or countersigned document is the commercial object.

Who owns which gap

Minimum fields fail when everyone assumes someone else will fill them. Assign owners by default:

Field tech — change trigger, baseline vs. proposed sentences, photos, measured quantities when tools are on site.
Estimator / office — pricing basis, unit rates, tax handling, document ID binding to the original estimate.
Dispatcher / PM — schedule impact flag and crew implications.
Approver named in authorization path — customer or GC signature / written OK before procurement.

When a field stays UNKNOWN past the review window, either gather the measurement or explicitly downgrade the draft to a non-priced clarification request. Shipping a priced change order with silent unknowns is how disputes are born. For after-hours intake that must not dispatch on thin notes, the same-site callback note minimum fields article uses a parallel gate pattern before AI dispatch language.

If two owners disagree—for example the tech wrote a quantity and the estimator distrusts it—freeze the AI draft and resolve the human conflict first. Models are good at sounding decisive; they are not good at mediating a field-versus-office measurement dispute.

Worked walk-through of a refused draft

Suppose the only note is: "Customer wants pump, price it." That note fails the gate: no job pointer, no baseline scope sentence, no quantity for drain line, no pricing basis, no authorization path. A compliant office response is not a clever prompt. It is a return to the tech with the minimum field list and a hold on drafting.

After the tech returns the filled table from the illustrative example above, the model may draft a description of the condensate pump addition and list UNKNOWN for drain length pricing. The estimator then replaces the unknown with a measured length and a unit rate from the bid sheet. Only then does a total appear—and only under a human-owned pricing step.

If the homeowner has not approved in writing, the draft header should say "draft pending authorization" rather than "change order ready to schedule." Language that implies crew commitment before approval is an operations risk, not a writing style preference.

Assumptions and limits

All examples are hypothetical. Trade rules, contract clauses, lien rights, and permit requirements vary by jurisdiction and by your signed agreements; this page does not interpret them. AI tools differ in retention and training policies—review your vendor terms before pasting customer data. This guide does not claim AdSense approval, legal compliance, or measured time savings. It also does not replace your attorney-reviewed change-order form. Check the stated assumptions before adopting the field list.

Sources and method

FTC: protecting personal information — retrieved 2026-10-19
FTC: start with security — retrieved 2026-10-19

AI-assisted writing with automated source, calculation and editorial checks. No hands-on testing or professional certification is claimed. Check the stated assumptions before applying this guide.