Guided fill and auto-fill
What this page is — the four aids that fill a document faster than typing each field: a step-by-step interview, AI auto-fill from a pasted brief, AI structuring of captured files, and AI-suggested titles and tags.
What it is for — so a thirty-field form is either walked through one question at a time, or filled from the email the request arrived in, with a person accepting each value.
The problem it solves — long forms are retyped from emails field by field, slowly and with errors.
Route: /org/papers/documents/<document> · Permissions: Edit draft documents / data values.;
Use AI drafting/extraction/Q&A for documents. for the three AI aids.
1. What it is
| Aid | Needs | Fills | Costs |
|---|---|---|---|
| Guided fill | Nothing | The same fields, one group at a time | Free |
| Auto-fill from source | Pasted text | Empty fields, as suggestions | 300 credits est. · 1,200 max |
| Structure captured content | A document converted from the Capture Inbox | Empty fields, as suggestions, plus a type-fit note | 300 · 1,200 |
| Suggest title & details | Nothing | Title, tags, category, one-line summary, as suggestions | 450 · 1,800 |
The three AI aids share one rule that makes them safe to use: nothing lands until you accept it.
Each suggestion appears as a chip under its field — value, a HIGH / MEDIUM / LOW confidence tag,
and the snippet it was taken from — with Accept and ✕.
They also share a second rule: they only fill empty fields. A value you typed is never overwritten.
2. Why you would use them
A well-designed type asks a lot. Faced with thirty fields at once, people skip the optional ones and guess at the rest.
- Guided fill makes the required set obvious. A yellow note lists the required answers still missing, and conditional questions resolve as you answer — so someone unfamiliar with the type fills it correctly the first time.
- Auto-fill removes re-typing. The deal was agreed in an email. Pasting that email and accepting eight chips is a minute; typing eight fields from it is ten, with transcription errors.
- The snippet makes each suggestion checkable. You are not trusting a value; you are confirming the sentence it came from.
- Structuring rescues captured paper. A scanned counterparty contract becomes a document with real field values, and the type-fit note tells you if you converted it to the wrong type.
- Consistent titles make the list searchable. "Master Services Agreement — Gatiro / Acme (2026)" beats "scan_0043".
3. Step by step
Guided fill
- Guided fill in the lifecycle bar — shown only while the document is editable and the type has fields.
- Answer each step. Step pills across the top let you jump around.
- Finish on the last step's green check. Answers save to the same form.
Auto-fill from source
- Auto-fill from source above the fields.
- Paste the brief, email or notes — or dictate them with the mic.
- Suggest field values.
- Accept or dismiss each chip.
Structure captured content — appears above the fields only on a document converted from a captured file. One click; accept or dismiss the chips; read the Type fit note first.
Suggest title & details — in the header. Review each suggestion; Use this title applies it.
4. What you provide
| Aid | Input | Required | Good input |
|---|---|---|---|
| Guided fill | Your answers | — | — |
| Auto-fill from source | Source text | Yes | Specific: numbers, dates, names, law |
| Structure captured content | — | — | The captured file's text is read automatically |
| Suggest title & details | — | — | The document's rendered body is read automatically |
A thin brief yields few suggestions. That is by design — the AI only proposes what it can ground in the source — not a miss.
5. Field reference — what comes back
Auto-fill and Structure (one suggestion per empty field it can fill)
| Field | Means |
|---|---|
field_name / label | Which field |
datatype | The field's type |
suggested_value | Typed correctly — ISO dates, plain numbers, a valid option for a select |
confidence | high, medium or low |
rationale | The source snippet it came from |
notes | A coverage note, or what to double-check |
type_fit | Structure only — whether the captured document looks like the chosen type, and what it resembles if not |
Suggest title & details
| Field | Means |
|---|---|
title | A concise descriptive title |
tags | 3–8 lowercase tags |
category | One of agreement, transactional, correspondence, certificate, policy, record, custom |
summary | One sentence |
It reads the current title only as a weak hint, and never invents facts not in the text.
6. Worked example
A sales lead has a three-year MSA to raise from this email:
"3-year MSA with Acme Corp, fees ₹50,000/mo, Net-30, starts 1 Aug 2026, governed by India law."
They open Auto-fill from source, paste it, and press Suggest field values. Four chips appear:
| Field | Suggested | Confidence |
|---|---|---|
monthly_fee | 50000 | high |
payment_terms | Net-30 | high |
start_date | 2026-08-01 | high |
governing_law | India | high |
counterparty_name already read "Acme Corporation Pvt Ltd" — typed earlier — so no chip was offered for
it, even though the email says "Acme Corp". That is the empty-fields rule protecting a better value.
They accept all four. Term (months) stays empty: the email says "3-year", but the field is a number and
the model did not convert it with confidence. They type 36.
Finally Suggest title & details proposes "Master Services Agreement — Acme Corp (2026)", tags
msa, acme, india, services, category agreement. They click Use this title. 750 credits in total;
under two minutes.
7. The admin contract
| Must be true | Where | What breaks without it |
|---|---|---|
| The role holds Edit draft documents / data values. | Role editor | No Guided fill, and chips cannot be accepted |
| The role holds Use AI drafting/extraction/Q&A for documents. | Role editor | No AI aids |
papers_field_autofill, papers_ingest_structure, papers_auto_metadata are active and enabled | Orbit AI Flow, /org/ai-utilities | "This utility is currently disabled", or the button is absent |
papers.ai_monthly_credit_cap not reached | System Config | monthly AI credit cap reached… |
| The type's fields have clear labels and datatypes | Field schema | Suggestions land on the wrong field or are not offered |
| Select fields list their options | Field schema | A select cannot receive a suggestion that is not one of its options |
On a document marked confidential, AI runs still work when pressed, and where a local model key is
configured they are forced on-prem and audited as ai_local_routed. Nothing about these aids runs
automatically.
8. Downstream
| After accepting a suggestion | Then |
|---|---|
| The field holds the value | Autosave, preview re-render, and an audited edit — attributed to you, not the AI |
| A conditional driver was filled | Dependent fields appear |
| A title was applied | The list and search use it immediately |
9. Don't confuse this with…
| Smart fields | Extracts configured values out of a finished document. These put values in |
| Drafting AI | Writes the document's wording. These fill its fields |
| Merge sources | Pre-fill automatically at creation, with no AI and no acceptance step |
| Classify | Picks the type and folder from your catalog. Suggest title & details only names and tags |
10. Troubleshooting
| Symptom | Cause |
|---|---|
| Auto-fill offered nothing | The source contains nothing matching an empty field — filled fields are never suggested |
| Few suggestions from a long email | It only proposes what it can ground. Vague prose grounds little |
| A suggestion went to the wrong field | Field labels are ambiguous. Dismiss it; improve the labels in the schema |
| No Structure captured content button | The document was not converted from the Capture Inbox |
| Type fit says it resembles another type | Reconvert to the right type before structuring |
| No Guided fill button | The document is not editable, or the type has no fields |
| A select would not take a suggestion | The value is not one of the field's options |