# The AI-Drafted Field Call, Each Answer Scored to Send, Rework, or Hand-Write

*Score every written field in one application on five dimensions, give each a Send, Rework, or Hand-Write call, and add the disclosure line wherever the rule requires one.*

- Canonical URL: https://www.refolk.ai/candidates/guides/ai-drafted-application-field-call
- Pillar: Positioning and materials
- Format: Framework
- Published: 2026-10-31
- Last reviewed: 2026-10-31
- Reading time: 21 min
- Keywords: can I use AI to write my cover letter for this job, company says no AI in application what counts, do I have to disclose AI on job application, certify application is my own work AI, how to tell if my AI cover letter sounds generic, employer AI policy for resumes and cover letters

## Key takeaways

- Published employer AI rules mostly police whose content an application is, not whether a tool touched it: the UK Civil Service, DWP, Anthropic and UKRI all allow refining and ban generation.
- Under a draft-it-yourself rule, an AI-drafted field cannot be reworked into compliance; the first draft has to be rewritten from your own notes, and only then may AI refine it.
- In a 600-person TopResume poll, 30.3% of hiring managers called AI use on resumes unacceptable against 25% for cover letters, so claim verifiability should weigh more on the resume.
- After an AI writing tool arrived on one hiring platform, the link between a tailored cover letter and callbacks fell by 51%, while more time spent editing AI drafts still tracked hiring success.
- Never score your own text with an AI detector: seven detectors flagged non-native English essays as AI 61% of the time, and one self-edit prompt cut detection of AI text to as low as 13%.
- In Refolk's index there are 97,201 US recruiter profiles against 7,112 in the UK, yet the detailed written-application AI rules come from UK public bodies, so most readers will be scoring under employer silence.

This guide is for anyone holding an AI-drafted resume, cover letter, and set of free-text answers for one specific posting. It shows you how to decide, field by field, whether each piece can go out as drafted, needs rewriting in your own words (with a disclosure line if the rules ask for one), or has to be written from scratch under this employer's rules. You get a five-dimension scoring model, the gates that override it, a procedure, and copy-paste templates for the score sheet and the disclosure line.

## Which of the six AI rule types does this employer use?

Employer AI rules for written applications come in six types: outright ban, draft it yourself then refine, polish or structure allowed with generic text banned, disclose significant use, attestation or acknowledgement, and silence. Tag the rule before you score anything, because the type sets the ceiling for every field on the form.

Five of the six types appear in named, dated public policies. These are the examples I score against:

| Employer | Rule type | Scope | Stated consequence |
|---|---|---|---|
| Anthropic, earlier form | Ban | Whole application | Not stated |
| Anthropic, replacement guidance | Own draft, then refine | Applications | Not stated |
| UK Civil Service | Polish or structure allowed; confirm accuracy | Applications | Application rejected |
| DWP | You must draft it yourself | Statements | Not stated |
| UKRI | Disclose anything beyond minimal use | Funding applications | Not stated |
| Amazon (reported) | Acknowledgement | Interviews and tests | Disqualification |

Source: the published policies of each body, as reported. Here is what each type means in practice:

- **Ban.** No AI anywhere in the stage the rule covers. Anthropic's earlier form said "please do not use AI assistants during the application process". The Scottish Government says never use AI tools to "answer questions or create statements on your behalf", which works as a ban on generated content.
- **Own first draft, then refine.** You write the first draft and AI may improve it. Anthropic's replacement guidance asks candidates to "create your first draft yourself, then use Claude to refine it". DWP says "You must draft your statement yourself".
- **Polish or structure allowed, generic text banned.** This is the UK Civil Service model. Its unacceptable uses include to "create generic responses and copy these into your application", and the Cabinet Office says the guidance applies "to all civil service roles across government".
- **Disclose significant use.** UKRI says applicants "do not need to disclose any minimal use", but they must not generate "an entire application, or sections of an application, without human involvement".
- **Attestation or acknowledgement.** You confirm, by checkbox or statement, that you followed a rule. The Civil Service asks candidates to confirm their information is accurate before submitting. Amazon's acknowledgement covers interviews and tests, not written applications.
- **Silence.** The ad, the careers pages, and the submit screen say nothing. No published example exists, because silence publishes nothing. You simply find it when you look.

How these types are spread across employers in general is not established publicly, so do not assume your employer resembles the table. Read its own pages.

Two patterns hold across every rule above. First, the rules police **whose content it is**, not whether a tool touched it. The Cabinet Office line is that applicants "should not use AI to generate full or false responses". Second, bans tend to soften. Anthropic replaced its ban within about five months, and the reversal was linked to "practical difficulties associated with enforcing a complete ban".

> **Note:** A remembered rule is a stale rule
>
> Rules change faster than application cycles. Quote the rule from the live page and record the date it shows; Anthropic's guidance carries a "last updated" stamp for exactly this reason.

## The five dimensions that score an AI-drafted field

Score each written field from 0 to 2 on five dimensions: the employer's rule for that field, any attestation that covers it, how far it relies on your voice and motivation, how verifiable its claims are, and how much of it is template language. A higher score means the field needs more of your own authorship before it goes out.

| Dimension | 0 | 1 | 2 |
|---|---|---|---|
| Rule (R) | Silence | Refine or polish allowed, with limits | Rule names this field type |
| Attestation (A) | No box | Accuracy confirmation | Own-work or AI-rule acknowledgement |
| Voice reliance (V) | Facts: dates, titles, tools | Evidence narrative | Motivation, "why us", values |
| Claim risk (C) | Every claim documented | True but inflated or rounded | Rests on memory, no document |
| Template share (T) | Most sentences carry specifics | About half generic | Most sentences fit any applicant |

### What each dimension proves, and how it lies

- **Rule (R).** This proves what the employer will hold you to. It lies when you read a rule written for one stage as if it covered another. The Fast Stream clause, "If we suspect that you have used Gen AI tools during your tests or assessments, your application may be withdrawn", says nothing about a personal statement.
- **Attestation (A).** This proves you will make a signed statement about the field, which turns a style question into an honesty question. It lies when you treat the box as routine. Current checkbox wording varies, and the exact text on live forms is not established publicly, so read yours.
- **Voice reliance (V).** This proves how much the reader is judging you rather than your record. A "why us" answer is almost pure voice. It lies when a fact-heavy field gets a high score because it reads smoothly. Smoothness is not voice.
- **Claim risk (C).** This proves whether the field would survive a reference check, an employment verification, or an accuracy confirmation. It lies when a number feels true because you have repeated it. Anthropic draws the line at "using Claude to invent experiences you haven't had", and an invented figure is an invented experience.
- **Template share (T).** This proves whether the text carries any signal. It lies when you judge by style. Semicolons and fluent prose are not evidence of anything. Count the sentences that name a project, a figure from your records, or a detail specific to this employer.

The dimensions are not equal. Some act as hard limits on the others, so check them in order, outermost first:

#### Order of precedence for a field call

1. **Employer rule for this stage** - A ban or a draft-it-yourself rule decides the call outright
2. **Attestation box** - Whatever you tick must be true of the field as sent
3. **Claim truth** - Every metric, title and date needs a document behind it
4. **Voice and template share** - Decide between Send and Rework once the outer layers pass

*An outer layer can force a call no matter how well the inner layers score, so read the rule before you judge the prose.*

### Why the resume carries extra claim weight

Hiring managers object to AI on resumes more often than on cover letters. In a TopResume poll of 600 hiring managers:

| Field | % saying AI use unacceptable |
|---|---|
| Skills assessment | 40.8 |
| Resume | 30.3 |
| Cover letter | 25.0 |
| Any stage | 14.5 |

That makes the resume 1.21 times as likely as the cover letter to be named off-limits. One plausible reason, which nobody has verified, is that resume lines are factual claims an AI can inflate. I act on that reasoning anyway because it costs little: on the resume, any claim score of 1 or more forces at least Rework. The same poll found that 52% of hiring managers call AI acceptable for proofreading or drafting support, and nearly 1 in 5 would reject an application that looks fully AI-generated. These are vendor polls, not probability samples, so treat them as direction rather than measurement. No survey gives a figure for motivation or short-answer fields.

## How a field score becomes Send, Rework, or Hand-Write

Run the gates first, then add the five scores: 0 to 3 is Send, 4 to 6 is Rework, and 7 to 10 is Hand-Write. **Send** means the draft goes out after a claim check. **Rework** means you rewrite it in your own words, keeping only content you can trace to your notes. **Hand-Write** means you write it from a blank page and use AI afterwards only if the rule allows refining.

### Gates that decide the call before scoring

1. **Ban or no-generation rule covering this stage.** The call is Hand-Write. No amount of editing turns an AI draft into one written without AI.
2. **Draft-it-yourself rule (Anthropic, DWP).** Hand-Write the first draft from your own notes, then refine with AI if the rule allows it. Your existing AI draft was not your first draft, so reworking it does not meet the letter of the rule.
3. **Attestation that excludes AI help for this stage.** The call is Hand-Write, and you tick the box only once the field is truly yours.
4. **Any claim with no record behind it.** Cut or correct it before scoring. This is not a call. It is a precondition for making any call.

### Overrides after scoring

- **Voice 2 and template 2 together.** Hand-Write, whatever the total. A generic motivation answer cannot be repaired by editing, because the missing material is your actual reason for applying.
- **Resume with claim risk of 1 or more.** At least Rework.

> **Rule:** The rebuild test decides every Rework
>
> A reworked field passes only if you can rebuild every sentence's claim from your own notes without looking at the draft. If a sentence's substance came from the model, it is not polish, it is substitution.

The two inner dimensions, voice and template share, decide most of the calls under permissive rules. This is how they combine:

#### Voice reliance against template share

Horizontal axis runs from Low template share to High template share. Vertical axis runs from Low voice reliance to High voice reliance.

| Quadrant | What it means |
| --- | --- |
| Voice-led and specific | Rework lightly and confirm every reason is genuinely yours |
| Voice-led and generic | Hand-Write from your real reasons for applying |
| Factual and specific | Send after checking each claim against records |
| Factual and generic | Rework by swapping stock lines for named projects and figures |

*The fields that need the most of you are the ones where the reader judges motivation and the draft says nothing specific.*

### A worked example

Take an application under a polish-allowed rule with an accuracy confirmation box, which is the Civil Service shape. Rule scores 1 and attestation scores 1 for every field.

| Field | R+A | V, C, T | Call |
|---|---|---|---|
| Resume | 2 | 0, 1, 0 = total 3 | Rework (resume claim override) |
| Cover letter | 2 | 1, 0, 1 = total 4 | Rework |
| "Why us" | 2 | 2, 0, 2 = total 6 | Hand-Write (V2 + T2 override) |
| Notice period answer | 2 | 0, 0, 0 = total 2 | Send |

The resume would have passed as Send on points, but one rounded-up headcount pushes it to Rework. The "why us" answer scores as Rework on points, but it is generic praise of the employer, so the override makes it a full rewrite.

### Where this model is thin

The bands and overrides are my judgement, not a validated instrument. The evidence supports their direction: rules target generated content, resumes draw more objection, and generic text loses signal. It does not support exact cut points. To check that a second reader would agree with you, have someone score the same field from the sheet without seeing your scores. If you differ by more than 2 points on any field, talk through the evidence lines until you agree.

## Scoring one application, field by field

Work through the application in this order: find the rule, classify it, list the fields, score them, check the claims, make the calls, disclose, and then test against the interview. Expect about an hour for a four-field application, with most of that time going on the claim check and the rework.

#### The field-call procedure

1. **Find and quote the rule** - Read the job ad, the careers page on AI, and the text beside the submit button (10 to 15 minutes). Done when the rule is quoted word for word with its date, or all three are confirmed silent.
2. **Classify the rule and its stage** - Tag one of the six types and the stage it covers: application, test, or interview (5 minutes). Done when the tag and stage word are written on the sheet.
3. **List every written field** - Write down the resume, cover letter, "why us" answer, and each short answer with its limit (5 minutes). Done when every prose box has its own row.
4. **Score each field on five dimensions** - Apply the gates, then score R, A, V, C and T from 0 to 2 with one line of evidence each (about 5 minutes per field).
5. **Check every claim against records** - Trace each metric, title and date to a document (15 to 30 minutes). Cut or correct anything without one before making any call.
6. **Make the call and do the rework** - Mark each field Send, Rework or Hand-Write, apply the overrides, and rewrite until a re-score lands in the Send band.
7. **Disclose and attest** - Add the disclosure line where the rule or form requires it, and tick boxes only once the field complies (5 minutes).
8. **Check against the interview** - Read each example aloud as you would tell it in an interview (10 minutes). Done when nothing written is something you could not say unprompted.

Step 5 is where most real damage gets prevented. If the resume was drafted by [Refolk](/candidates) from your own work history, the titles and dates already trace back to your records. Check the numbers anyway, because a draft can round up a figure faster than you notice. DWP's guidance gives the reason step 8 exists: interview answers must "align with your written application". A field you cannot defend aloud is a liability, however well it scored.

The score sheet below is plain text so you can paste it into a note or a spreadsheet:

**Per-field score sheet**

```
Employer and posting:
Rule (quoted):
Rule date shown on page:
Rule type: Ban / Own draft then refine / Polish allowed / Disclose / Attestation / Silence
Stage covered: Application / Tests / Interviews

Field:
Limit (words or characters):
Gates triggered: none / ban / own-draft / AI-excluding attestation / unrecorded claim
R (0-2):    evidence:
A (0-2):    evidence:
V (0-2):    evidence:
C (0-2):    evidence:
T (0-2):    evidence:
Total:
Overrides: none / V2+T2 / resume claim
Call: Send / Rework / Hand-Write
Disclosure needed: yes / no    line used:
Rebuild test passed (every claim from my notes): yes / no
Second reader total:
```

*Make one sheet per application. Copy the field block once for every prose box on the form.*

## When a disclosure line is required, and what it should say

Add a disclosure line when the employer's rule asks for disclosure of anything beyond minimal use, or when the form has a declaration field. Where the rule is silent, nothing I can point to requires one. Accurate claims and an honest answer if asked are the floor either way.

The public guidance gives two pointers on content. NHS Employers suggests forms ask applicants to "declare this, as well as what type of AI tools they have used and for what purpose". Australia's Fair Work Commission, a rule for legal filings rather than hiring, requires a party to "state in the document that GenAI was used and how". Both ask for the same three things: that a tool was used, what kind of tool, and what it did. There is no documented example of the exact wording any employer expects from candidates, so use the plainest version that answers those three.

UKRI's line tells you what counts as disclosable: minimal use need not be disclosed. A spelling and grammar pass is generally minimal. Restructuring or shortening text is a closer call. Generating sections of text is past the line.

**Disclosure line, full and short**

```
Full:
I wrote this [statement / cover letter / answer] myself. I used [type of tool, e.g. a generative AI writing assistant] to [check spelling and grammar / suggest a tighter structure / cut it to the word limit]. Every example, figure, title and date is my own and can be checked against my records.

Short:
Written by me; AI used only to [purpose]. All examples and figures are my own.
```

*Replace the bracketed parts with what you actually did; use the short version where the field has a character limit.*

On attestation boxes, the rule is simple: tick only when the field as sent meets what the box says. The Civil Service is explicit that "misuse of AI could result in your application being rejected", and Amazon's interview guidelines warn that "Failure to adhere to these guidelines may result in disqualification from the recruitment process". Whether anyone has had an offer withdrawn over AI use in a written application is not established publicly. That does not make the box safe to tick on autopilot. It means the consequence is unknown, not absent.

## How the call goes wrong: failure modes and false positives

The most common failures are reading the wrong rule, scoring with the wrong instrument, and letting permitted polish turn into substitution. Each one has a specific check, and running these checks matters more than adjusting the scores.

1. **Stale rule.** You score against a rule you remember. Anthropic went from a ban to draft-then-refine in about five months. *Check:* quote the live page and record its date.
2. **Scope confusion.** A test or interview clause gets read as covering the written application, which produces a false Hand-Write call. Fast Stream's withdrawal clause covers tests and assessments, and Amazon's acknowledgement covers interviews and tests. *Check:* find the stage word in the rule.
3. **Self-checking with a detector.** In Liang et al.'s study, seven detectors flagged 91 essays by non-native English speakers as AI-generated 61% of the time, and a single self-edit prompt cut detection of AI text from up to 100% to up to 13%. Worse, polishing those essays with ChatGPT significantly reduced misclassification, so dodging a flag pushes you toward more AI, not less. *Check:* never use a detector score as a dimension.
4. **Misreading surveys.** A widely cited Resume.io figure of 49% blurs recruiters who "suspect" AI with those who can "identify" it, and these are vendor polls. The 33.5% "spot it in 20 seconds" figure is self-reported, and no controlled study of how accurately recruiters spot AI-written resumes or cover letters was found. *Check:* read the question wording and the sample size before letting a number change a call.
5. **Polish turning into substitution.** "Refining" gradually becomes AI-written statements, which the Scottish Government forbids outright. *Check:* the rebuild test. Can you reconstruct every sentence's claim from your own notes?
6. **Style-tell false positives.** You rewrite a perfectly good field because it uses semicolons or reads fluently. NHS Employers lists "the over-use of semi colons" as a sign of AI text, which will also catch careful human writers. *Check:* look for content tells instead, such as no named project and no detail specific to this employer.
7. **Unverifiable claims.** A fabricated or inflated metric breaks the Civil Service accuracy confirmation and fails any later employment check. *Check:* every number has a document behind it.
8. **Sending unedited.** Cui, Dias and Ye found many AI drafts go out with minimal revision, yet more time spent editing AI drafts was associated with higher hiring success. *Check:* no field goes out without a re-score after editing.

> **Watch out:** A clean detector score proves nothing
>
> Detectors mostly measure how predictable the prose is, so they flag plain writing and miss lightly edited AI text. Score the content and the rule instead.

Failure 8 matters more than it looks. In the same Cui, Dias and Ye study, the correlation between a cover letter's textual alignment with the posting and callbacks fell by 51% after an AI writing tool arrived on the platform. Once tailoring became cheap, it stopped separating applicants. What still separates them is material a model could not have produced without you.

> A tailored letter is now the minimum. The signal left is what only you could have written.

## Who reads the application, and why silence is the usual rule

Most readers of this guide will apply under employer silence, so voice, claims, and template share will decide their calls. The detailed published rules come from UK public bodies, while most hiring volume sits in the US, where no private-employer rules for written applications were found.

| Segment | Profiles | Share "Recruiter" in top-25 sample |
|---|---|---|
| US Recruiter or TA | 97,201 | 52% (13/25) |
| UK Recruiter or TA | 7,112 | 56% (14/25) |
| US+UK resume writer or coach | 6,363 | n/a |
| US Recruiter or TA, "AI recruiting" headline | 45 | 76% (19/25) |

Source: Refolk's index of professional profiles. Title matching is noisy, and the samples include titles like "Founder & CEO".

**13.7x - US recruiter and talent acquisition profiles for every UK one in Refolk's index**

The written-application AI rules quoted in this guide come almost entirely from UK public bodies, so the larger US reader pool mostly sees silence.

Silence is not permission to send anything. It sets the rule score to 0 and leaves the other four dimensions in charge. A generic "why us" answer still scores Hand-Write under silence, because the override does not depend on the rule.

Do not try to read an individual reader's stance from their profile, either. In Refolk's index only 45 US recruiter headlines match "AI recruiting", about 0.05% of the total. That tells you almost nothing about how the other 97,000 judge AI-drafted text. The employer's written rule is the only stance you can quote.

**16.4 - Recruiters per resume writer or career coach across the US and UK in Refolk's index**

There are 97,201 plus 7,112 recruiter profiles against 6,363 writer and coach profiles, so most applications are judged by people who never see the drafting process.

Public-sector pressure is one reason UK rules are so detailed. Applications to competitive senior Civil Service roles rose 35% in a single year, a rise partly put down to AI-generated applications. Where an employer faces that kind of volume, expect it to write rules. Where it does not, expect silence, and score accordingly.

## Final checks before submit, and keeping the call current

Before you submit, confirm the rule, the claims, the boxes, and your interview readiness one last time. Then re-run the sheet for every new posting, because the rule, the fields, and the draft all change from one application to the next.

#### Before you press submit

- [ ] The employer's AI rule is quoted on the sheet with the date shown on the live page, or silence is recorded for the ad, the careers page, and the submit screen.
- [ ] The rule's stage word has been read, and test or interview clauses have not been applied to written fields.
- [ ] Every prose box on the form has its own scored row with one evidence line per dimension.
- [ ] Every metric, title, and date in every field traces to a named document.
- [ ] Every Hand-Write field was written from your own notes, not edited down from the draft.
- [ ] Every Rework field passes the rebuild test sentence by sentence.
- [ ] The "why us" answer names at least one detail specific to this employer and one real reason of your own.
- [ ] A disclosure line is present wherever the rule or form asks for one, naming the tool type and its purpose.
- [ ] Every ticked attestation box is true of the field exactly as it will be sent.
- [ ] You can tell each written example aloud with the same facts as on the page.

### Keeping the call current

- **Re-check the rule for every application, even to the same employer.** Anthropic's rule changed within months, and in one third-party check Amazon's rule could not be found on Amazon's own hiring pages. Pages move and get rewritten.
- **Re-score when the draft changes.** A field that scored Send for one posting can score Rework for the next once it has been retailored. If you use Refolk to tailor the resume and cover letter to each posting, treat each tailored version as a new draft and run the sheet on it, not on the master copy.
- **Keep the notes the rebuild test depends on.** A short running file of projects, figures, and the documents behind them makes Hand-Write fields fast and Rework fields provable.
- **Watch for new rule types in your sector.** If you start seeing a disclosure field or an attestation box you have not met before, add its wording to your sheet, classify it against the six types, and score from there.

A field call takes a few minutes once the sheet is set up. Getting one wrong on an attested field can sink an application that was otherwise strong.

## Frequently asked questions

### Can I use AI to write my cover letter for this job?

It depends on the rule this employer publishes for application materials. Under silence or a polish-allowed rule you can use an AI draft, as long as every claim traces to your own records and the letter carries specifics only you could write. Under a draft-it-yourself rule, like Anthropic's or DWP's, you write the first draft and can only use AI to refine it. Under a ban, or a rule against AI creating statements on your behalf, you write the whole thing yourself.

### The company says no AI in the application. What counts?

Read the stage word first. Some rules cover only tests and interviews: the Fast Stream withdrawal clause covers tests and assessments, and Amazon's acknowledgement covers interviews and tests. If the ban covers the application itself, treat every prose field as Hand-Write. Where the wording is ambiguous, the published rules mostly target content generated for you, so content written by you and lightly checked for spelling is the conservative reading. If it is still unclear, ask the recruiter in writing.

### Do I have to disclose AI use on a job application?

Only where the employer's rule or form asks you to. UKRI, for example, says minimal use need not be disclosed but significant help should be, and NHS Employers suggests forms ask what type of tool you used and for what purpose. No documented example shows the exact wording employers expect, so name the tool type, say what it did, and confirm that the examples and figures are your own.

### Can I tick a box certifying the application is my own work if AI helped?

Only if your field now complies with what the box actually says. A box that confirms your information is accurate requires every claim to be true and documented. A box that certifies the work is your own, or that you followed the employer's AI guidelines, requires you to meet that rule for the stage it covers. The Civil Service says misuse of AI can get an application rejected, so read the box wording before treating it as routine.

### How can I tell if my AI cover letter sounds generic?

Check content, not style. Go through it sentence by sentence and mark each one that names a project, a figure from your records, or a detail specific to this employer. If most sentences could be pasted into another applicant's letter unchanged, it is generic. Do not rely on semicolons, fluency, or an AI detector score: those tells misfire, and detectors flag plain non-native writing far more often than polished AI text.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/ai-drafted-application-field-call*