AI Resume Bullet Points and the Numbers You Didn't Give It

Using AI to make a resume sentence cleaner is one thing. Letting it upgrade a number is another. Changing “about 40 calls a day” to “50+ calls a day” may look like ordinary polish, but it creates a claim you could be asked to defend.

Two employers that publish applicant-AI rules draw that line clearly. Principal allows help with formatting, grammar and proofreading, but says AI should not replace the substantive account of your experience. West Monroe permits AI for a resume, cover letter and application when the content reflects the applicant’s genuine experience. Its policy says inaccurate use can have consequences beyond the application itself.

Numbers are where that distinction is easiest to lose. This guide shows how to keep them honest, then explains the safeguards in our own resume builder. We make that builder, so the product-specific sections below are a maker’s account of the code as it stood on October 31, 2026.

What the two employers allow, in their own words

Principal’s page for candidates gives three examples of permitted use: “Resume formatting, grammar, and proofreading”, “Improving clarity and structure of written materials” and “Interview practice and general preparation”. “Unless explicitly stated otherwise”, AI is not permitted during interviews, live or recorded, or during assessments. Its guidelines end with a request that covers the whole process: “We ask candidates to represent themselves accurately and authentically at every stage of the process.”

West Monroe sets its policy out as a table, one row per stage. For the resume it says “Use for grammar, structure, clarity, or research”, and for interview preparation it allows practice with simulated questions. Interviews and skills assessments are marked “No”. One sentence below the table matters for anyone who has used a tool: “If an applicant is asked to disclose their use (or non-use) of AI in the application process, they must answer truthfully.”

Two policies are not a survey. Other employers may forbid AI outright or say nothing at all, and we found no count of how many US employers publish rules like these. If the employer you are applying to has its own page, its words decide. What the two have in common is still useful: neither objects to a tool making a sentence clearer, and both object to a sentence that says more than you did.

Why AI reaches for numbers

Career advice asks for results, and results are often counted. Harvard’s resume guide says resume language should be “Fact-based (quantify and qualify)” and lists “Not demonstrating results” among the top five resume mistakes. Bullet generators build their pitch on the same idea. Rezi’s free generator promises “measurable impact” and says “Just enter your job title and company”. Its FAQ recommends adding specific tasks or achievements but says they are not required, and the page does not say where a measurement comes from when a job title and a company are all the tool has. Resume Genius’s generator tells users to “Add numbers or context to make each point more impactful” and to treat what it writes as “starting points, not final drafts.”

We have not tested either tool, and this is not a claim about their outputs. The point is simpler: when a tool is asked for impact without being given a measurement, it can either leave a gap or invent an answer. Only the gap keeps the claim yours.

One bullet, three rewrites

Here is an example we made up for this page; the person and the employer are fictional. A customer support representative at a utility company has this bullet on her resume:

Answered billing questions from customers by phone and email, about 40 a day.

The same job has a second bullet: “Trained 3 new hires on the billing system.” Suppose a tool offers her these three rewrites. We wrote them to show common problems; they are not the output of any product.

1. Resolved about 50 billing questions a day by phone and email, raising satisfaction scores by 15%.
2. Handled about 40 billing questions a day by phone and email for 3 product lines.
3. I answered about 40 billing questions a day in Zendesk by phone and email.

The first raises her own figure from 40 to 50 and adds a 15% result she never mentioned. The second keeps her 40 but borrows the 3 from the training bullet, where it counted new hires, and attaches it to product lines. The third keeps the facts but opens with a pronoun and names a help desk product she never wrote down.

On October 31, 2026 we ran the three through the check our builder applies to every bullet rewrite before showing it. The first came back as “Resolved about [X] billing questions a day by phone and email, raising satisfaction scores by [X]%.” The second came back as “Handled about 40 billing questions a day by phone and email for [X] product lines.” The third was set aside, because the other two did not name a tool her text lacks. On its own, without the other two, it would have come back without the “I” and with the product name still in it. Arithmetic gets the same treatment: a version that turned her 40 a day into “about 1,200 billing questions a month” came back with “[X]” in place of the 1,200, because she never wrote that number.

What a safeguard can catch — and what it cannot

The writing help in our builder starts from a set of instructions to the model. It may use only the bullet, its job title and organization, and a detail the person adds. It must never add a number, tool, employer, client or outcome that those do not state, and it must not round or combine numbers. The instructions also cover meaning: “helped” never becomes “led”, and limiting words such as “about” and “most of the time” stay in. A linked job posting may steer which facts come first and is never treated as a fact about the person.

Instructions are not a guarantee, so the server checks each reply in code before anyone sees it. Every number that does not appear in the bullet, the added detail, the job title or the organization becomes “[X]”, with the currency sign or percent sign kept around it. “1,200” and “1200” count as the same number. A number from one of the entry’s other bullets is replaced too, because moving it would put a result where it did not happen. The instructions ask for at most two gaps in a bullet, and the code drops any version that comes back with more than three. A leading “I” or “We” is removed. A version that only repeats the bullet is dropped, and so is one that says nearly word for word what another bullet in the entry says, at least when that bullet holds no number. A version that brings in a name the input does not contain, such as a tool, a product or a company, is shown only when no other version is left.

Look again at the first rewrite in the example. “Resolved” and “raising satisfaction scores” both got through the check. A safeguard can compare numbers and names against the text you supplied; it cannot know that answering a question is not the same as resolving it. Read every verb as closely as every figure. Our post on checking an AI resume rewrite for accuracy gives that review a line-by-line method.

Giving the tool your number first

The cleanest fix is to hand over the figure before the rewrite. In the builder, the AI help beside a bullet has an “Add a detail” field of up to 300 characters, and a number typed there counts as yours. “Improve this bullet” then brings up to two rewrites. Nothing on the page changes until you pick one and press “Use this”, and Undo takes that step back.

Any gap you accept stays visible. A “[X]%” prints exactly as it stands, so the builder’s resume score treats a gap in square brackets as a blocker, and a blocker keeps the whole score at 49 or below until you fill the gap or remove the phrase. The placeholders left rule on our builder page gives the exact definition. If you do not know the real figure, an estimate whose inputs you can explain is one option; our guide to defending the numbers on your resume shows how to word one. Cutting the phrase is the other.

The builder’s “Suggest a bullet” works the other way round. It writes a new line for a job or project that has a title, and it is told to write no number at all. The code drops any suggestion with a number that is not part of that title or the organization’s name, and any suggestion with a tool or other name that the entry, your skills list or the linked job posting does not already have. In English, a suggestion that claims a result or a lead role, such as “improved”, “reduced” or “led”, is thrown out as well. It shows two suggestions at a time, and nothing is added until you add one. A suggestion describes ordinary work for the job title, so it belongs on your resume only if it is true of you. The writing help section of the builder page covers both tools and their limits.

If an employer asks about AI

West Monroe’s rule is straightforward: if you are asked, answer truthfully. That is far easier when every number on the page came from you. For example:

I used an AI tool to tighten the wording. The figures are mine: about 40 billing questions a day came from our team’s weekly ticket report.

An interviewer can ask about any line, and a figure you accepted from a draft without checking may be one you cannot explain. A broad question can lead there as well, and our page on walk me through your resume covers that one.