How to Catch Made-Up Figures in AI-Drafted Proposals

Assume every number is unverified until you can point to where it came from. Pull every figure out of the draft into a list, tag each with its source (your own records, the client, a named public report, or a calculation), then check each against that source. Any figure without a source gets cut or replaced with a real one.

The figures to worry about are the believable ones. An AI tool rarely writes that you have placed a million candidates; it writes "an average time-to-fill of 18 days" or "94% retention", numbers precise enough to sound measured and modest enough to pass a quick read. They are generated because figures like them appear in thousands of proposals, not because anyone measured them in your business.

Follow me on Instagram@sagnikteaches

The kinds of figures AI makes up in proposals

A model producing a proposal is predicting what a convincing proposal contains, and convincing proposals contain numbers. The invented ones fall into a handful of types, and knowing them tells you where to look first. The mechanism behind all of them is explained in AI hallucinations explained for business owners; here is how it shows up in proposals specifically.

Connect on LinkedInSagnik Bhattacharya

• Invented internal metrics. Your retention rate, response time, client count or success rate, written as if taken from your systems. The most dangerous type, because the reader assumes you measured it.
• Zombie statistics. Industry figures that circulate online, quoted page to page, with no traceable original. "A bad hire costs up to 30% of first-year salary" is a classic: widely repeated, usually credited to a government labour department, and very hard to trace to any actual publication.
• Stale figures. True once, out of date now: last year's fee, a 2019 survey result, a headcount from before the restructure.
• Wrong population. A real figure about large corporations applied to a twelve-person firm, or a figure about one sector used for another.
• Unit and period slips. Monthly quoted as annual, 6-month retention described as 12-month, a percentage of revenue presented as a percentage of profit.
• Calculated claims. "This could save you $150,000", built by multiplying a questionable statistic by a guessed quantity.
• Precision theatre. 37.4%, 4.2 days. Decimals signal measurement; AI adds them because measured figures often have them.

Give the AI a numbers file before it drafts

The cheapest audit is the one you barely need. Before any proposal is drafted, give the AI a short file of figures it is allowed to use, each with a date and source, and tell it to leave a visible gap rather than invent anything else.

Subscribe on YouTube@codingliquids

A filled-in numbers file for an illustrative twelve-person recruitment agency:

The instruction that goes with it:

Use only figures from the attached numbers file. If a
sentence needs a figure that isn't in the file, write
[FIGURE NEEDED: what it would measure] instead of a number.
Do not use industry statistics unless I paste them in with
a source. Do not add decimals to any figure.

The draft it produces then reads like this (illustrative):

Over the past year we placed 214 people, and for permanent
office roles our average from brief to accepted offer was 26
days. For warehouse roles, we typically present a shortlist
within [FIGURE NEEDED: days to shortlist, warehouse roles].

That gap is the system working. The agency doesn't track days to shortlist for warehouse roles, so someone either pulls the figure from the system or rewrites the sentence without one. The alternative, a confident "within 5 days", is what the model would have written without the instruction.

Find every figure in the draft, including the disguised ones

Even with a numbers file, run a sweep, because drafts get edited, pasted into and "improved" after the first version. Don't rely on reading: eyes slide over numbers embedded in sentences. Make the software find them.

• In Word: open Find and Replace (Home, then Replace), select More, tick Use wildcards, search for [0-9], then choose Reading Highlight and Highlight All. Every digit in the document lights up.
• In Google Docs: open Edit, then Find and replace, type \d in the Find box and tick Match using regular expressions. Step through each match.

Then search separately for figures written as words and for figure-shaped claims, which a digit search misses: half, double, twice, a third, dozens, hundreds, most, majority, nearly all, every, leading, fastest, first. "Most of our clients stay with us for years" is a numerical claim in disguise, and it needs the same source as "72% renew".

List each hit in an audit table with four columns: the figure, the sentence it sits in, the claimed source, and a verdict. The table is the working document for the next three sections, and it is worth keeping with the proposal afterwards: if the client asks where 26 days came from, the answer takes one sentence.

Trace each figure to where it really came from

Every figure belongs to one of four source types, and each has its own test:

For public figures, the full technique for tracing a claim back through the pages that repeat it is in how to check sources and citations in AI research. The short version: if after ten minutes you can't find the original, the figure goes. A proposal loses nothing by dropping an industry statistic and loses a lot by including a false one.

A recruitment agency's 23-figure audit

Here is the whole routine on one illustrative proposal: the agency above pitching to a manufacturer that plans to hire fifteen warehouse and office staff over six months. An account manager drafted the nine-page proposal in a chat assistant from call notes, before the agency had a numbers file. This paragraph from the "Why us" section is typical of what came back:

With a network of over 45,000 candidates and an average time-to-fill of just 18 days, we place people faster than the industry average of 42 days. 94% of our placements are still in role after 12 months, and with a bad hire costing up to 30% of first-year salary, our approach could save you over $150,000 across 15 hires.

The sweep found 23 figures across the proposal. The audit took about 50 minutes, and the verdicts came out like this:

Almost every figure in the sample paragraph failed. "Over 45,000 candidates" was the all-time database, including people who registered in 2011; 3,100 have been active in the last 90 days. "18 days" had no source at all; the real average for office roles is 26. The agency measures retention at six months, not twelve, and the figure is 88%. The 30% bad-hire cost is the zombie statistic from earlier. The $150,000 saving multiplied that untraceable 30% by an average salary and fifteen hires, which quietly assumed every hire would fail without the agency. After the audit the paragraph read:

We have 3,100 candidates who have been active with us in the last 90 days, and in the past year we placed 214 people. For permanent office roles, our average from brief to accepted offer was 26 days. Of the 190 people we placed between September 2025 and March 2026, 88% were still in the role six months later.

It is less dramatic and far stronger. Every sentence survives the client's procurement team asking "how do you know?"

When the AI names a source that doesn't contain the number

The natural reaction to a doubtful figure is to ask the AI where it came from. That can help, and it can also produce a second invention. An illustrative exchange:

You: Where does "a bad hire costs up to 30% of first-year
salary" come from? Give the original source.

AI: This figure is widely attributed to a government labour
department estimate and is frequently cited in HR research,
for example in reports on the cost of employee turnover.

What you would do with that answer: nothing yet. It names a type of organisation, not a publication; "frequently cited" describes the zombie problem rather than solving it. When the account manager searched the named body's own website, no such publication turned up, only recruitment and HR pages quoting one another. Even when an assistant with web search returns a real link, open it and search the page for the number. Links to real pages that don't contain the quoted figure are common, and a figure that exists on the page can still be from a different year or about a different kind of employer. A named source isn't a source until you have read the number on it. Law firms face the extreme version of this with invented case citations; the checking routine for AI-invented case law applies the same rule under higher stakes.

Recalculating the savings and totals

Calculated figures fail in two ways: bad inputs and bad arithmetic. The agency's proposal had both. Beyond the $150,000 saving, its fee summary said: "For 15 hires at an average salary of $34,000, our total fee would be $91,800." Rebuilt in a spreadsheet:

• 15 × $34,000 = $510,000 in first-year salaries.
• At the standard 18%: $91,800. The arithmetic was right.
• But the terms of business give a 15% volume rate for five or more roles: $510,000 × 15% = $76,500.

The AI had used the right maths on the wrong rate, overstating the fee by $15,300. That error doesn't embarrass anyone; it simply loses the deal to a cheaper competitor, and nobody ever finds out why. Rebuild every calculated figure from verified inputs in a spreadsheet rather than asking the AI to check its own sums, for the reasons given in why AI is bad at maths.

Other sectors show the same pattern. An illustrative insurance broker's proposal claimed "clients who move their cover to us save an average of 22% at renewal". The broker had never calculated it. The honest version needed a real calculation from the previous year's moved policies, stated with its basis ("across the 40 business policies we moved last year, the median premium change was..."), or no figure at all.

Numbers that disagree across one proposal

Long AI drafts are often assembled in pieces, section by section, and the pieces don't always agree. A consistency pass catches figures that are individually plausible but contradict each other. A prompt for it:

Below is a proposal. List every number or commitment that
appears more than once or describes the same thing (fees,
timelines, headcounts, dates, response times). For each, show
every occurrence with its section heading. Flag any that
don't match. Do not fix anything.

[paste proposal]

Run on the agency's draft, it returned (illustrative):

FEE: "18% of first-year salary" (Commercials);
     "15% volume rate" (Our terms) - MISMATCH
SHORTLIST: "shortlist within 7 days" (Executive summary);
     "shortlists in week 2" (Timeline) - MISMATCH
HIRES: "15 roles" (Summary); "15 roles" (Timeline) - match

Checked against the proposal, both mismatches were real, and the tool found them in seconds. But it missed a third: the covering email said the agency had "over 200 placements a year" while the proposal body said 214 in the last twelve months, which agree, and "nearly 250" on the final page, which doesn't. The consistency prompt is a fast first pass; the digit sweep from earlier is what guarantees completeness.

Writing around a figure you can't back up

Cutting an invented number often leaves a hole in the argument, and the temptation is to fill it with a softer invention. There are three honest replacements, and each is usually more persuasive than the statistic it replaces.

The last row comes from an illustrative law firm pitch, where the AI had written the 300 figure and a success rate alongside it. The matter-management system could produce a real count of matters by type, but the partner chose examples instead, and dropped the success rate entirely after checking what the firm's professional rules say about advertising results. That is a good general reflex: before quoting any win rate, approval rate or success percentage, check whether your regulator or professional body restricts how results are advertised, because an invented figure there is a conduct issue as well as an accuracy one.

One more replacement is always available: nothing. Many strong proposal paragraphs contain no figures at all, just a clear account of what you will do, who will do it and by when. If the only figure you can find for a sentence is one you can't source, the sentence probably didn't need a number.

The number-only read before sending

The last step is a second person reading only the numbers. Not the whole proposal: the highlighted figures, each ticked against the audit table. For a nine-page proposal it takes about ten minutes, and it works because the reader isn't drawn along by the argument. They see "26 days" and ask one question: which report?

Give them this short list:

1. Every highlighted figure appears in the audit table with a verdict.
2. No figure is marked "untraceable" or "invented" and still present.
3. Every public statistic has a year next to it in the text.
4. Every calculated figure matches the spreadsheet.
5. Rounded figures round in the cautious direction ("over 200", never "nearly 250" for 214).
6. No decimals that weren't measured.

If the proposal passes, send it. If the same kind of invented figure keeps turning up week after week, the fix belongs upstream: add the missing figure to the numbers file, or add a line to the drafting instructions. Where proposals go through a formal sign-off, this read slots in as one line of the approver's checklist, the same way the approval step for AI-written quotes handles prices. For agencies using AI across business development, not only proposals, the tutorial on AI for recruitment business development covers the wider workflow these figures feed into.

Checking figures in AI proposals: more questions

Does using AI with web search stop it inventing figures?

It reduces invention but doesn't remove it. A search-enabled assistant can still pair a real link with a number the page doesn't contain, quote an old figure as current, or apply a statistic about large companies to a ten-person firm. Treat every cited figure as a lead: open the page, find the number, and check its date and who it describes.

Is it acceptable to round figures in a proposal?

Yes, if the rounding is honest and in the cautious direction. 214 placements can become "over 200"; it cannot become "nearly 250". Keep the exact figure and its source in your audit table so you can answer a client who asks. Avoid false precision too: a figure like 37.4% implies a measurement, so only use decimals you actually measured.

How long should a number audit take?

For a typical eight-to-ten-page proposal with twenty or so figures, allow 40 to 60 minutes the first time, most of it tracing public statistics. It gets faster once you keep a numbers file of verified internal figures, because those checks become a glance. The final number-only read by a second person takes about ten minutes.

Should I tell a client a proposal was drafted with AI?

There is usually no rule requiring it for a sales document, but check your sector's professional rules and any client procurement terms, which increasingly ask. What matters more is that every figure is true and sourced. If a client asks where a number came from, your audit table lets you answer in a sentence, whoever drafted it.

Further reads

• A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A five-minute routine for everyday AI output, not just proposals.
• How to Back Up Every Claim AI Writes in Your Marketing — The same discipline applied to marketing claims.
• AI Proposal Mistakes That Cost You the Job — The non-numerical proposal errors that also lose work.
• How to Write Business Proposals Faster With AI — A proposal drafting workflow that builds in these checks.
• How Accurate Is ChatGPT? What Owners Should Expect by Task — What accuracy to expect from ChatGPT by type of task.
• AI Error Log: Track Mistakes and Stop Them Happening Again — Record each invented figure so the prompt stops producing it.
• How to Set Up Human Review for AI Work Without Slowing Down — Four levels of human review matched to risk, how to make each check take under a minute, how many to sample, and when to relax or tighten.
• How to Write a Business Plan With AI, and What to Check Yourself — Draft each section of a business plan with AI from your own facts, build the financials yourself, and check the claims lenders and partners will test.
• How to Mention AI Use in Client Contracts and Proposals — Sample proposal wording and contract clauses for disclosing AI use to clients, matched to how much AI touches the work, with an architect-practice example.
• The Limits of AI: What It Still Gets Wrong in a Small Business — Eight things AI still gets wrong in a small firm, where each one bites, the warning signs, and the specific check that catches it.
• AI Itineraries: What a Travel Agent Must Check Before Sending — A seven-part checklist for AI-drafted itineraries, with why each check matters, how to verify it, the red flags of an unchecked draft, and a sign-off record.
• How Event Planners Use AI to Build Budgets and Run Sheets — Turn an event brief into a line-item budget and a timed run sheet with AI, while supplier quotes and spreadsheet formulas keep the numbers honest.
• How to Write a Salon Service Menu and Price List With AI — Price each service from your chair-hour cost, then use three AI prompts to structure, describe and stress-test the menu before it goes live.
• AI Engagement Letters for Accountants: Faster Drafts, Clear Scope — Which parts of an engagement letter AI should draft, five prompts with sample outputs, and the partner checks that keep scope clear and terms untouched.
• How Consultants Use AI to Write Proposals in Under an Hour — A proposal kit you build once and a timed five-block hour: brief, draft from the kit, price it yourself, let AI object as the buyer, then check.
• How Architects Use AI for Fee Proposals, Briefs and Paperwork — How a small practice uses AI for briefs, fee proposals, variation letters and site reports, with prompts, sample outputs and the parts AI must leave alone.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support (Find and Replace, wildcards, Reading Highlight); Google Docs Editors Help (find and replace with regular expressions). Checked September 2026.