Will an AI agent screen your CV at a small business?

What did Meta actually launch, and does Muse screen CVs?

On 29 September 2026, Meta launched Muse for Small Business — an AI agent that plugs into the software owners already run. It connects to Shopify and Stripe for commerce, QuickBooks for accounting, Klaviyo for marketing, Canva and Figma for design, and Slack, Asana and Notion for day-to-day work, alongside a business's Instagram professional account, Facebook Page and ad accounts. It is not a recruiting product. Meta pitched it at growth and admin: analyse sales and campaigns, draft a plan, flag emails that need a reply, spot unusual expenses. It's free with usage limits, with paid tiers for heavier users, and nothing publishes, sends or spends without the owner's sign-off. So no, Muse is not going to reject your CV this week.

Read the pitch again, though, because the framing is the story: give the agent a goal, and it works toward it. Meta says 200 million small businesses can be found on Facebook, and it built this after hearing that owners are short on hours, not ideas. Now think about which task eats the most hours at a 12-person company with no HR department. It's hiring. Sixty CVs land in one inbox, the owner reads them between shifts, and the whole thing drags on for three weeks. The moment that owner has an assistant that already knows what the business sells and how it talks, those CVs become the obvious next thing to paste in.

Here's the part most CV advice hasn't caught up with. Screening at a small company was always the loosest, most human part of hiring — a quick skim, a gut call, a name someone recognised. Replace the skim with a language model and it doesn't get more mechanical; it gets more interpretive. My honest take is that this is harder to game than a big-company ATS, not easier. An ATS looks for strings and hands the recruiter a list. An agent reads your whole document, forms an impression, compresses it into two sentences, and that summary — not your CV — is what the owner actually reads. You're writing for a reader who paraphrases.

Does a small company use an ATS, or just an AI assistant?

Most small companies don't run a real applicant tracking system. Published estimates of small-business ATS adoption swing between roughly one in five and one in three, and the width of that range tells you how soft the numbers are. What's consistent is the gap: near-universal adoption at enterprise scale, patchy and inconsistent below about 50 employees. If you applied by emailing a PDF to a person's address, or through a form on a Squarespace site, or via a LinkedIn post that said "DM me", no ATS touched your file. A human downloaded it. And increasingly, that human opens a chat window before they open the attachment.

That changes what failure looks like. An ATS rejects you quietly and consistently — same rules, same parse, every time. An AI assistant is inconsistent by design: the owner's prompt on Tuesday isn't the prompt on Thursday, the model may only see the first page if the paste got truncated, and a document that extracts as scrambled text produces a confident summary of nonsense rather than an obvious error. Nobody tells you. You get the polite no, or nothing at all. The practical consequence is that you can't optimise for one system's quirks anymore; you have to be legible to both, which in practice means being legible to plain text.

There's a second difference that cuts in your favour. Big-company screening is ruthless about titles and years — the requisition says five years, you have four, you're out. A small-business owner asking an agent to compare six candidates is usually asking a fuzzier question: who can do this job on Monday, who will stick around, who has handled the adjacent stuff too. A model is genuinely good at that kind of reasoning, provided the evidence is on the page. Vague CVs that survived human skimming because the layout looked sharp now fall apart, because the agent has nothing concrete to summarise.

How does an AI agent read your CV differently from an ATS?

An AI agent reads your CV as one long stream of extracted text, in the order the file gives it up — and then rewrites it in its own words. That second step is the one that catches people out. Before any reasoning happens, your PDF has to be converted to text, and that conversion is where designed CVs die. Two-column layouts interleave: your job title from the left column lands next to a skill tag from the right, so "Operations Manager" becomes "Operations Advanced Manager Excel". Text boxes float loose. Icons carry no text at all, so a phone glyph next to your number means nothing. Headers and footers sometimes vanish, taking your contact details with them.

An ATS at least fails loudly enough that a recruiter can spot a mangled profile. A chat window doesn't. Hand a model garbled text and it will smooth it into plausible prose, quietly inventing the connective tissue and flattening the specifics you worked hardest on. It may merge two employers into one. It may report your most recent role as the one that happened to appear first in the extraction. Nothing flags an error, because as far as the model is concerned there wasn't one. If your CV is a designed template and you've never seen what a parser actually pulls out of it, that's the single highest-value thing to check.

Position matters more than it used to, as well. When an owner pastes six CVs into one conversation, the window fills fast and attention concentrates at the top of each document. Whatever sits in your first fifteen lines does most of the work. There's no keyword-match percentage being calculated anywhere — no score, no threshold, nothing you can tune to a number. What exists instead is a judgement, built from whatever the extraction handed over. Plain single-column structure, real section headings, dates in a consistent format, and no critical information trapped in an image or a sidebar.

How do you write a CV an AI agent can quote back to the owner?

Give the agent a summary worth quoting, because it's going to write one either way. Four lines at the very top, under your name and contact details: your role and years, the stack or tools you actually use, the scale you've operated at, and one number that proves it. "Operations manager, seven years in multi-site hospitality. Rota and payroll for 40 staff across three sites. Cut agency spend 31% in one year. Level 3 food safety, full UK driving licence." That's the paragraph the model will lift almost verbatim, because it's already compressed. Write it yourself or let the machine write a worse one for you.

Then mirror the ad's vocabulary explicitly, in the bullets themselves. If the posting says "stock control", don't write "inventory management" and assume the model will connect them — it usually will, but the owner reading the summary won't see their own phrase come back, and that recognition is half the battle. Keep one term per concept, use it consistently, and put it in the bullet where you actually did the work rather than stranding it in a skills list. Spell out abbreviations once: "Xero (accounting)" costs you two words and removes any doubt. Then run the finished file through a free CV analysis to see how it scores on clarity, structure and ATS compatibility before you send it anywhere.

Practical file hygiene still decides more outcomes than it should. Export a real text PDF, never a scan or a screenshot — if you can't select the text with your cursor, neither can the agent. Name the file with your name and the role, because that string often becomes the only label in the owner's folder. Keep it to two pages. Put dates in month-year form on every entry and explain gaps in one short line rather than leaving a hole the model will flag as unexplained. If you're rebuilding from scratch, an AI CV builder that exports clean single-column PDF and Word files saves you fighting a template.

How do you tailor your CV to a vague small-business job ad?

Treat the badly written ad as your keyword source anyway — clumsy wording is still the wording the owner will recognise. Small-business job ads are often written in fifteen minutes by the person who'll be your manager, and they read like it: three responsibilities, one line about culture, "must be a self-starter who can wear many hats". There's no competency framework hiding behind it. So mirror what's there literally, including the odd phrasings, and then fill the obvious gaps yourself. If the ad lists "bookkeeping and general admin" for a 12-person firm, it also means invoicing, chasing payments, supplier calls and whatever the owner currently does at 10pm.

Answer the generalist question explicitly, because it's the real question. At a company this size, nobody is hiring for a narrow slot — they're hiring someone who reduces the number of things the owner personally touches. Show range with evidence: the three functions you've covered, the systems you set up rather than inherited, the time you onboarded your own replacement. Concrete beats broad. "Ran the switch from spreadsheets to Xero, trained four people, closed month-end in three days instead of nine" tells a model more about your adaptability than any sentence containing the word "adaptable" ever will.

Two more things worth doing before you hit send. First, include the local specifics a small employer cares about and a big one automates away — your notice period, whether you can work the shifts advertised, tickets and licences, your actual commute or time zone. An agent asked to shortlist will surface exactly these, and a blank space reads as a no. Second, test your own file the way the employer will: paste the text into a chat and ask it to summarise you in two sentences and name your weakest area. If that summary isn't the one you'd want the owner to read, the CV needs another pass, not the model.

Frequently asked questions