Of Meta's AI features for small businesses, three are worth a trial now: text variations for ad copy (low risk, easy to review), image expansion for Stories and Reels placements (check every crop), and Meta Business Agent in messaging, only with handoff rules set and the first weeks read. Test Advantage+ campaigns against your own setup before handing it your budget.
"Worth switching on" means something specific here: a feature earns its place if it saves you time or money on a job you already do, if you can check its output before a customer sees it, and if you have a rule for when to switch it off. Feature names and availability change often, so the facts below date from November 2026 and every claim about availability is worth confirming in your own account.
How this verdict was reached, and what it is not
The verdicts rest on Meta's published descriptions of each feature, its developer documentation, and the limits it states itself. I have not run your ads, and no review of a platform's AI can predict how it performs in your market, with your product and your budget. So each feature below ends with a test: what to run, for how long, and the result at which you keep it. The scorecards shown are illustrative layouts with made-up numbers to show you what to record; they are not measured results and should not be quoted as such.
One rule applies to all of them. Switch on one feature at a time, so a change in results has one cause. If you switch on five at once and sales rise or fall, you will not know why.
Meta Business Agent: the four controls decide the verdict
Meta Business Agent is an AI agent that replies to customers for your business: it answers questions, asks customers for details, recommends products and supports sales. Meta's page says it is available to eligible businesses in select markets, with more coming, and it does not name them. It focuses on the WhatsApp Business app, and the page's FAQ says the agent can also represent your business in Messenger and in ads you create. Whether it works on Instagram messages is not confirmed on the pages I could read, so check your own account. To find it in the WhatsApp Business app, open Tools and look for Meta Business Agent; if the entry is not there, it is not available to your account.
What matters for a small business is the four controls Meta describes:
| Control | What it does | What to set at the start |
|---|---|---|
| Knowledge | Learns from your Facebook Page posts, past WhatsApp chats and data you add, such as your website and price lists | Give it a short, accurate FAQ and current prices. Remove old posts that give outdated information |
| Personality | Sets the voice; rewriting an answer updates future replies | Write three sample replies in your own words and edit early answers |
| Audience | Everyone, new customers only, or customers coming from ads | Start with customers from ads or new customers, not existing ones |
| Handoff | Topics the agent leaves to you; when you step in it learns from the conversation | Complaints, refunds, price negotiation, health, safety, legal topics, anything with a deadline |
Cost needs care. Meta says the WhatsApp Business app stays free and Meta Business Agent includes a free monthly allowance, without publishing the number. Beyond it, Meta One subscriptions apply per business profile or phone number, with a possible seven-day trial; the tier prices announced in September 2026 were reported as $14.99, $49.99, $149 and $499 a month, and Meta says plans and availability vary by region, app and account. On the separate WhatsApp Business Platform (the API route that software providers use), Meta's developer pricing page has charged for agent messages since 1 August 2026 at $2 per million tokens, with no free window; Meta estimates a typical message at 20,000 to 25,000 tokens, or roughly 4 to 5 cents. Confirm the live figures before you decide, since this is the part most likely to have moved.
Verdict: switch on with a trial, for a narrow audience. It earns its place if you answer the same ten questions all day (opening times, prices, how to book). It is a poor fit if most of your messages are complaints or complicated custom jobs. For how a restaurant or a hotel should judge this, see whether a restaurant should switch on Meta's agent; the general picture is in what Meta AI in WhatsApp means for your business chats.
A two-week test you can run
- Before switching on, log 30 recent conversations: how many were simple questions, how many needed a person, and how long the first reply took.
- Switch on with Audience set to customers from ads or new customers, and Handoff set as above.
- Read every conversation for the first week, and a sample of half for the second.
- Score each as: correct and helpful, correct but clumsy, wrong, or should have been handed off.
An illustrative scorecard for a physiotherapy clinic that tried it on new-patient enquiries (made-up numbers, to show the layout):
| Week | Conversations | Correct and helpful | Correct but clumsy | Wrong | Should have been handed off |
|---|---|---|---|---|---|
| 1 | 40 | 26 | 8 | 3 | 3 |
| 2 | 46 | 35 | 6 | 2 | 3 |
The decision rule matters more than the numbers: keep it if "wrong" stays under about one in twenty, nothing "wrong" involves price or clinical advice, and every "should have been handed off" leads to a Handoff rule you can add. In this example the wrong answers would be read for what they were: if one gave an old price, the Knowledge source needs correcting before anything else. Stop the trial at once if the agent states a price, a result or a clinical claim you did not provide.
Two policy points to hold in mind. WhatsApp's Business Messaging Policy says that a business using automation on the Business Platform must have "prompt, clear, and direct" escalation paths, such as handing the chat to a person, a phone number or an email address, so never remove the handover path. And if you use a third-party chatbot, not Meta's own, ask the provider whether Meta's platform terms allow it, since the terms bar AI providers whose general-purpose AI is the main product rather than an incidental feature.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhat to put in the Knowledge source before launch
The agent is only as accurate as what it learns from, and Meta says it learns from your Page posts, past chats and the website and price lists you provide. That means old promotions and one-off answers can become "knowledge". A short clean-up pays for itself. Write a one-page sheet for the agent and check it against your real prices: opening hours and holiday changes, what you do and do not offer, prices or price ranges for your five most-asked services, booking and cancellation rules, and what to say when someone asks something not on the sheet (for example, "I will ask the team and they will reply today").
An illustrative before and after for a bookkeeping practice. Before: the agent's source includes a Page post from two years ago advertising a "free first month" that no longer exists. A new customer asks about it and is told it applies. After the clean-up: the post is hidden, the sheet says "no current offers", and the agent's reply to the same question is "I cannot see an offer at the moment; the team can tell you what is available". The difference between those two answers is a customer who feels misled and one who does not.
Advantage+ campaigns: a trial against your own setup, not a handover
Meta's Advantage+ campaigns apply AI across audience, placement and budget. The page names three: Advantage+ sales campaigns (formerly Advantage+ shopping campaigns), Advantage+ app campaigns and Advantage+ leads campaigns. Meta's page cites better results (20 percent improved CPA for sales campaigns, 10 percent lower cost per qualified lead for leads campaigns and 7 percent improved CPA for app campaigns), but the footnotes behind them were not shown on the page, so read these as Meta's claims for the conditions it tested. For some objectives Advantage+ is the default (Meta's leads page says every leads campaign is Advantage+ On by default), and audience and placement controls set at account level still apply.
Verdict: test it, with a ceiling and a comparison. It suits an account with a clear conversion event and enough volume (many small accounts do not have it) and fits worst where your leads are slow to qualify or your margin is thin. The generic test is a split: run an Advantage+ campaign beside your current setup with the same budget and the same creative for a fixed spell, and compare cost per qualified result. For the question of how long to wait before judging, how much to spend on AI-run ads before judging results sets the budget; whether a small business should use Meta Advantage+ gives the longer argument.
An illustrative layout for an IT support firm testing lead ads over three weeks:
| Spend | Leads | Qualified leads | Cost per qualified lead | |
|---|---|---|---|---|
| Current campaign | $300 | 20 | 5 | $60 |
| Advantage+ leads campaign | $300 | 28 | 5 | $60 |
The point of the illustration is the middle column: Advantage+ produced more leads, but not more qualified ones. A campaign judged on raw lead count would have won. So define "qualified" before the test, and count it yourself from the leads, not from the platform's dashboard. Stop rule: if cost per qualified lead is worse by more than a fifth after the agreed spend, or if lead quality falls, go back to your own setup.
Text variations for ad copy: low risk, if a person reads it
Meta's generative text feature produces alternative versions of your primary text (the main copy above the image), drawn from the text you supply, your previous ads or your Page content. In Meta's developer documentation, you must supply a primary text, the opt-in applies per ad, and an ad opted in through the API is created paused so you can read the generated text before setting it live. That design is a good model for how to use it anywhere: write one version you would be happy with, let the tool offer variations, and read each before it runs.
Verdict: switch on, with a read-through rule. The typical fault is a claim you cannot support (a guarantee, a number, a promise of speed). If you sell anything with advertising limits, such as health, finance or housing, the variations need a compliance read as well. A simple test: take 10 ads, count the variations you would publish unchanged, edited and rejected. If fewer than half are usable, the saving is small. For writing the original copy so the variations start well, see how to write Facebook and Instagram ad copy with AI.
Image expansion and background generation change your picture
Image expansion generates new pixels to extend a picture into other shapes. The documentation lists it for the Instagram standard, Instagram Stories, Instagram Reels, Facebook Reels and mobile feed placements. It is most useful when you have a square or landscape photo and want a tall version for Stories and Reels without cropping the product. Background generation is narrower: it creates new backgrounds for product images and works only with dynamic product ads or Advantage+ catalogue ads in mobile feed, for eligible images.
One setting to check before anything else: Meta's Advantage+ creative page says some enhancements are on by default in Ads Manager and can be turned off at any time, and that the previews show samples rather than every variation that may run. So open the creative settings on each ad, switch off what you have not decided to test, and treat the preview as a spot check rather than a full proof.
Verdict: image expansion, switch on and check every crop; background generation, only if you run catalogue ads, and check the product. Generated pixels can warp edges, hands, logos or text that sits near the border. The check is simple and manual: open each preview in the placement you will use, zoom in on the logo, the product edges and any text, and confirm the product you are selling is the product shown. An expanded background that adds a second, imaginary item to your product photo is a misleading ad, and the responsibility is yours.
An illustrative test for a bakery's square product photos: expand 12 images for Stories; count how many look right at full size, how many need a manual crop, and how many are distorted. If more than one in four need rescuing, the saving over a quick edit in a photo editor is small, and the manual route is better.
The AI label and what it means for your ads
Meta labels AI-generated content with "AI info". Its help page says ads fully created or materially edited with Meta's own generative AI tools will carry the label, rolled out gradually, usually in the three-dot menu, or beside "Sponsored" when the image shows a photorealistic AI-generated person. The same page still says ads made with third-party AI tools get no such label, while a Meta newsroom update of 1 June 2026 says it will detect those ads automatically and label them in "About this ad". The two statements sit side by side, so do not promise customers that no label will appear, and do not remove a label's effect by editing around it. If your brand relies on photography being real, such as a food business or a clinic showing results, avoid AI-generated images of people or outcomes altogether, whether a label appears or not.
Meta One and the paid tier: buy for a reason, not for the AI
Meta One is a subscription for businesses. The pages I read describe benefits in groups: brand protection (a verified badge, a custom web address, impersonation protection), operations (including a Business Agent that replies around the clock and books appointments) and personalisation. A subscription applies per business profile or phone number and may include a seven-day trial that renews automatically unless you cancel.
Verdict: buy it if the verified badge or impersonation protection is the reason, not because the AI is promised. If an agent is the only reason, test the free allowance first and see whether you ever reach it. Put the trial's end date in your calendar the day you start, because trials that renew on their own are the common small-business surprise.
A summary table you can take to your next marketing meeting
| Feature | Verdict | Test | Stop rule |
|---|---|---|---|
| Meta Business Agent (WhatsApp, Messenger) | Trial, narrow audience, handoff set | Two weeks, read all conversations in week one | Any invented price or advice; wrong answers above about one in twenty |
| Advantage+ campaigns | Test against your own setup | Same budget and creative, compare cost per qualified result | Cost per qualified result more than a fifth worse, or lead quality falls |
| Text variations | On, with a read-through | Ten ads: publish, edit, reject | Fewer than half usable, or any unsupported claim |
| Image expansion | On, with a manual check of each crop | Twelve images for Stories | More than one in four distorted |
| Background generation | Only for catalogue ads, checked | Twenty products, compare the product shown | Product looks different from the real item |
| Meta One | Buy for the badge and protection, not the AI alone | Use the free allowance first | Cancel before the trial ends if unused |
Setting up the trial so the results mean something
A small business can ruin a good test by changing three things at once. Keep to a short routine: write down today's numbers (cost per lead, reply time, share of messages handled by staff), change one feature, give it a fixed period, and write down the same numbers at the end. Keep a one-line note of anything else that changed in the period, such as a sale, a holiday or a new price, since any of them can look like an AI effect.
Tell your staff what is on. A receptionist who does not know an agent is answering overnight messages will contradict it the next morning, and a customer will notice. Add a short line to your profile or first reply saying that messages may be answered by an automated assistant and that a person can take over; this is both honest and consistent with Meta's expectation of a clear route to a human.
Finally, keep a record of what the agent said. Export or screenshot the conversations the agent handled for the first month, so that if a customer disputes something, you know what was said. Check the data and privacy settings of anything you connect before you put customer information into it. If you want help working out which of these features deserves a trial on your own account, and what the stop rules should be, that is part of what an AI implementation consultation covers; the starting point is the contact page.
Questions small businesses ask about Meta's AI tools
Is Meta Business Agent available to every business?
No. Meta's own page says it is available to eligible businesses in select markets, with more to follow, and does not list them. Look for Meta Business Agent under Tools in the WhatsApp Business app; if it is not there for your account, it is not available to you yet.
Does Meta label ads made with its AI tools?
Yes. Meta's help page says ads fully created or materially edited with its own generative AI tools get an "AI info" label, rolled out gradually. The same page says ads made with third-party AI tools get no label, while a June 2026 newsroom update says Meta will detect and label them in "About this ad". Do not rely on a label being absent.
Can I stop Meta's AI changing my ad images?
Yes. Meta says Advantage+ creative enhancements can be turned off at any time, but some are on by default in Ads Manager, so check the creative settings on each ad rather than assuming they are off. Review the preview before publishing, and switch off any enhancement that distorts your product, logo or text.
What does Meta Business Agent cost?
On the WhatsApp Business app, Meta says it includes a free monthly allowance, and its published figures give no number. Beyond that, Meta One subscriptions are reported from $14.99 a month. On the WhatsApp Business Platform route, Meta's developer pricing page lists $2 per million tokens, which it estimates at about 4 to 5 cents a message.
Should a small business let AI answer customers without checking?
Not at the start. Review the first weeks of conversations, set the handoff control for price disputes, complaints, health or legal topics, and keep a person able to take over. WhatsApp's messaging policy also requires clear escalation paths, such as a person or a phone number, when automation is used.
Further reads
- How Meta and Google Use AI to Run Your Ads: A Plain-English Guide — How Meta and Google use AI to run ads, in plain words.
- Can a Takeaway Take Orders on WhatsApp With AI? — Can a takeaway take orders on WhatsApp with AI?
- WhatsApp for Small Hotels: Meta's AI Rules, Templates and Handover — Meta's AI rules, templates and handover for small hotels.
- Salon AI Chatbot: ManyChat, Meta Business Agent or Booking-App Bot? — Meta's agent against ManyChat and booking-app bots.
- AI for Meta Ads: What Agencies Hand to Advantage+ and What They Keep — What agencies hand to Advantage+ and what they keep.
- How to Write Instagram Captions With AI That Sound Like You — Write organic Instagram posts that sound like you.
- ChatGPT Ads Cost for a Small Store: CPC, CPM and a Test Budget — What ChatGPT ads cost a small store in clicks and impressions, plus a sixteen-day test budget with written stop rules.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Meta Business Agent and Meta One pages on whatsappbusiness.com, Meta's Advantage+ campaigns page, Meta's Marketing API documentation for generative AI creative features and Meta's AI info labelling pages (all read in November 2026); test scorecards are illustrative.