Quick answer
Vendor pages from Google, AWS and HubSpot document AI that, for example, drafts email, writes up meetings, transcribes voicemail, pulls typed and handwritten text from documents and answers questions about a contact or company. In a Census survey (17 Nov 2025 to 8 Feb 2026), 17.9% of all US businesses reported using AI in the last two weeks.
Key takeaways
- The five vendor pages read here document AI for email drafts, meeting notes, voicemail text, text pulled from documents and questions about a contact or company. These are the vendors' descriptions, not test results.
- Three of the five pages say the output can be wrong or missing: Google's Gmail, Meet and Voice help pages. The Meet sentence comes from a troubleshooting section that lists causes. The Textract developer guide and HubSpot's Agent Hub marketing page state no such limit on the pages read; they may carry limits elsewhere.
- Census Bureau survey, 17 Nov 2025 to 8 Feb 2026: 17.9% of all US businesses used AI in any business function in the last two weeks. It is self-reported use, not a measure of results or of AI agents.
- This site's suggestion: match each task to a check, such as reading before sending, listening to the audio for names and numbers, or spot-checking extracted totals against the paper.
What can AI do for a small business today?
5 vendor pages describe, for example, AI that drafts email, takes meeting notes, transcribes voicemail, pulls typed and handwritten text from documents and answers questions about a contact; 3 of the 5 also say results can be wrong. Those are the vendors' descriptions, read on 7 October 2026, not test results, so each task needs a check.
These are 5 pages and 6 task rows, chosen because each had a vendor sentence we could quote. Tasks for which we found no such sentence (scheduling, review replies, phone answering) are left out, not ruled out.
The table below links each task to this site's guide for it. For how the pieces fit into a whole, see the AI automation guide.
Which tasks do vendors document, and what limit do they state?
Each row below is one vendor's own sentence about the feature, next to what the same page says about mistakes. Where the page read has no warning, the table says so. Availability conditions the page attaches to the feature are summarized in each Google row.
One vendor per task, as examples of documented features, not a list of tools to buy; other vendors may offer the same. Three rows are Google products because its help pages state both the feature and a limit.
| Task (live guide) | What the vendor's page says it does | What the same page says about errors | Check this site suggests |
|---|---|---|---|
| Email replies (guide) | Google, Gmail Help, for Help me write: "Generate a new email draft, such as a birthday invitation or an introduction to a potential business contact." The page lists it for an eligible Google Workspace or Google AI plan (supported languages) and for a personal Google Account (supported languages, US only). | "Gemini features may suggest inaccurate or inappropriate information." | Read every draft before it goes out. |
| Meeting notes (guide) | Google, Meet Help: "Automatically capture meeting notes organized in Google Docs and share with your team." The page says the feature is only available with a work or school Google Account or an eligible Google AI plan, for meetings spoken in one of its listed languages. | "Sometimes, the meeting summary can be incomplete, inaccurate, or not generated." This sits in the page's troubleshooting section, which lists causes such as a policy breach, connectivity problems or a meeting shorter than the recommended length. | Compare the action items with your own memory the same day. |
| Calls and voicemail (guide) | Google, Voice Help: "You can listen to voicemails or read voicemail transcripts." For work or school accounts the page lists the languages Google Voice transcribes. | "Voicemail transcripts may be incorrect or missing." | Listen to the audio for any name, number or address. |
| Documents (guide) | Amazon Web Services, Textract developer guide: "Detect typed and handwritten text in a variety of documents, including financial reports, medical records, and tax forms." | No error limit on the page we read, which calls the service's analysis "powerful and accurate"; it does list quotas such as file size and supported languages. | Check extracted totals against the paper for a sample. |
| Customer questions (guide) | HubSpot, Agent Hub (marketing page): "Customer agent resolves common questions 24/7, with answers grounded in real contact and contract history, so your team is free for the conversations that need a human." | No error warning on the page we read. | Read the conversations it handled, and what it passed to a person. |
| Leads (guide) | HubSpot, Agent Hub (marketing page): "Data agent answers custom questions about any contact or company using the signals already in your CRM, so reps know exactly who to reach and why." | No error warning on the page we read. | Confirm a contact's details yourself before you reach out. |
- The voicemail row covers reading a transcript; the live guide covers answering services, and Google's page says nothing about AI answering calls.
- The Textract page is written for people adding the service to applications, so it is a developer description, not an owner's setup guide.
- HubSpot's page is a marketing page whose sentences do not state a plan or edition.
- The leads row is about answering questions from data already in a CRM, not about finding new leads.
- A missing warning describes only the page we read: the Amazon and HubSpot pages are overview and marketing pages that may carry limits elsewhere, and this table does not say those products have none.
What do businesses actually use AI for?
In the Census Bureau's Business Trends and Outlook Survey (AI supplement, 17 November 2025 to 8 February 2026), 17.9% of all US businesses used AI in any business function in the last two weeks. For the last six months, 14.3% used it in sales and marketing and 7.6% in customer service. All three are shares of all businesses, self-reported, from that single survey; the two-week and six-month figures cover different periods and should not be added together.
The match to the table is this site's, and loose. Sales and marketing is the nearest Census function to the leads row, and customer service the nearest to customer questions. No Census function this site registered matches email replies, meeting notes, voicemail or documents, so this site prints no share for them. The survey records use, not results or value, and does not measure AI agents. The full breakdown by industry is in what businesses use AI for, and by size in AI use by business size and industry.
How do you decide which task to try first?
This site's rule of thumb, not a finding: pick a task you already do every week, where a person can check the output before it reaches a customer and a mistake is cheap to fix. Then use the vendor's own stated limit, where its page gives one, as the first thing to test. A row with no error warning is not a verdict on the tool; it means the page read gives no guidance on errors, so you write the check yourself.
This site's reading of the conditions the five pages state, row by row:
- Gmail drafting fits you if you have an eligible Google Workspace or Google AI plan in a supported language, or a personal Google Account in the US only. First test: whether a draft is inaccurate or inappropriate, which the page says may happen.
- Meet notes fit you if your account is set up through work or school or you have an eligible Google AI plan, and your meetings are in a listed language. First test: whether the summary is incomplete, inaccurate or not generated.
- Voice transcripts: the page lists transcription languages for work or school accounts, so check that yours is there. First test: whether a transcript is incorrect or missing.
- Textract fits you if someone builds software for you, because it is a developer service you add to an application. The page states no error limit, so first test extracted totals against the paper.
- HubSpot fits you if the vendor confirms your plan includes it: the page read is a marketing page that states no plan or edition, so ask the vendor. The page states no error limit, so first test what the agent handled and passed to a person.
The Census shares do not pick the task: they show where businesses report using AI, not whether it worked.
Ask any seller these before paying:
- What does your documentation say the output gets wrong, and where is that written?
- Which plan, account type and language does the feature need? (Google's pages attach plan, account and language conditions to the features above.)
- What happens to a case the AI cannot handle: does it pass it to a person, and with what context?
Does a documented feature mean it will work for my business?
No: a documented feature means a vendor says the feature does the task, not that it performs well for your business. Nothing here measures accuracy, time saved or cost. Another product may behave differently, and features change; every sentence here was read on 7 October 2026.
What is a sensible next step?
This site's suggestion: pick one row, run it on ten of your own real items, and apply its check. Ten items is enough to see whether the errors the page warns about show up in your material. If they do, keep a person in the loop; if the row had no stated limit, your ten items are the only evidence you will have.
How was this page made?
This page was drafted with Claude, an AI assistant made by Anthropic, and checked against each vendor's published pages. Anthropic is not one of the vendors covered. Vendor pages were read from outside the United States on 7 October 2026.
Our data
Nationally, 17.9% of all US businesses used AI in any business function in the last two weeks; in the last six months 14.3% of all businesses used AI in sales and marketing and 7.6% in customer service (Census Bureau BTOS AI supplement, 17 Nov 2025 to 8 Feb 2026, self-reported).
Method: Shares of all businesses taken from the Census Bureau's Business Trends and Outlook Survey AI supplement tables (2026), as registered in this site's reviewed study. Any-use covers the last two weeks; the function items cover the last six months. One cross-section, one respondent per business. The survey does not measure AI agents or value.
FAQ
Can AI run a small business by itself?
The pages read here do not say that. They describe AI doing single tasks, such as drafting an email or transcribing a voicemail. This page found no source for AI running a business on its own.
What can AI do for a business with no technical staff?
The Gmail, Meet and Voice help pages describe features switched on inside tools you may already use, subject to plan, account type and language. Textract's page describes a service for adding to applications, so it is aimed at developers.
Does a missing warning mean the tool is accurate?
No. A page that states no error limit tells you only that the page read gives no guidance on errors, so you write the check yourself. AWS's page describes Textract's analysis as "powerful and accurate" and still gives no error limit.
Is a vendor’s feature description proof it works for my business?
No. It tells you what the vendor says the feature does. Only a trial on your own emails, calls or documents shows whether it works for you.