Nov 21, 2026Español
GuideGuides · AI agents · Nov 21

What AI agents can and can't do for a small business in 2026

Agents now draft invoices, run marketing plans and chase late payments. They also make compounding mistakes and can be tricked by text they read. What to check first.

Guides · AI agents

The 30-second version

  • An AI agent is a model that decides its own steps and uses tools, such as your invoicing or email software, to finish a task, rather than following a fixed script.
  • Real small-business agents now draft and send invoices, prepare payroll drafts and run marketing plans, usually with an owner approval step built in.
  • The main risks are compounding errors, prompt injection from emails or web pages, and agents holding more access than they need; least privilege and human approval for money moves are the core defenses.

"Agent" has become the word every software vendor uses in 2026, from accounting suites to email marketing tools. For a small business owner, the useful question is narrower: what can one of these systems do safely in my business today, and where should I keep my hands on the wheel?

This guide explains what an agent is, shows real examples available to U.S. small businesses, lays out the risks and ends with a checklist to run before an agent gets access to your accounts or payments.

Part 1: What an AI agent actually is

A chatbot answers. An agent acts. It takes a goal, decides which steps to take, calls tools such as your invoicing software, calendar or inbox, checks the results and keeps going until the job is done or it gets stuck.

Anthropic's engineering team draws a helpful line between two kinds of systems. A workflow runs a model and tools along predefined code paths: the steps are fixed by whoever built it. An agent lets the model direct its own process and choose which tools to use. Anthropic's advice is to start with the simplest option, which is often a single model call or a fixed workflow, and use agents only for open-ended tasks where that flexibility is worth it. It notes that agents trade higher cost and latency for better performance, and that their autonomy creates room for errors that compound from one step to the next.

Many products sold as "agents" are closer to workflows with an AI step. That is not a bad thing. Predictable is usually what a small business wants.

Part 2: What agents can do today, with real examples

Invoicing and collections. In July 2026, Intuit expanded the QuickBooks connector in Claude and the plugin in ChatGPT beyond read-only access. According to Intuit, they can now create, update, send and delete invoices and estimates, set up recurring invoices, send reminders on overdue invoices, add customers and products, and generate payment links for businesses with QuickBooks Payments. Intuit says every action stays reviewable and human-verified, with the typical flow being prompt, review, then send. Payroll and lending access through those assistants is limited to questions for now.

Marketing. Constant Contact's Catalyst, announced in November 2026, builds a marketing plan from a business's online presence and runs it across email, social, SMS, ads and Google Business Profile, after the owner approves the plan and content. It is in early access. We covered it in our story on Catalyst.

Customer service. AI agents on help desks now answer common questions and complete simple workflows such as order lookups or rebooking, handing off to staff when unsure. Our chatbot setup guide covers that starting point.

Spending. Card platforms have started issuing virtual cards with preset limits that an AI agent can use, as in Cardless AI's launch. Tools for agents to complete payments on their own are still being built.

Part 3: What agents can't do well, or shouldn't do

  • Know your business by default. An agent works from the data and instructions you give it. If your customer list or price book is messy, its output will be too.
  • Guarantee accuracy. Models still make things up. In a multistep task, one wrong assumption early on can carry through every later step.
  • Act everywhere. Some companies block outside agents. United Airlines said in November 2026 it will prohibit unauthorized personal AI agents from booking or managing its flights, citing account security and unintended purchases. See our report. Check a vendor's terms before pointing an agent at its site.
  • Carry responsibility. If an agent sends a wrong invoice or a misleading email, the business that deployed it answers to the customer and, potentially, to regulators.

Part 4: The risks that matter most

The OWASP GenAI Security Project, a widely cited open security community, publishes a top-10 list of risks for applications built on large language models. Two entries map directly onto agents.

Prompt injection is ranked first. It happens when content the model processes changes its behavior. In the indirect version, the agent reads a web page, file or email that contains hidden instructions and follows them. OWASP notes the instructions need not even be visible to a person. An agent that reads your inbox and can also send email is exactly the setup an attacker would target.

Excessive agency describes an agent that can do damage because it has too much power. OWASP breaks the causes into three: tools the agent does not need, permissions broader than the task requires, and autonomy to take high-impact actions without independent checks. Its prevention list includes giving the agent only the tools it needs, using the narrowest permissions possible such as read-only access, acting under the specific user's scoped credentials, requiring a person to approve high-impact actions, and enforcing permissions in the downstream system rather than trusting the model to police itself.

Part 5: Checklist before an agent touches accounts or payments

  • Scope: Write down the one task the agent is for. Turn off every connected tool that task does not need.
  • Permissions: Start read-only. Use a separate user or API key for the agent so its actions are logged under its own name and can be revoked in one step.
  • Approval: Require human sign-off for anything that sends money, issues refunds, deletes records, changes bank details or contacts customers in bulk.
  • Money limits: If an agent can spend, give it a dedicated virtual card with a low cap, a merchant restriction and an expiry date, never your main card or bank login.
  • Audit trail: Confirm the product logs what the agent did and why, and read those logs weekly at first.
  • Untrusted input: Be cautious about agents that both read outside content, such as email or web pages, and can take actions. That pairing is where prompt injection does damage.
  • Vendor terms: Check data retention, whether your data trains the vendor's models, and who is liable for errors.
  • Stop switch: Know how to pause the agent and disconnect it from each account.

Anthropic's guidance for builders points the same way: test in a sandbox first, have the agent check real results at each step, and build in human checkpoints and stopping conditions. For a small business, that translates into a short pilot on low-stakes work, a person reviewing every action for the first few weeks, and more autonomy only where the record shows it has been earned.

By the numbers

LLM01OWASP's rank for prompt injection among LLM risks

LLM06OWASP's code for excessive agency

July 28, 2026QuickBooks gave its Claude and ChatGPT connectors invoice actions

What it means for your business

Treat an AI agent like a new hire on probation: give it one narrow job, read-only access where possible, and a person who approves anything that sends money, deletes records or contacts customers. Start with tasks where a mistake is cheap and visible, such as drafting invoice reminders, before anything that pays bills.

Common questions

What is the difference between an AI chatbot and an AI agent?

A chatbot answers questions. An agent can take actions through connected tools, such as creating an invoice, sending an email or updating a record, and it decides the steps itself. Anthropic's engineering team separates fixed workflows, which follow predefined code paths, from agents, which direct their own process and tool use.

Can an AI agent pay my bills or move money?

Some tools can generate payment links or prepare payments, but most products small businesses use today keep a person in the approval loop for money. Security guidance from OWASP says high-impact actions should require human approval, and that is the setting to insist on.

What is prompt injection?

It is when text the agent reads, such as a web page, file or email, contains instructions the model follows. OWASP lists it as the top risk for applications built on large language models, and it matters most when the agent can send email or change data.

Sources

This story is prepared with AI assistance and reviewed and approved by a person before it is published. How we work. Spotted a mistake? Tell us.