An AI assistant for business is a chat tool you prompt: Claude, ChatGPT, Microsoft 365 Copilot, Gemini inside Google Workspace. It drafts, summarizes, researches and answers on demand, in a window you have open. It does not hold a recurring job in your business. Four questions tell you which of your work sits on each side of that line.
The handoff test. Ask these about one piece of recurring work:
- Does it start without anyone asking?
- Does it have to reach a system of record, rather than a file you paste in?
- Does anyone besides you need to see that it ran, and what it did?
- If it stopped quietly, how long before someone noticed?
Questions one and four decide which side the work falls on. Answer no to the first and "same day" to the last, and an assistant can own it. Answer yes to the first, or weeks to the last, and the work needs an owner instead. Questions two and three do not change the verdict; they tell you how much the handoff will cost.
What is an AI assistant for business?
It is a general-purpose chat tool, sold per seat, that you direct one request at a time. The current ones remember you, read your files and connect to your apps. Anthropic's pricing page, read on October 14, 2026, lists memory across conversations and app connections on the free tier of Claude, and the $17 a month Pro plan on annual billing, $200 up front or $20 billed monthly, adds the ability to hand off and schedule tasks.
That last item matters for how you read everything else written about this. Most "assistant versus agent" explainers draw the line at scheduling, and that line moved. A $17 subscription can now put something on a clock, which means the difference businesses keep running into has to be somewhere else.
What can an AI assistant actually do for a business?
Everything where you are in the room. In our experience the honest list is short and useful: draft the thing, rewrite the thing, summarize the document you handed it, research a company before a call you already booked, turn your notes into a brief, check your reasoning, write the formula or the query you half remember.
Notice what those have in common. You decided the work needed doing, you supplied the context, and you read the output before it went anywhere. The assistant made your task shorter. It did not take the task off your week, because nothing about the arrangement remembers the task exists.
That is a real benefit and it is worth paying for. It is also why a lot of businesses buy seats, use them happily, and six months later still have the same list of things falling through. The seats were never pointed at that problem.
Where does an AI assistant stop being enough?
At the first job nobody is awake for. Below is the handoff test run against ten jobs small businesses actually have.
| Recurring job | Starts on its own | Needs a system of record | Someone else needs the record | Silence noticed in | Verdict |
|---|---|---|---|---|---|
| Rewrite a job post before you publish it | No | No | No | Same day | Assistant |
| Summarize a vendor contract you paste in | No | No | No | Same day | Assistant |
| Research a prospect before a call you booked | No | Sometimes | No | Same day | Assistant |
| Turn your meeting notes into a client recap | No | No | Yes | Same day | Assistant |
| Rework a proposal you already wrote | No | No | No | Same day | Assistant |
| Triage overnight web form submissions | Yes | Yes | Yes | Hours | Needs an owner |
| Build Monday's pipeline report from the CRM | Yes | Yes | Yes | A week | Needs an owner |
| Chase the invoices that went past 30 days | Yes | Yes | Yes | Weeks | Needs an owner |
| Reconcile new signups against billing nightly | Yes | Yes | Yes | Weeks | Needs an owner |
| Collect project status from four people weekly | Yes | Yes | Yes | A week | Needs an owner |
The two middle columns move around, and they do not change the verdict. The prospect research touches your CRM on a good day and a browser tab on a bad one. The client recap goes to somebody outside the building, which makes it worth proofreading, and still leaves it a job you start by pasting in your notes.
What settles each row is the first column and the last. Every job in the bottom half starts on a clock or an event, and every job in the bottom half can fail for days without anyone finding out. A chat window answers neither. Forget to open it on Monday and nothing in the product notices.
Before you read the bottom half as a shopping list, check whether those jobs need a model at all. Several of them have fixed inputs and rules someone already wrote down, which makes them cheaper and steadier as plain automation. AI Agents vs Automation vs RPA has a four-question routing test for exactly that call.
AI assistant vs AI agent: what actually changes
The clearest evidence for where the line sits comes from the three companies selling you the assistant. All three put the standing-job version in a different product.
Microsoft's own documentation for Microsoft 365 Copilot describes agents as "scoped or focused versions of Microsoft Copilot that act as AI assistants and can automate business processes," and you build them in Copilot Studio, which Microsoft calls "a graphical, low-code studio for building and managing AI-powered agents and workflows." The sentence that gives the whole thing away is further down that page: "Some agents can be given their own account so they can work proactively on tasks and take part in shared business processes, such as onboarding a new employee or coordinating a recurring meeting."
The vendor wrote the handoff into its own documentation, and the giveaway is the account. A seat belongs to a person. A job that runs while that person is asleep needs an identity of its own.
OpenAI splits it the same way. ChatGPT is the assistant; the thing that holds a job gets built against the API, where the agents guide tells you to "choose a runtime based on where you want orchestration to run and who should manage the state between tasks." Google separates them in the admin console, where Gemini in Workspace apps and Gemini Enterprise agents are different sections with their own licenses and access controls.
Read those three together and the difference stops being philosophical. Crossing the line means adding an account, a schedule, a permission scope, an approval point for anything consequential, a record of what happened, and a named person who hears about it when the thing breaks. None of those six arrive with a seat license, and somebody has to keep doing all six.
What a standing job looks like once you run one
We can be specific here, because the agents that run our own company are on the other side of that line. As of October 14, 2026 we have thirteen agents, and twelve of them hold scheduled work: 72 enabled jobs, 71 of them on cron expressions and one on an interval.
The more useful number is in the status column: 59 of those jobs last finished clean, three last finished with an error, and ten have not run yet. Three of them failing at any given moment is normal, and that is the whole reason a standing job is a different kind of product. We watch each job's last run status, its last error and a heartbeat file, because the failure that costs money is the job that goes quiet rather than the job that answers badly. A wrong answer argues with you. A job that stopped firing at 6 a.m. says nothing at all.
If you want the version with real jobs attached to real roles, the agents that run our company goes agent by agent, and how thirteen of them hand off work covers what happens when one job's output is another's input.
Two other things decide whether a job survives contact with your business. The first is whether the systems it has to read are reachable and trusted, which AI Agent Integration works through system by system. The second is which actions stay human. We do not let our own agents send money, make hiring calls or take legal positions, and we have more reason than most to trust them; Human in the Loop for AI Agents goes action by action through what to gate.
How much does an AI assistant for business cost?
Tens of dollars per person per month, which is why the price is the least interesting part of the decision. Claude Pro is $17 a month on annual billing or $20 monthly, as published on Anthropic's pricing page on October 14, 2026. Check each vendor's own pricing page for the rest, because these plans are renamed and repriced often enough that a roundup from last quarter will be wrong.
Comparing one subscription against another is the smaller question. The bigger one is a seat, which buys capability your team still has to point at something, against a standing job, which buys an owner. Our own numbers for the second kind are public: the build starts at $1,000 one time and includes three coaching sessions for your team, and hosting and maintenance start at $200 a month, as shown on our pricing page the day this was published. Market ranges for the other routes, development shops and in-house builds included, are collected in How Much Does AI Agent Development Cost?.
If you are weighing the four shapes this market sells, and the lists blur them badly, AI Agents for Business sets the four side by side with each one's published price and the column the roundups skip.
Frequently asked questions
Is an AI assistant the same as an AI agent?
No, though the capability gap is narrower than most articles claim. Both can remember, connect to apps and run on a schedule. The difference is ownership: an agent has its own account, a defined job, a permission scope, an approval point and somebody who hears about failures. Microsoft's documentation draws the same line when it notes that some agents get their own account so they can work proactively.
Can an AI assistant work while I am offline?
Partly. Scheduled tasks on a paid plan will fire without you, so the simple version of this works. What does not work offline is the accountability: nothing routes the result to a colleague, logs what was done for an auditor, or raises a hand when the task silently stops. If any of those three matter, the work belongs on the other side of the handoff test.
Which AI assistant is best for a small business?
Mostly whichever one sits in the software your team already lives in, which is why the roundups all reach different conclusions and none of them is wrong. If your documents are in Microsoft 365, Copilot has a shorter path to them; if they are in Google Workspace, Gemini does. The choice is worth about an hour. The jobs that failed the handoff test are worth considerably more.
Do I need both an assistant and an agent?
Usually, and they are not competing for the same budget line. Assistants make your people faster at work they are already doing. Agents take recurring work off the list entirely. A business with neither should start with seats, because they are cheap and immediate. A business with seats that is still dropping follow-ups has learned what seats do not fix.
Is my business data safe in an AI assistant?
That depends on the plan you bought and the systems you connect, and it is worth reading your vendor's own data terms rather than a summary of them. The practical control is scope: decide system by system what the tool may read, and keep the consequential actions behind a person. What Access Should an AI Agent Have to Your Business Systems? lays out a permission ladder you can apply to an assistant just as well.
Start with the jobs that failed the test
Run the handoff test over your own recurring work this week and you will end up with two lists. The short one is work your assistant seats should already be shortening, and if they are not, that is a training problem rather than a buying problem.
The other list is the interesting one. Pick the job on it that annoys you most, and write down five things: what starts it, which systems it has to read, what it should produce, who sees the result, and the one action that must stay human. That page is the whole input to a first conversation.
When you have it, talk to us about that one job. It starts with 30 minutes on your workflows, and the point of the call is to decide whether the job is worth handing off at all.
