This spring, every software company we follow is offering an agent: one that answers email, one that books appointments, one that researches competitors overnight. The demonstrations are impressive, as demonstrations generally are.
We have worked with these models since obtaining early access to GPT-3, and have been building custom GPTs and agent workflows for some time. This experience gives us a clear view of what currently works for a small business and what remains a showcase.
What is an AI agent?
A chatbot responds; an agent acts. It can read an inbox, look up information, fill in a form, call another program, check the result and decide on the next step, several steps in a row, with less supervision.
The clearest explanation we have read is Anthropic's Building effective agents, published at the end of 2024. It distinguishes workflows, where the steps are defined in advance and the model fills them in, from agents, where the model determines the steps itself. Its main recommendation is to start with the simplest solution that works and add autonomy only when necessary.
The article was written for developers. Its advice applies even more strongly to a four-person business in a village.
Which tasks can small businesses delegate to an agent today?
Routine ones: repeatable, somewhat tedious, and inexpensive to correct if the agent makes an error.
- Sorting the inbox and drafting replies to common questions (hours, prices, directions), for a person to review and send.
- Extracting details from invoices, forms or order emails into a spreadsheet.
- First drafts of review replies, social posts and newsletters, in both languages, which someone then rewrites in their own voice.
- Research, such as summarizing what other campgrounds in the region charge, or what a new regulation states, with clickable sources.
- Moving information between tools that do not communicate with each other: the website form, the calendar, the accounting software.
None of these tasks is remarkable, but together they save hours every week, which is what a small business actually needs from technology.
Which uses are still mostly demonstrations?
An agent that runs customer service without supervision. An agent that makes purchases with the business's credit card based on its own judgement. Any process that requires clicking through a dozen websites in sequence and completing every step correctly.
These work some of the time. In a demonstration, that is sufficient. In a business, the single instance in which the agent books the wrong date for a wedding party is the one customers remember. We recommend waiting, or keeping a person to approve every action that costs money or cannot be reversed.
How should an agent be supervised?
Like a new employee. This is the most useful framework we have found.
No business would hire someone on Monday and allow them to answer every client, spend from the business account and reorganize the files without oversight. A new hire receives a few clear tasks, guidance on how the work should be done, and close review for the first few weeks, decreasing as trust is earned.
An agent requires the same: written instructions with examples of good and poor answers, a narrow initial task, and someone who reviews the output daily at first. Unlike a new employee, an agent does not learn from its mistakes unless its instructions are changed, so the review process never disappears entirely.
Who should manage the customer relationship?
A person. On this point, our position is firm.
In a small community, the relationship is the business. Residents know the owner of the hardware store by name. If a regular customer writes with a complaint and receives a polished but slightly inaccurate AI reply, the loss of trust outweighs the time saved. The agent can prepare the draft; a person should decide what is sent, particularly when something has gone wrong.
An agent should also never be presented as a person. If a chatbot answers on your site, disclose it. Customers accept it readily when informed, and react poorly when they discover it afterwards.
Why must an agent's actions be logged?
Because the log is what allows errors to be detected. This practical rule is frequently overlooked, and it is the one that prevents the most damage.
Every action an agent takes should leave a readable record: what it received, what it decided, what it sent, what it changed. Most reputable tools offer this. Enable it, then review it, daily at first and weekly afterwards.
The log reveals an agent that has been answering a question incorrectly for three weeks. It also shows what the agent does well, so that more can be delegated to it.
A note for Quebec businesses: if the agent handles customers' personal information and sends it to a service hosted outside Quebec, Law 25 applies to that transfer. Determine where your data goes before connecting anything to your inbox.
Where should a small business start this summer?
With a single task. The tools are improving month by month, and some of what is a demonstration today is likely to become usable under supervision. We nonetheless recommend proceeding gradually, since a better model does not change who is accountable for errors.
For a small business in the region, we would choose one task for the summer, most likely the inbox. Let an agent sort and draft for a month, read every draft, and record what it gets wrong. By the end of the month, the business will have learned more than any demonstration can show.
Effective AI automation begins with this kind of small test. A productive first discussion on the subject devotes as much time to what should not be delegated as to what should.