Which AI agents should a small CPG brand use, and how should it choose?

Choose an AI agent by the job, not by the product.

An AI agent is a piece of AI that does a task for you on its own, without you typing each step. Ask four questions about the job. Does it have to run on a schedule without you? Does it touch money or customers? Does it need your own data connected? Who gets told when it fails?

Your answers send you to one of three outcomes: use a ready-made agent, build a small scheduled job, or do not automate it.

What is an AI agent doing that a chat window is not?

A chat window waits. You type, it answers, it stops. An agent acts. It can open your files, update a spreadsheet, send a message, and it can do all of that at 5:00 AM while you sleep.

That is the whole value and the whole risk. Nobody is watching when an agent acts, so you choose one the way you choose someone you will not be supervising.

The decision tree

Walk the job through four questions, in order. Stop at the first one that gives you an answer.

1. Does the job have to run on a schedule, without you? If no, use a chat window. Paste in the work, read the answer, done. Most jobs end here. If yes, go on.
2. Does it touch money or customers? That means anything that emails a customer, changes a price, moves an order or pays an invoice. If yes, a person must approve every result before it leaves the building. If you cannot put an approval step in, do not automate it. If no, go on.
3. Does it need your own data connected? Your sales numbers, your inventory, your retailer portals. If no, a ready-made agent is probably enough. If yes, go on.
4. Who gets told when it fails? If the honest answer is "nobody" or "I would notice eventually", do not automate it yet. If a named person gets a message the same day, build a small scheduled job.

That gives you three outcomes.

Use a ready-made agent. The job is generic, needs none of your data, and a failure costs you an annoying morning, not a customer.

Build a small scheduled job. A scheduled job is a short set of written instructions that a computer runs at a set time, say every day at 6:00 AM. This is the right answer when the job needs your data and a named person hears about every failure.

Do not automate it. The job touches money or customers with no approval step, or nobody hears when it breaks.

What should an agent prove before you trust it with a job?

Three things, in this order.

First, it gets a sample right where you already know the answer. Give it last month's task, with the result you got by hand, and compare.

Second, it says so when it is unsure. An agent that always sounds confident will be confidently wrong on a Tuesday.

Third, it tells you when it breaks. This matters most. A job that fails silently is worse than no job. A green check can also lie. I once had 121 automated tests pass on a system that was producing junk. (An automated test is a small check a computer runs on its own to confirm the system works.) The tests asked whether the machinery ran, not whether the result was any good. I wrote that up in Green tests measure the wrong thing.

Treat AI like a hire covers the trial period for any agent. AI that audits the AI covers having a second AI check the first one's work.

When is a ready-made agent enough and when do you build?

A ready-made agent is enough when the job looks the same at every brand: summarizing a call, drafting a first reply, tidying a list.

Build when the answer depends on your numbers: your deductions, your reorder points, your retailer rules. A ready-made agent can only see what its maker connected. A small scheduled job sees what you point it at, runs when you tell it to, and reports to whoever you name. You describe the job in plain English, and Claude, the AI assistant I use, writes the instructions. If you would rather have that done for you, the page for it is https://nordicnavigators.com/workflow-build.

What are the red flags in an agent vendor's pitch?

• The demo runs on their data, never yours.
• There is no answer to "what happens when it fails, and who is told?"
• It is "fully autonomous", meaning nobody checks its work, and that includes messages to your customers.
• Nobody can tell you how to undo what it did.
• Success is counted in tasks completed, not in results you can check.
• It will not say which of your data it needs or where that data goes.

Grade the pitch in thirty seconds

In the older days you would read the vendor's pitch twice, book a call, and ask your questions one at a time, writing the answers down. Then you would ask for two customer references, phone them, and compare it all against what the next vendor told you. That is a week of evenings. And a week later you have your answer, though you still would not be sure you had asked the right things.

Today, give this to Claude and you have your answer in thirty seconds. (A prompt is just the instructions you type in; you can paste this one as it is.) Put the vendor's pitch, whether it is an email, a web page or a deck's text, right under the instructions.

You are grading a sales pitch for an AI agent on behalf of a small food and beverage brand owner who is not technical. The pitch is pasted below these instructions. Do not be polite. Use plain English, and explain any technical word in the same sentence you use it.

Score the pitch pass, fail or unclear on each of these eight lines. For each one, quote the sentence from the pitch that is your evidence, or write "not mentioned".
1. It names the exact job it does, in one sentence.
2. It says whether the job runs by itself on a schedule or only when I ask.
3. It says which of my data it needs, and where that data goes.
4. It says what happens when it fails, and who is told.
5. It keeps a person's approval in front of anything sent to a customer or anything involving money.
6. It says how I undo what it did.
7. It measures success by results I can check, not by the number of tasks done.
8. It shows a demo on my kind of business, not only on its own data.

Then give one verdict: USE IT, BUILD MY OWN SMALL SCHEDULED JOB, or DO NOT AUTOMATE THIS. Give two sentences of reasoning. Then list the questions I should email the vendor before I pay anything, most important first.

The pitch:

Start with the pitch already sitting in your inbox. If the vendor will not answer those questions in writing, you have your answer.