How to choose an AI assistant: a buyer's checklist

To choose an AI assistant, write down the job you need done, then run each candidate through three of your own real tasks, one of them long. Also get plain answers on long conversations, memory, your data, actions, sources, pricing and export. The one that does your work well and answers plainly beats the one with the best demo.

What job do you want an AI assistant to do?

Describe the job in one sentence: who it is for, what goes in, what comes out, and how often. "Draft replies to customer emails in our tone, about ten a day, for a person to check" is a job. "Be helpful" is not.

The job decides what matters. A writing helper needs good edits and a consistent voice. A research helper needs sources you can open. A coding helper needs to work with your files. A recurring task may need a schedule, and a team needs shared access. Write down what you will not accept, too, such as anything sent without your approval.

If the job involves doing things and not only answering, you are shopping for an agent, a different category covered in AI agents vs. chatbots.

How does it handle long conversations, and does it remember?

Today's assistants can use only a limited amount of text at once, called the context window. When a conversation outgrows it, the product has to drop older material, summarize it, or stop. Ask which one yours does and whether it tells you. Quality can also slip in a long conversation before the limit is reached, which is why keeping a long AI conversation on track takes habits.

Then ask whether it remembers across conversations. Some assistants have a memory feature and some start fresh each time. If it remembers, check that you can see what it saved, edit it, and clear it. Remembering across chats is a different thing from fitting more text into one chat, as AI memory vs. context explains.

Speed matters if you will use it dozens of times a day: long answers take longer, and speed can change with load.

What happens to your data, and can you delete it?

Ask these before you paste anything sensitive:

• How long are conversations kept?
• Are they used to improve or train models, and can you switch that off?
• Can the provider's staff read them, and when?
• Can you delete a single conversation, and can you delete your whole account?
• Does deleting remove the data everywhere, or only from your view?

The answers should be in the privacy policy, the terms and the settings page. Personal and business plans often have different terms, so read your plan's. If you cannot find an answer, treat it as unknown. A longer version of this list is in the AI assistant privacy checklist.

Can it take actions, and can you check its work?

First find out whether it only answers or can also act: send email, change files, run commands, browse. If it can act, ask whether it stops before a consequential step, whether the approval shows exactly what it will do, and whether you can see a record afterward. A good approval shows the actual recipient and message, not a vague "send email?" prompt. Why an AI agent should ask before it acts explains why that step matters.

Then look at claims. When the assistant states a fact, does it show where the fact came from, with links you can open? A citation is only a start. Open it and check that the page says what the assistant says it does. Any assistant can be confidently wrong, so the question is how easy the product makes checking. How to fact-check an AI answer walks through the habit.

What will an AI assistant cost, and what else should you check?

Pricing usually takes one of three shapes:

• Usage-based. Common with developer tools: you pay per token, the unit of text going in and out. Because the earlier conversation is typically sent again with each message, long conversations cost more per turn. What a token is covers the units.
• Subscription. A fixed monthly price for one person, usually with usage limits.
• Per seat. A monthly price for each person on a team, which grows with every person you add.

Surprises tend to hide in the same places: limits that pause you mid-task, higher tiers for stronger models or longer inputs, and extra charges for search or tools. Price your test tasks, then multiply by how often you will do them. At ten drafts a day, the email job above is about two hundred a month.

Three more things to check:

• Integrations. Does it connect to the tools you already use, and what can each connection read or change?
• Export. Can you download your conversations and anything it made in a form you can reuse, so you can leave?
• Admin controls. For a team: roles, control over who sees whose chats, and a way to remove someone's access.

How do you test an AI assistant with your own tasks?

Pick three real tasks you finished recently, so you already know what a good result looks like, and run the same prompts on every candidate. For the customer-email job, that could be:

• A long one. Work through twenty past support emails in one conversation, over an hour or more. In the first message, set a rule such as "replies under 80 words, British spelling, no exclamation marks", and see whether the twentieth reply still follows it.
• One that hinges on a fact you know. Ask about a detail in your own field that is easy to get wrong, such as the date a rule changed. See whether it shows a source, admits doubt, or states a wrong detail with confidence.
• One with your own material. Give it a real customer email, with names removed, and your tone notes. Count the edits the reply needs before you would send it.

Then score each candidate on edits needed, whether the rule held, whether sources held up when opened, and time taken. Finish with the exit: delete one conversation and export another. Use non-sensitive material until you have your answers on data.

What are the red flags when choosing an AI assistant?

• Vague claims. "Smarter" or "human-like" with nothing you can try.
• No deletion option, or no way to remove your data.
• No record of actions, or an agent that acts without asking.
• Sources that do not exist, or that do not say what is claimed.
• A price you cannot work out before you commit.
• No export, so your work is stuck.
• A request for wide access on day one, such as all your email or files.

Where NAMELESS stands

NAMELESS is an AI model that is still in training and not yet available, so there is nothing for you to test today. It is built on our own architecture, with unlimited context and no rereading of the conversation, extremely fast replies, stronger logic and reasoning, and a cost of pennies compared with today's leading assistants. Until you can run the tests above on it, treat all of that as our claim. There is no launch date yet. You can join the waitlist to follow progress.