AI assistants

The useful question is not which assistant is best. The useful question is what you need to produce: a draft, a challenge to your argument, an analysis of a question, or a comparison of two candidate replies. ChatGPT, Claude, Gemini, and Microsoft Copilot all sit in the same class of chat assistants. They take a prompt, optional pasted context, and return text. Ranking them as if one winner exists for every workplace task is a different job from the one in front of you, and it delays the email you still have to send.

Name the artifact first. A client update, a one-page brief, a list of risks in a proposal, a rewrite that keeps every claim tied to a source you already have. Once the artifact is named, the prompt writes itself: what to produce, for whom, in what tone, and what must remain empty if the sources do not support it.

If you catch yourself typing "which AI is best for work," stop and restate the actual deliverable. The ranking question feels like progress. It is not. It produces a debate you cannot verify and a draft you still do not have.

Assistants are a starting point for most workplace AI use. They are good at brainstorming, pressure-testing an idea, unpacking a fuzzy question, and producing a first pass of text. They are not a filing system, a connector layer, or a voice capture tool. When the same thread has to hold a long project, move that work into AI Projects. When a repeated instruction should travel with you, put it in AI Skills rather than retyping it. When the input is spoken notes, dictation AIs get the words onto the page first. When the assistant needs live tools or files instead of a paste, that is MCP. This page stays on the chat itself.

A quality draft cites the prompt you used and the sources you pasted. If you have no source for a claim, the claim stays empty. Do not ask the assistant to "add citations" after the fact and hope it invents plausible ones. That is the failure mode that turns a useful first pass into something you cannot send.

Paste the material before you ask for the draft. Meeting notes, the three paragraphs from the research pack, the last email in the thread, the policy excerpt, the numbers from the spreadsheet you already checked. Then state the rule in the prompt: use only what is in this paste; if a point is not supported, leave it out or mark it as missing.

That rule is the difference between a draft you can stand behind and a fluent paragraph that sounds finished. Assistants complete patterns. If the pattern of a good email includes a statistic or a named source, they will often supply one. Your job is to make the allowed evidence visible and to forbid the rest.

Keep the prompt next to the paste. You will need both when you review, and you will need both if you later compare two tools on the same job. A comparison that does not freeze the prompt and the sources is not a comparison. It is two different conversations.

Use the chat to produce a first pass, or to attack a pass you already have.

For a draft, say what the reader must walk away knowing, who they are, and which pasted sources are in play. Ask for a structure you can scan: short paragraphs, claims tied to the paste, a blank where evidence is missing. Then read the result the way a colleague would if your name were on it. Check every factual sentence against the paste. If the model cited something you did not provide, delete it. If it softened a risk you asked it to keep, put the risk back.

For a challenge, paste your own draft and ask for the opposite of polish. Where is the argument thin? Which sentence would a skeptical reader reject? What did you assume that the sources do not say? A challenge is useful when you already know the content and need pressure, not when you are still collecting facts.

Do not treat the chat pane as sent. The thread is a workbench. The email, the ticket, the slide, the message in the real channel is the artifact. Copying a reply straight into send is how invented citations and off-tone lines leave the building with your name on them.

One pass is usually enough to see whether the assistant understood the job. If the structure is wrong, fix the prompt and the paste rather than arguing in follow-ups. Follow-ups that say "make it better" without naming what is wrong produce longer text, not a better artifact.

Compare tools when you need to see how two competent drafts handle the same constraint, not when you want a winner to adopt forever. Freeze the prompt. Freeze the paste. Run the same job in a second assistant in the same class: ChatGPT, Claude, Gemini, or Microsoft Copilot. Read both as a reviewer, not as a shopper.

Look for disagreements that matter. One reply cites only the paste; the other adds a confident extra fact. One keeps the missing-evidence gap visible; the other papers it over. One matches the tone of the thread you are actually in; the other sounds like a template. Those differences teach you how to brief the next prompt. They do not license a ranking of vendors.

If both replies agree and both stay inside the paste, pick the one you can edit fastest and move on. Comparison is a judgment exercise. It is expensive in attention. Use it when the send is high-stakes or when you are still learning how these tools fail. Skip it when the job is a routine rewrite you can check in a minute.

The comparison artifact, when you need one, is a short note to yourself: the prompt, the sources used, where the two replies diverged, and which lines you kept. That note is the quality outcome. A public ranking is not.

The assistant does not own the send. You do. After the draft or the comparison, you are left with text that must still pass a human check: Does every claim have a source in the paste? Does the prompt you used still describe the job? Is anything empty that should stay empty? Would you put this in the thread without a disclaimer?

Here is a small picture of the whole loop. You need to email a partner a status note based on two internal memos and yesterday's call notes. You name the job: a six-sentence update, no new commitments, every number taken from the memos. You paste those three documents. You ask for a draft that cites them in plain language and leaves a blank if a figure is not in the paste. The first reply invents a completion date that appears in neither memo. You delete that sentence. You run the same prompt and paste in a second assistant because the tone still feels off. One version hedges every line; the other is direct and stays inside the sources. You take the direct version, restore one caveat from the call notes the model skipped, and you send the email. The ranking of which vendor is "best for work" never entered the task.

That is the standard. Pick the job. Paste the sources. Draft or challenge. Compare only if you need the disagreement. Keep the prompt and the sources with the output. Leave gaps empty. You still send.