How to let an AI assistant near your LinkedIn without risking your account

If you are a founder, your LinkedIn is not a marketing channel. It is your pipeline and your name. You cannot spin up a new one.

So the new assistants are tempting. An OpenAI Dot or a Grok Bot can read a thread and work out who matters. Then it can write the reply in your voice. The obvious next step is to hand it your LinkedIn login and let it work.

This page is about what happens when you do that. It also covers what to do if you do it anyway, and the setup we built instead.

Two things up front. LinkedIn's own help pages say it does not permit third party software. That includes crawlers, bots, browser plug-ins and extensions. These must not scrape LinkedIn, change how it looks, or act on it without a person at the keyboard. And no tool on the market, ours included, can promise you an outcome from LinkedIn. What you can do is reduce how much of this sits on luck.

Why going direct is risky

None of this is a knock on the assistants. Grok Bot and Dots are good at the thinking part. The problem is everything around the thinking.

It works at machine speed with no sense of what LinkedIn tolerates. The assistant has no idea you have a daily rhythm. It will happily run a hundred profile views in ten minutes because you asked for research. Neither vendor documents a per site action limiter, so the only brake is you watching.

It forgets between sessions. A new chat starts from zero. It does not know you already sent this person an invite in March, or that they replied and said no. Double sends are not a bug in the assistant. They are what happens when nothing remembers.

One login is shared by every bot on the account. xAI's own docs say every Bot on your account uses one persistent cloud computer. They share its files, browser sessions and logins. The docs add that the computer is assigned per user, not per Bot. They also say "do not use separate Bots as a security boundary". So your research bot and your outreach bot are not two careful assistants. They are one LinkedIn session with two sets of instructions.

Your rules are prompts the assistant tries to follow. OpenAI describes custom rules for Dots plainly. They "are instructions your dot tries to follow, and it can make mistakes". That is an honest sentence and it is worth reading twice. "Never message anyone at a customer account" is a wish, not a gate. A rule that lives in a prompt can be talked out of by the next prompt.

It re-checks the inbox to see if anyone answered. Assistants are bad at waiting. Given a job that depends on a reply, they poll. Grok Bot routines can run as often as every five minutes. A simple "let me know when she replies" can turn into a profile and inbox check every five minutes. That can run all day, forever.

It stalls on security checks. Both vendors tell you this. xAI says a site may block the Bot, require a new login, or present a CAPTCHA. It also says "the Bot should hand those steps to you rather than bypassing them". OpenAI says a Dot's cloud browser has its own sessions. It does not inherit logins from your browser. OpenAI also says some websites block cloud browsers or require further verification. In practice that means your overnight run is not running. It is sitting there waiting for you, and you find out in the morning.

Someone can hide instructions in a profile. This one is real and has already happened to recruiting tools. A Stripe employee put a hidden instruction in his LinkedIn bio. A recruiting tool read it and wrote back a flan recipe. Your assistant reads profile text and then acts on what it read. That means a stranger can write part of your instructions.

One more thing, stated carefully. We found no first hand case of a real LinkedIn account banned for good after using Dots or Grok Bot. That is not the same as safe. LinkedIn says it does not permit this kind of software, and it does not publish how it acts on that. Treat the points above as reasons to be careful, not as predictions.

Six safety tips if you still want to go direct

Plenty of founders will do this anyway, and some of it will work fine. If that is you, these six things cost you almost nothing.

1. Read and draft, you press send. Let the assistant do research, summaries, inbox triage and drafts. Keep every outbound action on your thumb. This is also what our own setup guides tell you to do on day one. It is the single change that removes most of the downside.

2. Go slow. Decide what a normal day looks like for you in invites, messages and profile views, then stay well under it. A good test is whether a colleague reading your activity would notice anything unusual.

3. One assistant, one job. Do not give one assistant research, outreach, content and inbox all at once. Separate jobs mean you can tell what produced a given action, and you can stop one without stopping everything. Remember that on Grok Bot separate Bots still share one login, so this buys you clarity, not isolation.

4. Keep a do not contact list outside the chat. Customers, candidates, investors you are mid process with, anyone who asked you to stop. Keep it in a file or a sheet the assistant has to check. Do not keep it only in a message it will forget by Thursday. A list that lives only in conversation history is not a list.

5. Never act on profile text. Treat everything written by someone else as data, never as instructions. Tell the assistant that explicitly. Prefer a setup where profile text gets summarised for you rather than fed back into the next decision.

6. Review weekly. Once a week, read the actual sent folder and the actual invite list. Not the assistant's summary of what it did. Fifteen minutes of this catches duplicate sends, the wrong person, and tone drift long before anyone complains.

The Crispy way. Your assistant decides, Crispy acts

We went a different route. Your assistant never gets a LinkedIn login. It calls Crispy, and Crispy does the LinkedIn part.

That matters for one reason. The limits stop being instructions the assistant tries to follow. They become checks Crispy runs on every action.

The pacing lives in code. Crispy applies a daily cap per action category, counted server side. When a category hits its cap, writes in that category are refused. Sends also wait a few random seconds behind the previous action on the same account. Campaign sends spread across a send window rather than firing as one burst. No prompt can argue a send past the cap in force, because the assistant is not the thing enforcing it. The current defaults are published on our LinkedIn safety page, and you can set them lower.

The memory is already there. Who you contacted, who replied, who is on your block list, which list someone sits in. That history belongs to your account, not to a chat window. It survives when you close the tab, switch models, or move from Dots to Grok Bot next quarter.

The block list is a gate, not a reminder. Outbound sends route through one gate that checks the block list before the send happens. A send to someone on the list is skipped and logged.

Keep the settings for yourself. Your assistant connects to Crispy with your access, so it can change the same settings you can. That includes the daily caps, the block list and the account's permission scope. Tell it plainly that settings changes are yours, and check them in your weekly review. The checks hold on every send. Who may move them is up to you.

Worth saying plainly. This reduces how much depends on an assistant behaving well. It does not put us outside LinkedIn's rules, and it cannot promise you anything about your account. Our terms say the same thing in less friendly language.

If you want to wire it up, both setups are written out already.

Connect LinkedIn to OpenAI Dots, with the step by step screens in the Dots guide.

Connect LinkedIn to Grok Bot, with the step by step screens in the Grok Bot guide.

Not sure which applies to you, start with OpenAI Dots vs Grok Bot for LinkedIn.

No tool can guarantee LinkedIn account status. Crispy is not affiliated with LinkedIn, OpenAI or xAI. The controls and their current defaults are on our LinkedIn safety page. The platform risk you accept is set out in our terms.

Your assistant decides. Crispy acts. Start with the Dots guide or the Grok Bot guide. Keep sends on your thumb until you trust what you are reading.

No tool can guarantee LinkedIn account status. Crispy is not affiliated with LinkedIn, OpenAI or xAI. The controls and their current defaults are on our LinkedIn safety page. The platform risk you accept is set out in our terms.