AI for LinkedIn selling: what to automate, what to read, what to keep

Most AI features sold to salespeople write things. Openers, follow-ups, comments, whole sequences. It is the easiest thing to demo and the least valuable thing on offer, because the bottleneck in social selling was never producing sentences. It is knowing which of your two hundred conversations deserves the next twenty minutes.

This is a split that holds up in daily use: what a model should decide, what it should only read, and what has to stay yours.

Start by separating readings from facts

Before deciding what to hand over, separate the two kinds of thing a sales tool knows.

A fact is recorded. You sent a message on the 14th. They accepted your invitation. You wrote a note. A tag exists. Nothing needs interpreting, and asking a model about it is both wasteful and risky — it can be wrong about something that was simply written down.

A reading is interpreted. What their reply means. How well this profile fits your ideal customer. Whether this thread is waiting on you. Those genuinely need judgement, and a model is good at them.

Most bad AI features come from mixing the two: asking a model to work out whose turn it is in a conversation when the last message’s sender is right there in the data. The rule is worth writing on the wall — if it is recorded, read the record. Only ask a model about things that are actually ambiguous.

What a model should decide for you

Four readings earn their place, because each one saves a decision you would otherwise make sixty times a week.

What a reply means. Not positive or negative — what the person is doing. Asking about the offer, pushing you to later, pointing you at a colleague, being politely done. Plus what they want alongside it: a call, a document, an answer to an objection. Those two are different questions, and forcing them into one is the most common failure in the category: “no to the call, send me a deck instead” reads as a refusal if you make the model pick one label. Reply intent detection goes into this in detail.

Whether anything is waiting on you. The only sort order that beats recency in an inbox. A thread whose next move is theirs steps aside; the ones where you owe an answer rise. Inbox triage covers how that works and why the question has to be phrased carefully.

Who fits your offer. Written as a paragraph about your ideal customer rather than configured as a scoring matrix, and read against every contact and every new lead. The output should be a level, not a number — more on that below.

Whether your draft has an obvious defect. An unfilled variable, the wrong language, a contradiction with the thread, a question of theirs left open. Note what is not on that list: anything about style.

What a model should read but not decide

Two things are worth letting a model summarise while the decision stays yours.

The state of a relationship. Forty messages into three sentences, plus a suggested next step. Useful before you open a thread you have not touched in a month. Not something to act on blindly, because the reason you are writing to this person today is usually not in the data.

A reply draft. The most useful writing AI in a sales inbox is the one that drafts your next reply from the thread and your own tone of voice — and the most important thing about it is that it lands in the composer greyed out, as a draft you edit, not as a message. Nothing is sent until you press send.

Worth insisting on one detail: a draft has to know how old the thread is. A conversation that went quiet three months ago needs a fresh opening, not “just following up on my last message”. Date every turn you give a model, or it will write a follow-up to a dead thread.

What has to stay yours

Why now. The reason you are writing to this person today — their post, their new role, a customer you share — is the only part of the message a prospect cannot get from anyone else. It is also the part a model has no access to.

Facts outside the thread. What you said on the call, the price you quoted, the date you promised. If it is not in a note, it did not happen as far as the AI is concerned. Write the note; the next draft gets better.

The send button. Every time. Not because a model cannot write a good message, but because the cost of one bad automated message to the wrong person is out of proportion to the seconds a click saves.

Who to stop talking to. An explicit no should be honoured by you and recorded, never re-evaluated by a model that might read it as a soft maybe next quarter.

Four rules that keep this honest

These are the things that decide whether a team trusts AI features or quietly turns them off.

Levels, not percentages. A reading shown as “73%” reads as a measurement it is not. Five levels — from very low to excellent — say the same thing without the false precision. Keep the same scale everywhere: a lead badged “high” before import must badge “high” after it, or nobody will believe either number.

Mark everything a model produced. A filter, a score, a suggested stage, a campaign condition. One consistent mark, so a user can tell at a glance which parts of a screen are interpreted. A filter you know to distrust is worth more than one that quietly disagrees with you — and the moment you hide the difference, every wrong reading costs you credibility on the right ones too.

Prefer typed answers to prose. A model that returns a probability, one label from a closed set, or a level on a rubric cannot invent a field and leaves nothing to parse wrongly. A model asked for JSON containing a score and a justification can hand you a score for the wrong person and a confident sentence explaining it. Where software consumes the answer, the answer should be a type.

Give every closed set an escape hatch. A classifier with no option for “this means nothing” will invent a meaning. Measured: “thanks, have a good day” reads as a confident not interested the moment you remove the neutral option — and if your sequence branches on that, it just stopped for no reason.

What automation is still for

None of this replaces sequences. A model that reads well does not open conversations; a sequence does, at a pace that does not get your account flagged, with steps you wrote.

What changes is where the sequence branches. Instead of “did they reply, yes or no”, a step can route on what the reply meant — the document to whoever asked for one, the answer to whoever objected, silence and a reminder to whoever said later. Same automation, aimed properly.

And the sequence still does not stop just because someone replied. That is a condition you add deliberately, because “replied” and “replied something that means stop” are not the same event — which is the whole reason reading replies is worth doing at all.

The test to apply

Before switching on any AI feature in a sales tool, ask what it does when it is unsure.

If the answer is “it picks the most likely option anyway”, it will confidently mislabel your politest prospects and stop sequences that should have run. If the answer is “it says nothing and leaves the record blank”, you have something you can build a week around.

Everything else — how fluent the writing is, how many models it runs, how fast it answers — matters far less than that one behaviour.