Using AI in Sales: The Line Between Input Work and Output Work

Advice on using AI in sales tends to come in two flavours. One says AI will do most of the job and you should hand it everything. The other says AI writes generic rubbish and you should keep it away from customers. Both are describing real experiences. They are describing different tasks.

Here is the distinction that explains nearly every AI-in-sales success and nearly every horror story: AI is reliably good at input work and unreliably bad at output work. Input work is everything that goes into a sale — research, notes, summaries, preparation, finding the calls worth reviewing. Output work is the words a buyer actually reads or hears. Almost every story of AI embarrassing a salesperson is a story of it being used on the wrong side of that line.

This guide is for the rep deciding what to do with AI on Monday morning. It gives one test that sorts any task, applies it to fourteen real sales tasks, names the four input jobs worth automating first, and ends with a routine you can start this week. Where we have written about a specific use in depth, it links there rather than repeating it. Rolling AI out across a whole team is a different problem — an operating model rather than a set of habits — which we covered in AI enablement.

The test: who sees the mistake first?

AI makes mistakes. It summarises a call slightly wrong, gets a job title out of date, invents a plausible fact, writes a sentence that sounds like everyone else's. That does not make it useless. It means the question for any task is not "is AI good at this?" but "when it gets this wrong, who finds out first — me, or the buyer?"

• If you see the mistake first, the task is input work. An error costs you a minute of checking. The model's speed is pure gain, and its mistakes are caught before they matter.
• If the buyer sees the mistake first, the task is output work. An error costs you credibility with the one person whose opinion decides the deal, and you often cannot tell it is an error until they react.

Use AI freely where you will catch its mistakes. Be careful where the buyer will catch them for you.

Why output work goes wrong

There are three separate failure modes on the output side, and they are worth telling apart because they need different fixes.

The average of the genre. A language model writes by producing the most likely next words given everything it has read. Ask it for a cold email and you get a very likely cold email — which is, almost by definition, the kind buyers have already received most often. It is fluent, polite and correctly structured, and it reads like the hundred others in the inbox, because in a real sense it is their average. Buyers have been trained by volume to recognise it, often in the first line.

The register. Models default to a tone somewhere between a marketing department and a helpful assistant: enthusiastic, slightly formal, fond of "I hope this finds you well" and "I'd love to explore". Real salespeople who are good at their jobs rarely sound like that, and buyers notice the gap even when they could not name it.

The invented fact. Models fill gaps with plausible material. A congratulation on a funding round that did not happen, a role the prospect left last year, a product feature your company does not have. On the input side this is an annoyance you catch while checking. On the output side it is sent to a buyer under your name.

There is also a fourth problem that is not about quality at all. AI removed the cost of producing sales messages without removing the cost of reading them. When every seller can send ten times more, each message is worth less to the person receiving it — we worked through that arithmetic in using AI for sales prospecting. "Do more with AI" is the wrong instruction for the output side even when the output is good.

Fourteen sales tasks, sorted

Task	Side	What to do
Account research before a call	Input	Automate, but require a source and date for every fact
Summarising a call into notes	Input	Automate; check names and numbers, which transcription gets wrong
Updating the CRM after a call	Input	Automate; it is the task reps most often skip
Finding which calls to review	Input	Automate; let it pick, you listen
Preparing a deal for a review	Input	Let it summarise the history; you make the judgement
Listing likely objections for an account	Input	Automate, then practise answering them out loud
Practising a call against an AI buyer	Input	Mistakes are private, so this is input work even though it is a conversation
Generating discovery questions	Mixed	Let it list; you choose three and rephrase them in your own words
Recap email after a call	Mixed	Let it draft the factual recap; you write the opening and the ask
First draft of a proposal	Mixed	Let it structure; you verify every claim, number and commitment
Comparing yourself with a competitor for a buyer	Output	Write it yourself; an invented claim about a competitor is costly
Cold emails	Output	Write it yourself; use AI to critique and cut
LinkedIn messages and comments	Output	Write it yourself
Reading AI suggestions during a live call	Output	Mostly avoid; reading competes with listening

Two rows deserve a comment.

Practice is input work, even though it is a conversation. The test is who sees the mistake, not whether words are involved. In a rehearsal against an AI buyer, every fumble is private and free. That is why practice is one of the strongest uses of the technology and substitution — an AI holding the real conversation for you — is one of the weakest. We set out that contrast in AI SDRs, and how practice tools work in AI sales training.

Live suggestions are output work in disguise. A prompt that appears on screen mid-call is input to you, but you use it within seconds, in front of the buyer, with no time to check it — and reading competes for the attention you should be spending on listening. We looked at what those systems can and cannot see in AI guided selling.

The four input jobs to automate first

If you only change four things, change these. Each is cheap to verify, low cost when wrong, and gives back time every single day.

1. Pre-call research, with sources

Before every call, have AI pull together who the person is, what the company does, what has changed recently and what that might mean for the conversation. Require a link and a date for every fact; anything without a source is an inference and stays in your notes, never in your mouth. Ten minutes of reading a sourced brief beats an hour of tab-switching, and a wrong fact you catch here costs nothing.

2. Call notes and the CRM

Turn the transcript into notes, next steps and a CRM update immediately after the call. Check names and numbers: speech-to-text is weakest on exactly the proper nouns sales depends on. We covered what a transcript can genuinely tell you about a call in sales call transcript analysis.

3. Deal and pipeline summaries

Before a pipeline review or a deal strategy session, have AI summarise the history of each deal from emails, notes and transcripts. It does the reconstruction; you do the judgement about what it means. Do not let the summary's confident tone substitute for the judgement.

4. Objection preparation

Give AI the account, the person's role and what you sell, and ask for the objections most likely to come up. Then do the part AI cannot do for you: answer each one out loud, more than once, until the answer is yours. A list of objections you have only read is recognition. An answer you have said three times is recall, and only recall survives a live call. You can rehearse them against an AI buyer if there is nobody to practise with.

The thing AI makes harder

Here is the side effect almost nobody mentions. When AI takes over the research, the notes, the summaries and the follow-ups, the only part of the job left that is entirely human is the conversation itself. That part now carries more of the difference between a good rep and an average one, not less.

And it is getting less practice. Writing your own notes after a call used to be a small rehearsal of the deal — you had to decide what mattered. Reading an AI summary is easier, and it is not the same mental work. A rep can arrive at a call having read an excellent brief and still not have thought about the deal.

Two habits counter this:

• Close the brief and say it. Before the call, without looking, say out loud the three things that matter most about this buyer and what you want from the call. If you cannot, you have read the brief but not absorbed it.
• Practise the conversation, not just the preparation. The time AI saves on input work is best spent on the part it cannot do: being good on the call.

A week-one routine

Monday — research. For tomorrow's calls, have AI prepare a sourced brief on each account. Check every fact you intend to mention.

Tuesday — notes. After each call, generate notes and a CRM update from the transcript. Correct the names and numbers. Notice how much time it saves.

Wednesday — follow-up. Let AI draft the factual recap of what was agreed. Write the opening line and the ask yourself. Compare the two halves.

Thursday — objections. Ask for the likely objections for your next important call. Answer each out loud at least twice.

Friday — audit. Count how many things AI produced this week reached a buyer without you rewriting them. Each one is a place you crossed the line. Decide whether it was worth it.

By the end of the week you will have kept most of the time savings and very little of the risk, which is the whole point.

Common questions about using AI in sales

How should salespeople use AI? For input work — research, call notes, CRM updates, deal summaries, objection preparation and practice — where its mistakes are caught before a buyer sees them. Be much more careful with output work, the words a buyer reads or hears, where AI tends to produce generic, slightly wrong or invented material under your name.

What sales tasks should you not use AI for? Anything where the buyer will be the first to see a mistake: cold emails sent unedited, LinkedIn messages, claims about competitors, commitments in proposals, and reading AI suggestions during a live call. AI can still help with these by critiquing or structuring, but the final words should be yours.

Is it OK to send AI-written emails to prospects? It is risky. Models produce the most likely version of an email, which is the kind buyers have seen most often, in a tone few good salespeople use, and occasionally with invented facts. Use AI to draft factual recaps or critique your own writing, and write the opening and the ask yourself.

How do I stop AI from making up facts about a prospect? Require a source and a date for every fact it reports, and treat anything unsourced as a guess that stays out of your conversation. Check any fact before you mention it to the buyer, especially job titles and recent company news, which go out of date quickly.

What is the best first use of AI for a sales rep? Pre-call research with sources, followed by call notes and CRM updates. Both save time every day, both are easy to check, and mistakes in either are caught before they reach a buyer.

Does AI make sales calls easier? It makes preparing for them faster. It does not make the conversation itself easier, and by taking over research, notes and follow-ups it leaves the live conversation as a larger share of what separates good reps from average ones. Spend some of the time saved practising it.

How do I use AI to prepare for a sales call? Ask for a sourced brief on the person and the company, the likely objections for someone in that role, and three discovery questions. Then close the brief, say the key points out loud, and practise answering the objections rather than just reading them.

Should sales managers ban AI-written emails? A ban is hard to enforce and throws away the useful parts. A better rule is the input-output line: AI may draft, research, summarise and critique, but no AI-written words reach a buyer without the rep rewriting them. Then review a sample of sent messages to check.

A note on sources

This guide contains no figures for hours saved, productivity gains, reply-rate changes or adoption rates. Those numbers are everywhere in this topic and come almost entirely from vendor surveys and self-reported estimates, where respondents are selected by interest in AI and "time saved" is rarely measured against a baseline. We have chosen not to repeat any of them.

The description of how language models generate text is deliberately plain: they produce likely continuations of what they have been given, which is why an unspecific request produces typical output. The observation that transcription is weakest on proper nouns comes from our own work building voice practice, where it is a known and handled problem. The input-output line, the "who sees the mistake first" test, the task table and the routine are ours — a working rule for deciding, task by task, where AI helps a salesperson and where it quietly costs them.

For short answers to the broader questions people ask about AI and selling — including whether it will replace salespeople — see our AI in sales FAQ.