Where should AI sit in a sales team?

We've spent years giving sellers more technology, more data and more ways to reach people. Somehow we've also given them more admin, more reporting and more work to do before they can have a conversation. Now AI arrives and the first instinct in some businesses is to automate the conversations as well. I think that's getting the opportunity backwards.

Look at the work surrounding a customer meeting. There's research, previous notes, pricing, stakeholder information, the CRM update, the follow-up and possibly a proposal. Some of it matters enormously. But how much of it needs somebody doing it manually, every single time? That's the question I'd ask before buying another tool.

Take a simple illustration. Four meetings requiring 30 minutes of preparation each add up to two hours. If a properly checked workflow reduced that to five minutes per meeting, you'd potentially return 100 minutes. Those are hypothetical timings, not a promised saving. The point is to measure the work and find out what is genuinely possible.

AI should take the admin off the seller and leave the call to them.

My test is straightforward: what happens if the output is wrong, and who carries the consequences? A draft meeting brief can be checked before a call. An inaccurate claim about your product sent to a prospect is a different matter entirely. Both might involve AI. They should not have the same level of autonomy.

That distinction matters because the buyer sees your company's name, not the clever workflow behind the message. If your automated outreach gets their role wrong or invents a reason to speak, you've damaged trust before a seller has even opened their mouth.

I'd begin with the repetitive work around the sale. Research and preparation, meeting notes, finding approved case studies and spotting patterns across conversations are all worth investigating. But check the inputs, permissions and quality before you assume a tool can do them safely.

1. 01Account and contact research. Pulling together what a company does, what has changed recently and who is involved, from sources your team trusts, so the seller starts with context already gathered.
2. 02Meeting preparation. A short brief built from CRM history, previous notes, open opportunities and your own product knowledge, ready before the call starts instead of assembled in a five-minute scramble.
3. 03Notes and next steps. Capturing what was agreed and writing it to the right fields, with the seller checking it before it is saved, so the record reflects the conversation while it is still fresh.
4. 04Finding approved knowledge. Answering 'what is our position on this?' or 'where is the latest case material?' from your approved playbooks and documents, instead of a scramble through old folders and chat threads.
5. 05Patterns across conversations. Reviewing far more calls than a manager could ever hear, then pointing to the themes, risks and coaching moments worth a person's attention.

None of those improvements requires AI to pretend to be the seller. The person still needs to understand the buyer, decide what matters and conduct the conversation. The technology should help them turn up better prepared.

There's a reason I keep coming back to human capability. Buyers have questions they haven't fully worked out themselves. They change their minds. They go quiet when something doesn't add up. A good seller notices, asks and listens. Automating the exchange doesn't automatically reproduce the judgement.

And consider the commercial consequences of everyone automating outreach. When producing a thousand plausible messages becomes cheap, producing another thousand stops being an advantage. A relevant conversation with someone who has done their homework becomes more valuable.

I'd rather see AI help a seller prepare a useful point of view than help them send a generic one to more people.

Please don't make the first step another login. Most teams already work across a CRM, inbox, calendar and document system. Look at what those tools can do, which licences are in place and what IT actually permits before you add anything.

Then get specific. Where does the information come from? Is it approved and current? Who can access it? What does the output need to contain? Who checks it? The answers are less glamorous than a product demonstration, but they're what make a workflow usable.

You may discover you need a specialist product. Equally, you may find the capability already exists inside something you're paying for. Defining the work first makes that decision much easier.

Measure the task before you change it. Then look at the time saved after checking and corrections, whether people actually use the new process and whether the work meets an agreed standard. A dashboard full of AI activity tells you very little about any of those things.

Finally, follow the capacity. If you give a seller five hours back, where do those hours go? More conversations? Better preparation? More practice and coaching? Time returned is valuable, but it isn't automatically revenue. I'd want evidence of what changed.