AI in sales works best in the groundwork around a conversation and worst at judging people, even when those people are your own salespeople. The return figures doing the rounds almost all come from companies that sell the software.
In the sales process, AI does six things that appear on every vendor’s page: scoring leads, researching companies, drafting outreach, summarising calls, planning follow-up and adjusting forecasts. That is where the real time savings are, and it is work nobody misses.
Above that sits a layer that most content on the subject barely mentions. Call analysis of your own team touches two provisions of the AI Act, and customer data in a sales tool touches the GDPR.
Further down, this article covers where AI in sales goes wrong in practice, which tasks do lend themselves to it, and in what order to start.
What does AI do in the sales process?
In the sales process, AI does six things: scoring leads based on behaviour and company characteristics, summarising company research before a call, writing a first draft of outreach, transcribing calls and summarising them, planning and reminding you about follow-up, and enriching forecasts with deal data instead of gut feeling.
Lead scoring is the most frequently mentioned of these and also the most misunderstood. A model ranks prospects based on website behaviour, email interaction, company characteristics and the pattern of customers who signed before. The result is an order in which to call, not a judgement on who may or may not become a customer.
Forecasting works on the same principle. Instead of relying on what salespeople report about their own pipeline, the model weighs deal activity, cycle time and historical conversion per stage. If a deal stays in the same stage too long, that is a signal rather than a surprise at the end of the quarter.
The common thread is that AI takes over the work around the conversation, not the conversation itself. A salesperson who is well prepared sooner does not sell better because the model is clever, but because they come to the table better prepared.
Some of this borders on marketing and sometimes even falls under it. This article starts where the lead comes in.
How reliable are the figures on AI in sales?
The figures on AI in sales almost all come from companies that sell AI for sales. Percentages such as forty per cent more closed deals, thirty to fifty per cent higher response rates and forecasts that are twenty to thirty per cent more accurate circulate without any study design, sample or measurement period attached.
Measured effects are more modest. In a field experiment with 5,172 customer service agents at one company, Brynjolfsson, Li and Raymond found fifteen per cent more issues resolved per hour, published in the Quarterly Journal of Economics in 2025. That fifteen per cent is throughput, not time saved, and the study is about customer service rather than sales, but it remains the closest measured commercial context available.
At the broad end is a survey by Humlum and Vestergaard of 25,000 workers in eleven occupations in Denmark: 2.8 per cent of working time, about an hour a week, and that figure is self-reported.
The difference between those figures and the percentages from the tool demo is the difference between a measured effect and an expectation. So calculate with the time you spend today on quotes, CRM entry and follow-up, because that is a number you do know.
What are you not allowed to do with AI on your own sales team?
AI that you point at your own sales team falls under two provisions of the AI Act, and call analysis touches both. Conversation intelligence that records, transcribes and rates calls on quality is more than a coaching tool: it is a system that evaluates employees’ performance and behaviour.
Such a system falls under Annex III, point 4 of the Act and therefore into the high-risk category. Those obligations were due to apply on 2 August 2026 and have been moved to 2 December 2027 by the Digital Omnibus, an amendment in Regulation (EU) 2026/1744 that entered into force on 27 July 2026.
The second provision has no postponement. If the tool also analyses your salesperson’s emotional state, for example from voice, speaking pace or tone, that touches Article 5(1)(f). Inferring people’s emotions in the workplace is prohibited, with a narrow exception for medical and safety purposes, and that prohibition has applied since 2 February 2025.
The Act looks at what the system does, not at how you introduce it. Announcing call analysis as a learning tool rather than a monitoring tool does not change the classification.
What are you not allowed to do with your prospects’ data?
AI in sales runs on customer data, and that falls under the GDPR as soon as there is a person behind it. A company name is not personal data; a contact person with a job title, email address and call notes is.
Three questions determine whether you have arranged it properly. On what legal basis do you process this data, what data processing agreement do you have with the tool’s supplier, and how long are call recordings and transcripts kept?
In practice, that means no customer data in a free account and no client files in a personal environment. Systematic processing of call recordings also requires a data protection impact assessment (DPIA).
Where does AI in sales go wrong in practice?
AI in sales goes wrong in four places, and all four can be prevented.
The first is the made-up detail in outreach. A model that researches a prospect and writes an opening line about them sometimes mentions an acquisition, a job change or a product launch that never happened. This is called AI hallucination, and in a first contact with a prospect, one such error costs you the whole conversation.
The second is polluted data driving the scoring. Lead scoring learns from what is in the CRM, so a CRM in which deals are updated inconsistently produces a model that neatly reproduces past mistakes and attaches a score to them.
The third is personalisation that reads as automation. Since Google’s December 2025 core update, I have seen on client sites that content without a human contribution has dropped structurally in traffic. A search engine that recognises interchangeable text does so with considerably more effort than a recipient who gets five such emails a week.
The fourth is measuring on volume. More messages sent and more activities logged are not a result, and a team managed on those numbers optimises for noise rather than pipeline.
Which sales tasks are best suited to AI?
The sales tasks best suited to AI are the tasks between conversations. Summarising company research, writing up call notes, updating the CRM, drafting a first version of a quote and preparing a follow-up email all fall into this group.
One question separates these tasks from the rest: is a judgement made about a person that affects that person? For a summary, the answer is no. When rating a salesperson, or deciding on a candidate or a customer with consequences for that person, the answer is yes, and then AI is limited to preparation.
A third category sits in between: anything the customer sees directly. A quote, a proposal and a first email may be drafted by AI, but they do not go out without someone having read them. From my years at major banks, I bring the four-eyes principle to AI work, and this is exactly where that principle pays for itself.
Frequently asked questions about AI in sales
Besides the AI Act, other rules also apply to AI in sales, especially around email and automated decisions.
Can you send AI-written cold emails to prospects?
You can only send AI-written cold emails to prospects if the recipient gave prior consent, just as with any other commercial email. In the Netherlands, the Authority for Consumers and Markets (ACM) makes no distinction between consumers and businesses in its explanation of Article 11.7 of the Dutch Telecommunications Act. An exception applies to existing customers for similar products, provided they can unsubscribe. In every message, make clear who the sender is and offer a simple way to unsubscribe.
Can AI decide on its own whether a prospect becomes a customer?
AI may not decide without human involvement whether a prospect becomes a customer if that decision produces legal effects for someone or similarly significantly affects them. According to the European Commission, the GDPR only allows such a fully automated decision in limited cases, such as explicit consent or necessity for a contract. In those cases, the person concerned has the right to human intervention and may contest the decision. A lead score that only sets your call order does not fall under this, as long as a salesperson makes the decision.
How do you get started with AI in sales?
Getting started with AI in sales takes five steps, and the first is taking stock rather than buying. Ask the team what is already being used and for what, because in most sales teams this has long been happening out of management’s sight, a pattern known as shadow AI.
Then assess what you point at your own team and what at your prospects, because those are two different legal questions with two different answers.
Next, pick one task between conversations, measure how many hours it saves, and scale up only when that number holds. Finish by recording which data may go where and who checks what before anything goes to a customer.
AI in sales therefore clashes exactly where a judgement about a person comes into play, and works everywhere around it. If you want to set out that boundary for your team in a way of working you can show to a customer, an accountant or a works council, the AI Literacy Course covers how to record it in an internal AI policy.
