Beware the AI Sales Trap

Every sales and revenue leader I’ve spoken with in the past 18 months has been AI-shopping. They are all looking for that killer app to 10x their sales efforts, and yield some shiny stats for the board on the ROI of their AI plans.

The core idea is hard to resist for any sales leader who has been neck-deep in spreadsheets for 10+ years, juggling pipelines and targets.

Ideas to automate outreach, meaning fewer SDRs needed, ideas to eliminate inconsistency and build that predictable pipeline, to book all your meetings automatically, to personalise messaging at scale for maximum relevancy. Underpinning it all is that dream of multi-million-dollar ARR with no salespeople.

In theory and practice, AI automation in B2B sales does have huge potential to grow the bottom line. With greater efficiency meaning more sales, and more time to do the stuff that works.

Except in the many cases where AI transformation in sales departments has been pretty lacklustre. Project failures even spawned the nomenclature “AI Boomerang” to explain the re-hiring of humans where AI replacements fell flat, or where vibe-coded SaaS replacements ended up more of a time sink than a cost saving (see: the legacy Build vs. Buy arguments SaaS operators hinged their sales on, when cloud subscription offerings first pushed up against on-premise owned software in the ’10s).

The problem, if we can call it that, is that AI can comfortably automate 99% of your sales process.

As a quick example, using tools like Clay, one can automate lead research and enrich leads, and then push them directly into CRMs with tasks attached. One can synthesise and systematise successful historical outreach, turn that into a schema to plug into LLM word generators like Claude Sonnet, and push out personalisation via an email sequencer like YesWare.

One can prompt an AI chatbot to conduct discovery via an agent chat API, and score leads depending on responses. Schedule and automate follow-ups so a nudge is never missed. Prepare customised pitches using Agents based on theses from commonly seen pain points. Automatically generate and send a proposal based on AI notetaker records. And so on.

Machines have basically solved all of it, which means there is a real temptation to let machines do all of it.

But as the adage goes – just because you can, doesn’t mean you should.

I would argue, dear reader, that the challenge is not to automate more, but to automate the bits machines can do better than people and ruthlessly protect everything else – to free people up for the 1% of sales that only they can do.

The companies actually seeing returns on AI initiatives are using it for the repetitive, data-intensive grunt work that’s stealing time from the stuff that actually matters. They’re protecting the 1% of the sales process where human judgement, relationship building, and deep customer understanding are the main levers that close deals.

I’ve seen the impact of this first-hand, working with over twenty-five APAC-based companies across insurance, finance, hospitality, law, and education, where ROI was the #1 and only metric to judge an AI’s impact.

The 99% Problem

99% of your sales effort is wasted.

Here’s what every* cold outreach operation looks like: a rep sends 100 prospecting emails to an identified ICP, they get 3-5 replies, and have one actual worthwhile conversation.

Put another way, those other ninety-nine contacts (multiplied by the touchpoints in your sequence) are a waste of time.

(* “but our sales team convert better than 1%!” I hear you say. Perhaps, but even a 10% conversion rate is still 90% wasted effort ;))

Humans are the ones unfortunately responsible for the waste, and most serious sales operators battle through repitition to get it done, because the effort-to-reward ratio of sales for the past 30+ years almost seems like a natural law.

The reality, sans AI, is that most of your best salespeople are spending hours researching, writing personalised emails that end up looking templated anyway, waiting for responses that never come, scheduling follow-ups at inconsistent times, and chasing down conversations that went silent three months ago. It’s data-intensive work. Low-judgement work. Repetitive work.

And humans are terrible at it!

Not because your salespeople are bad, but because it’s work meant for robots. Pattern matching at scale. Remembering the 6-month budget renewal. Sending the follow-up at 10 AM on Thursday instead of Tuesday at 4 PM. The stuff that closes deals when done right, and loses deals when done wrong. It’s not because of poor execution or lack of care, but because humans simply forget.

This is where we talk about systems, not heroics, because it’s where the real cost lives; in systematic gaps that need top operators to be on top of their game to fill, all the time.

Most of a sales leader’s job before AI automation was to get their sales team to behave more like robots than humans – to follow the process every time, or else. To simply tolerate the 99% time-sink because it was the only way. That definitely has changed.

What AI Should Actually Do for Sales

Let me be clear: I’m saying deploy AI where it has a competitive advantage and protect the places where you do.

Automate the volume and waste

Lead scoring and routing is an easy win. AI can crunch firmographics, engagement signals, job changes, and company news to flag the 1% of leads that are actually market-fit and ready to spend time on. This top of funnel research should absolutely be automated by AI. It works here because an AI system doesn’t need to understand or engage with the person, it needs only to recognise the pattern.

Data enrichment and research, similarly. The grind work. Company intelligence, contact and stakeholder information, technographic data, plans and strategies – the stuff that was sucking your team’s time instead of actually engaging with real people. Let AI do it.

Outreach personalisation and sequencing. This is a more tricky and contentious one (see: LinkedIn posts highlighting failed automated outreach and empty {merge_field} tags in messages), but it works when executed thoughtfully. AI can research a prospect in seconds, write contextual outreach, adjust the sequence based on engagement, and handle multi-channel follow-ups without ever missing a timing window. The key: this is volume-based, rule-driven work, and so your rules need to be good and tested in the field.

— As an aside: there’s a Rubicon here for any sales or organisational leader, in whether enabling AI to converse directly with your customer base is a wise idea. When you give any LLM the ability and means to speak to your customers without human oversight, you are inviting risk. It’s no use after-the-fact blaming the OpenAI integration for saying something dumb, or talking out of scope, or generally being unhelpful. Your company name is still on the communication.

If you want an AI to engage with your clients, you need strong agent harnessing. Otherwise you end up like Chipotle, whose support chatbot went viral for happily writing Python on request, or the car dealership whose website bot agreed to sell a car for a dollar. Guardrails need to be strong – to protect both the customer experience and your brand’s perception. This is especially true in sales where sensitive or impactful advice is discussed early in the process; industries like Healthcare, Insurance, sometimes Hospitality.

You cannot and should not tell an AI to blindly “write an outreach email to this person x and say y” because ultimately AI writing is a sum of the average. Without guidance, even the best LLM will spit out a three-paragraph Hubspot-esque template email with an inane greeting, a surface-level personalisation, a “we do this” line, and a signoff asking for a meeting. We’ve all seen, received and likely sent these exact type of outreach emails.

But if you can test 1,000 actual outreaches, pull out what works (e.g. strong Challenger lines that trigger conversation), and then systemise it? Much better, and plausible at scale. A person reading an email and being influenced by its tone is a human interaction. Triggering that email at exactly the right moment because the algorithm understands the pattern is what AI does better.

Initial qualification and templated discovery. Automating these works because they follow a structured, repeatable framework. “What’s your current solution? How many seats do you have? When does your contract renew?” AI can ask these. People can provide better answers when there’s no pressure, and we can feed this back into a virtuous loop with the ongoing lead scoring and discovery frameworks like MEDDIC/MEDPICC

The commonalities underlying here are data-intensive, pattern-based, and low-judgement tasks. In these situations, an AI doesn’t really need to understand the person, and it doesn’t need wider context than you have given (which has an impact on tokens/cost of the output), it only has to recognise and surface the pattern.

AIs are genuinely better at this than any human can be unless you happen to have some hyper-focused quant already on staff.

The Trap: Automating the Human Work

This is where I see teams crash. Trying to replicate a virtual salesperson. This is not the way.

Where it goes wrong, often, is in trying to automate the work that shouldn’t be automated, usually in the name of efficiency.

Automating complex discovery: instead of having a rep dig into the customer’s situation and problem – asking follow-up questions, tell me more questions, connecting the dots between stated needs and unstated concerns – teams deploy AI to run templated and impersonal discovery sequences. Unfortunately, the customer usually feels it. They know they’re talking to an automation (as a side note, your bot should indeed confirm it is a bot when asked.. don’t hide it).

Automating relationship building: especially in Enterprise deals, where trust is a big component of the actual product. If a customer gets an algorithm handling them or sending out LinkedIn requests en-masse, where there’s no actual person thinking about their specific situation and leveraging their expertise as a viewpoint, the deal loses energy. Good deals need momentum, momentum needs clear steps and trust. Trust needs a human.

Automating objection handling: most sales orgs have an ‘approved’ response to objections like “I don’t have budget”, but such generalised objections are surface-level. Often, the real objection is the political risk or internal capital to get something done inside the customer’s organisation. Most enterprise deals die because the buyer is afraid of getting the blame for a non-solution. An objection is mostly a sign you’re not aligned, and a misunderstanding needs to be uncovered and reconciled. Linking up the disparate signals and personalities across any deal and reflecting where the blocker is – AI has a hard time with that.

Automating closing and deal strategy: a good partnership, especially in B2B sales, is based on ongoing success and achievement of defined, important metrics. A good close comes as a result of good, ongoing discovery and an understanding of the customer. The eventual deal being agreed shouldn’t be materialised by AI, because it’s a result of the context of those conversations happening over time. As well, to get most significant deals over the line, it’s often a case of making sure every important person (in sales parlance; stakeholder) on the same page, pulling in the same direction to that agreed beneficial outcome. If you’re relying on AI to craft a tailored proposal as your last step, you’re likely missing those cues along the way which – I would argue – is a big part of a salesperson’s job, gathering those clues to plot the best next steps forward.

Being in the room: while The Challenger Sale proudly declared the relationship sale dead (which, when you think about it, is itself a classic Challenger line), most experienced salespeople will tell you that having a good personal relationship with your counterparty on the other side of the deal certainly makes any landing smoother. Increasing wariness over AI impersonalisation and generated content gives the edge to any salesperson who has met in person. As the adage goes, and still holds: all things being equal, people do business with their friends, so make things equal and then make friends. Being in the room seems to be more important than ever. Perhaps in the near future we can enlist humanoid robots to talk to prospects at conferences. Could work.

The near future – MCP sales?

As of writing (Q4 2026), more companies are making their products buyable inside Claude, Gemini and ChatGPT using Model Context Protocol (MCP) integrations. For now, the first movers are B2C – trading platforms, retailers, hospitality providers – but there’s no obvious reason B2B companies can’t do the same.

AI is already summarising your product’s strengths and weaknesses and ranking options for buyers. It’s why smart marketers are spending their time seeding the right positioning in both their primary sources and high-authority third-party sources. We talked about those strategies here, if you’d like to read on the topic.

Your product is already being pitched on your behalf by an LLM agent tuned to the buyer’s preferences. No salesperson is in that conversation. So for simpler, transactional purchases, it’s worth looking seriously at building an MCP server that lets buyers go straight from research to purchase without a human stepping in.

A self-test heuristic

To give some simple practical advice, you can ask yourself: does it need judgement, or some context that hasn’t been written down, or trust? If a task requires none of these things, automate them. If it requires one or more, protect it, and have your people razor sharp on handling it.

For an insurance partner in APAC, this focused strategy on the 1% yielded a 4x return on spend via new sales generated, and 15% increased proposal to close ratio.

So here’s the grand irony – for the past 30 years or more, the best sales leaders have spent their time trying to turn people into machines. Now the machines are here, perhaps we can liberate those uniquely human qualities of trust, understanding and empathy to achieve our quarterly targets. And the game goes on.

Have a think about your own 99/1 split. What’s the 1% that is key, non-automatable stuff? If in doubt, talk to us and we’ll figure it out together.

Header photo Ann H @ Pexels: https://www.pexels.com/photo/close-up-shot-of-a-wooden-robot-toy-on-a-black-surface-4102557/