AI in the sales stack (2026)
AI is now embedded in every layer above rather than sold separately: Salesforce ships Agentforce agents alongside Sales Cloud, HubSpot ships Breeze, Gong ships AI agents inside its platform, LinkedIn ships Account IQ and Lead IQ inside Sales Navigator. The interview-relevant skill is not the product names. It is being able to say credibly where AI earns its keep and where it destroys credibility — because that judgement is the scarce thing.
What it genuinely does well
Call summarisation and CRM field population. The strongest, least contested use: the system attends, summarises accurately, drafts the follow-up, and increasingly writes structured values back into fields. It attacks the oldest problem in sales — admin crowding out selling.
Research briefs. Condensing an annual report, recent news, leadership changes and public strategy into a page before a first call. Ten minutes of preparation becomes one, and the quality floor rises for everyone.
First-draft outreach. A competent draft of an email, opener or follow-up, edited by someone who knows the buyer, is real leverage.
Forecast hygiene and deal risk. Comparing claimed forecasts against observed engagement and flagging the gaps — no economic buyer on any call, no activity for three weeks, single-threaded, close date moved twice. This is AI doing what humans do badly: noticing patterns across hundreds of deals without fatigue or favouritism.
Coaching at scale. Every call scored against a rubric, with a searchable library of how the best people handle the hard moments.
Voice from the field"Every revenue leader has AI. Almost none of them are moving the number with it."
— Eilon Reshef, Chief Product Officer and co-founder, Gong (June 2026)
That is a vendor announcing a product, so read it with the appropriate scepticism — but the observation underneath is widely shared. Adoption is near-universal; measurable revenue impact is not.
Where it embarrasses people
Fabricated personalisation. The most common and most damaging. Asked to personalise, a model produces a plausible reference — congratulating someone on a promotion they did not get, citing a paper they did not write, noting an expansion that never happened. The recipient does not conclude the tool was wrong; they conclude you are careless or dishonest, and in a small UK vertical like health tech they mention it. Rule: never send a factual claim about a person or company you have not verified at source.
Generic sequences at scale. Everyone got the same drafting tools, outreach converged, and buyers adapted by ignoring the genre — cold email reply rates have declined markedly as inboxes filled with machine-written messages. The infrastructure hardened too: Google and Yahoo's bulk-sender requirements (SPF, DKIM and DMARC authentication, one-click unsubscribe, spam complaints below a low threshold) became enforced practice, with Microsoft following for Outlook-family addresses. Blasting AI-written volume from a poorly configured domain now gets mail rejected outright and burns the domain your colleagues send from. The lesson is not "don't use AI" — it is that volume is no longer the lever it was, and relevance is.
Over-trusting scores. Deal scores, lead scores and intent signals are probabilistic summaries of historical patterns. As a second opinion they are valuable: "the system thinks this is at risk — is it right?" As a verdict they cause two failures: deprioritising a real deal because the score is low, and forecasting one as safe because the score is high while the economic buyer has not appeared in six weeks. A score is not evidence. Exit criteria are.
Unedited AI in front of the customer. Anything a model writes about your product may be wrong about your product. In regulated markets that is not embarrassing but a compliance issue — a claim about clinical performance, regulatory status or data handling that your product does not support is not a typo.
Ethics and etiquette
No settled norms exist yet, but a defensible position does. Assistance is fine; deception is not — drafting or tightening a message is ordinary tool use and nobody expects a disclosure label on a well-written email. Do not simulate an observation you did not make: "I read your Q3 results and noticed…" must mean you read them. Apply the defensibility test: if they replied "how did you know that?", could you answer without embarrassment? If not, cut the claim. Mind what you paste in — prospect names, notes and transcripts are personal data, and putting them into a consumer tool outside your employer's approved list can be a genuine breach, so "does the company have an AI policy?" is a fair day-one question. And own the output: "the AI wrote it" is not a defence available to you.
Field noteThe most frequent modern etiquette failure is the AI note-taker joining a meeting the customer did not expect to be recorded — often a bot that silently appears in the participant list. In a sensitive conversation (an NHS information-governance discussion, a commercial escalation, anything touching a customer's own confidentiality obligations) that discovery can cost you the room. The professional pattern is boring and effective: mention it in the invite, restate it in the first ten seconds, offer to turn it off, and turn it off instantly if anyone hesitates.