There is a strange contradiction in the way sales teams are using AI. We have access to technology capable of analysing huge amounts of information, challenging our assumptions and helping us think more clearly. Yet much of the conversation still revolves around writing better prospecting emails.
That feels like an extraordinary waste. I rarely look at AI and think "how can this save me five minutes writing an email?" I am much more interested in a different question: how can AI help me make better commercial decisions? Interrogate deals more effectively. Prepare better for pipeline reviews. Understand accounts more deeply. Identify weaknesses in bids before the customer does. Help sales managers become better coaches.
For the best part of a year, the way I got AI to do any of that was prompts. Long, carefully engineered instructions. "Act as a Chief Revenue Officer." "Do not agree with me." "Score this from 1 to 10." I built a small library of them and leaned on it constantly.
I don't write them any more. Not because the thinking behind them stopped mattering — it's the same five questions I was asking a year ago — but because the mechanism underneath them changed.
What Actually Changed
The prompt era had a cost nobody talked about much: the compiling. Before you could ask AI to challenge your assumptions on a deal, you had to go and assemble the deal yourself. Pull the CRM notes. Copy the meeting notes. Dig out the emails. Find the tender documents. Paste all of it into one enormous block of text, then paste a carefully worded prompt on top.
That's forty minutes of admin before the thinking even starts. Most people did it once, got a genuinely good result, and quietly stopped — not because the prompt was wrong, but because gathering the input was too much friction to repeat every week. A prompt is only as useful as your willingness to keep feeding it.
The new way removes the compiling. AI connected directly to the CRM, to call transcripts, to email — can go and get its own evidence. I don't write "here is everything I know about this account, now act as a strategic director" any more. I ask "what's the state of this account" and it goes and pulls the pipeline history, the notes, the last call, itself.
Old way
New way
The five things I still ask it to do every week are identical to a year ago. What's gone is the gap between having the question and getting it answered. Below, each one shows the old prompt I used to write, and what I actually type now.
01
Interrogating a Deal
Salespeople are naturally optimistic. That's usually a strength — you need it to operate somewhere your outcomes are always uncertain. But optimism becomes dangerous the moment it enters the forecast. A deal can feel good, the customer can sound positive, and none of that necessarily means the opportunity closes. So one of the most useful ways I use AI is to deliberately introduce scepticism.
Old way — the prompt
"Act as an experienced Chief Revenue Officer reviewing this opportunity. Your job is not to agree with my assumptions — challenge them. Identify the strongest evidence this deal will close, the weakest assumptions, missing stakeholders, single-threading, commercial and procurement risk, and signs the close date is artificial. Score it 1–10." Then paste in every CRM note, email and meeting note by hand.
New way — what I actually type
"Challenge my assumptions on this deal — don't agree with me." It's connected to the CRM, so it pulls the stage history, the activity log, the stakeholder list and the linked call notes itself, then runs the same interrogation: strongest evidence, weakest assumptions, missing stakeholders, single-threading, artificial close dates, and a confidence score. The scepticism instruction still matters. I'm just not the one feeding it the raw material any more.
02
Preparing for a Pipeline Review
Most pipeline reviews start too late. The manager opens the CRM, the salesperson opens the CRM, and together they slowly discover what's in the pipeline. That's data collection, not a review. The manager should arrive already informed — which opportunities haven't moved, which close dates look unrealistic, which deals lean on one stakeholder.
Old way — the prompt
"Review this salesperson's current pipeline as a sales director preparing for a coaching conversation. Identify stalled deals, weak evidence, unusual value or date changes, vague next steps. Don't summarise the data — give me the five most important questions to ask." Then export the pipeline to a spreadsheet and paste it in.
New way — what I actually type
"Prep me for this week's pipeline review." It reads the open deals straight from the CRM and comes back with what needs attention and the questions to ask — no export, no spreadsheet in between. I still don't want it to conduct the review. I want it to make me better at conducting the review. That difference hasn't moved.
03
Building an Account View
Most account plans are terrible — static documents full of information everyone already knows, updated twice a year and read by nobody in between. Account planning should answer one question: how do we become more strategically important to this customer?
Old way — the prompt
"Act as a strategic account director. Build a commercial view of this customer covering their priorities, our current position, stakeholders and influence, whitespace and risk. Give me our three strongest positions, three biggest vulnerabilities, and the ten actions we need to take." Then gather every scrap of account history yourself first.
New way — what I actually type
"Build me the current account view for this customer." It pulls what we hold in the CRM and layers in public information about the customer's own priorities and pressures on top. The plan stops being a document updated twice a year and starts being something I can re-ask any time the picture changes — because asking again costs nothing.
04
Red-Teaming a Bid
Salespeople and bid teams get too close to their own response. You know what you meant, why the solution is good, why the answer works. The evaluator only sees the words on the page. One of the best uses of AI is asking it to act as a hostile evaluator rather than a supportive one.
Old way — the prompt
"You are evaluating this tender response on behalf of the customer. Be deliberately critical. Identify unsupported claims, generic wording, weak evidence, missing outcomes, and areas a competitor could score better. Score it against the published methodology, then rewrite only the weakest sections."
New way — what I actually type
"Be a hostile evaluator on this response — where could we lose marks?" — with the response and scoring criteria attached. This is one place the mechanics haven't fully caught up: nobody's tender documents live in a system AI can reach on its own yet, so there's still a paste. What's changed is the instruction is short, because it's a conversation now rather than a one-shot script — I can push back on what it flags and ask it to go again.
Worth saying plainly: not everything has moved to "just ask." Anything that lives outside a connected system — a document on someone's desktop, a PDF a customer sent — still has to be handed over by hand. The shift isn't that friction has disappeared everywhere. It's that it's disappeared everywhere the business actually keeps its data.
05
Coaching, from Patterns Rather Than Events
Most coaching today is event-driven. A deal goes wrong, a meeting goes badly, pipeline is low — and then the manager starts coaching. The problem is that isolated events mislead. Patterns matter more, and patterns only show up across months of opportunity history, call notes and win/loss data.
Old way — the prompt
"Act as an experienced sales coach. Don't analyse individual deals in isolation — look for recurring patterns. Identify what they consistently do well, where opportunities repeatedly weaken, whether discovery is deep or superficial, whether they challenge customers. Give me three coaching priorities for the next 90 days." Then spend an evening pulling together months of call notes to feed it.
New way — what I actually type
"Look for patterns in this rep's last quarter — deals, calls, pipeline movement — and give me three coaching priorities." It has the pipeline history and the call transcripts already, so it can look for the pattern the moment I ask instead of the moment I've found the evening to assemble one. That's the difference between coaching happening quarterly and coaching happening whenever it's actually useful.
The Real Opportunity Was Never the Wording
There's an understandable temptation to judge AI by how much time it saves — fewer emails written by hand, less admin, faster proposals. Useful things. But I think they miss the bigger opportunity, and the shift away from prompts makes that easier to see, not harder.
The prompt was always a proxy for the thing that actually mattered: knowing which question to ask. "Challenge my assumptions." "Don't summarise, tell me what to ask." "Look for patterns, not events." Those instructions are still exactly the discipline that makes AI useful in sales — I haven't stopped needing them, I've just stopped needing to retype them alongside a pile of hand-copied evidence every time.
The salesperson still needs to build trust. The manager still needs to coach. The customer still needs to make a decision. AI doesn't replace any of that — it means I turn up to it better prepared, more often, with less of my week spent as my own data-entry clerk.
// What's actually worth paying attention to
Not smarter prompts. A system that's actually plugged into how the business runs.
That's the part that changed — and the part worth building toward, whatever tools your own team ends up using.