AI won’t sell for you. It gives back the eight hours a week you currently spend not selling.
Most “AI for sales” advice is written for high-volume transactional teams — blast more emails, book more meetings. If you run a long technical sale with a handful of serious opportunities a year, almost none of it applies. This is what actually works when the deal takes nine months and the buyer is an engineer.
What AI is genuinely good at — and what it quietly ruins
The useful distinction isn’t “sales tasks” versus “non-sales tasks.” It’s whether the output is checkable. AI is excellent where you can verify the result in seconds and terrible where being subtly wrong costs you a relationship you spent two years building.
High leverage
Output is checkable, volume is high, and being 90% right is genuinely useful.
Low leverage / actively harmful
Output is hard to verify, stakes are high, or the value was the human effort itself.
The long-cycle deal, with AI in its proper place
A nine-month technical sale, and what should be machine-assisted at each stage. Notice the pattern: AI clusters at the ends — preparation and administration. The middle, where the deal is actually won or lost, stays human.
Identify
Weeks 1–3
Account research, org mapping, trigger events, technical context. Build the picture before the first conversation.
Qualify
Weeks 3–8
Is this real? Who decides? What’s the actual driver? AI can prep the questions. It cannot read the answers.
Scope
Weeks 8–20
Technical discovery, site visits, engineering judgement. The stage that wins the deal — and the least automatable.
Propose
Weeks 20–30
First drafts, structure, consistency across documents. You supply every number and every commitment.
Close & keep
Weeks 30–40+
Follow-up discipline, CRM hygiene, handover notes, account history that survives you leaving.
The stages where AI helps most are the ones nobody enjoys. That’s not a coincidence — it’s the whole opportunity.
Five prompts that earn their keep
These are structured for long-cycle technical selling, not SaaS. Each one is written to produce something you can act on rather than something that reads well. Open any of them for the full prompt and the reasoning behind its shape.
01The pre-meeting briefWalk in knowing more than the person who invited youSaves ~90 min›
The mistake is asking for “information about the company.” You get a Wikipedia summary. Ask instead for the specific things that change how you run the meeting — constraints, pressures, and the questions you’d be embarrassed not to have considered.
I'm meeting [ROLE] at [COMPANY] next week about [WHAT YOU SELL]. Long sales cycle, technical buyer. Give me: 1. What this company actually makes or does, in plain terms — including how they make money. 2. Anything in the last 18 months that suggests pressure or change: expansion, new plant, safety incident, leadership change, lost contract. 3. The three most likely operational constraints someone in [ROLE] is dealing with right now. 4. Five questions I could ask that would show I understand their world — not generic discovery questions. 5. What I'm most likely to be wrong about here. Flag anything you're inferring rather than sourcing.
Point 5 is the one that matters. It surfaces your own blind spots, and it’s the line most people delete.
02Notes to CRM recordThe habit that quietly saves entire dealsSaves ~4 hrs/wk›
CRM data quality is the root cause of almost every pipeline visibility problem. The reason reps don’t update the CRM isn’t laziness — it’s that turning a messy conversation into structured fields takes twenty minutes they don’t have. This collapses it to two.
Turn these rough meeting notes into a CRM entry. Output exactly these fields: - Summary (3 sentences max, factual) - People met (name, role, apparent influence) - Stated need vs underlying need - Explicit next step, with owner and date - Risks or red flags - What I still don't know Rules: do not invent detail. If something wasn't said, write "not established". Keep it in my voice, not marketing language. NOTES: [paste your raw notes]
“Not established” is doing real work here. A field that honestly says unknown is far more useful than a plausible guess sitting in your pipeline for six months.
03The pre-mortemFind out how the deal dies, before it doesSaves lost deals›
Most deal reviews ask “how do we win this?” That question invites optimism. Inverting it — assume it’s already lost, explain why — surfaces the risks people are too invested to say out loud.
It's nine months from now and we lost this deal. Here's what I know today: [deal context, stakeholders, timeline, competition, what's been agreed, what hasn't] Write the post-mortem. What killed it? Give me the six most likely causes, ranked by probability, not by how uncomfortable they are. For each: the early warning sign I'd see in the next 30 days, and the one action that would most reduce the risk. Be blunt. Include causes that are my fault.
Run this the week you submit a proposal, not the week it goes quiet. By then the warning signs have already been and gone.
04Proposal first draftStructure from the machine, substance from youSaves ~3 hrs›
The value is never the prose — it’s the structure and the completeness check. A good technical proposal fails because something obvious was left out, not because the writing was flat. Explicitly ask for the gaps.
Draft a proposal for [CUSTOMER] covering [SCOPE]. Their stated problem: […] What we agreed in discovery: […] Commercial terms: […] Structure: problem as they described it → what we propose → why this approach over alternatives → scope boundaries → what we need from them → commercial summary → next step. Hard rules: - Use ONLY figures I've given you. Where a number is needed and I haven't supplied it, write [TBC — Itai] and list it at the end. - No superlatives, no "cutting-edge", no filler. - Write for an engineer who is sceptical. Then list what's missing that a buyer would expect to see in a proposal like this.
The [TBC] convention is the safety mechanism. It makes fabricated numbers structurally impossible rather than something you have to catch by reading carefully.
05The Monday pipeline interrogationReplaces the report nobody buildsSaves ~1 hr/wk›
Export your open deals to CSV and interrogate them conversationally. Faster than building a report, and it catches the deals that are quietly rotting — which is the failure mode that actually costs you the quarter.
Attached is my open pipeline export. Tell me: 1. Which deals have had no activity in 21+ days, sorted by value. 2. Which have had their close date moved more than twice — treat these as at risk regardless of what stage says. 3. Where is the pipeline concentrated? If my top two deals vanished, what's left? 4. Which deals are in a late stage but missing a defined next step. 5. Three questions I should be asking about this pipeline that I'm probably not. Just the analysis. No encouragement.
Strip customer names before uploading if you’re using a consumer AI account. Check your company’s data policy first — this is the one place in this guide where you can create a real problem.
Four things to never hand over
Not out of principle. Out of self-interest — each of these costs more to repair than it ever saves.
× Technical commitments in writing
Throughput, tolerances, cycle times, compliance claims. A generated number that turns out to be wrong doesn’t just lose the deal — it can end up in a contract. Every technical figure gets checked by a human who owns the consequence.
× Outreach to your whole market
If your total addressable market is a few hundred companies, it is a finite resource. Automated sequences spend it all at once. In industrial B2B your reputation arrives before you do, and it takes years to rebuild.
× Anything a customer will read as personal
Condolences, congratulations, thanks after a site visit, apologies when something went wrong. These are the moments the relationship is actually built. Outsourcing them is the one shortcut that’s worse than not sending anything.
× The decision to walk away
Qualifying out is the highest-value judgement a BDM makes, and it rests on information that was never written down. A model working from your CRM only sees the optimistic version of the deal — because that’s the version people record.
What you actually need
Three tools, not thirty. Most “AI sales stack” articles list forty products because listing products is easier than explaining a workflow. If the five prompts above are all you ever use, you’re ahead of almost everyone.
A frontier AI assistant
The thinking layer
Where all five prompts above run. Pick one with a long context window so you can paste a whole discovery document, and web search so research prompts pull current information rather than training data.
A CRM that holds the record
The memory layer
AI output has to land somewhere permanent or the leverage evaporates. Prompt 02 is worthless if the structured record it produces gets pasted into a notepad. You need stage dates and activity history captured reliably.
An automation layer
The plumbing
For the small number of things worth making automatic: stalled-deal alerts, routing new enquiries, pushing meeting notes into the CRM. Keep this deliberately thin — every automation is something that can silently break.
Get the B2B Sales Stack Cheat Sheet
The stack these five prompts plug into — one page covering CRM, automation and prospecting. Sent to your inbox.
The bottleneck was never the writing
AI gives you hours back. Whether that turns into revenue depends entirely on whether the system underneath — pipeline, process, follow-up — is capable of using them.
Some links on this page are affiliate links. If you sign up through them we may earn a commission at no extra cost to you. The prompts above work with any AI assistant and any CRM — nothing here depends on you buying anything.