"We want to make sales more efficient, but we don't know where to start." This is one of the most common questions Mihata receives from the SMEs we support. With the rise of generative AI, the work that surrounds sales — company research, proposal drafts, email writing, meeting summaries — has indeed begun to change in a meaningful way. That said, throwing everything at AI rarely produces good results. "What you delegate to AI versus what humans must own" is the central question that decides outcomes.
This article breaks the sales process into five phases, then for each phase shares practical prompt examples, a rollout path from PoC to organization-wide adoption, KPI design, and operational cautions — as concretely as possible. The structure is designed to be hands-on, so your team can start moving today.
Where in Sales Can AI Help? Break the Process Into Five Phases
Trying to "AI-ify" the entire sales process at once usually stalls midway. A more realistic approach is to first split sales into five phases and, for each one, separate "what AI is good at" from "what humans must own."
Phase | Main Tasks | Where AI Helps | What Humans Must Own |
|---|---|---|---|
1. Targeting | Customer list cleanup, prioritization | Generating segment hypotheses, articulating scoring criteria | Strategic judgment, final visit order |
2. Research | Industry, company, key-person research | Summarizing public information, organizing key issues, hypothesizing customer pain points | Fact-checking, recency verification |
3. Proposal | Building proposals and plans | Outline drafts, skeleton structure, per-section drafts | Originality, logic, pricing, final polish |
4. Email and outreach | First-touch / follow-up / thank-you emails | Generating drafts, adjusting tone, summarizing | Phrasing tuned to the relationship, confidentiality care |
5. Reflection | Meeting minutes, next-action extraction | Summarizing audio and text, organizing key points | Interpretation, judgment, sense of how to share internally |
Laying it out this way makes one thing clear: AI is not a replacement for thinking — it is an assistant that buys back time for thinking. The rest of this article walks through each phase with concrete prompt examples.
Phase 1: Targeting — Use AI to Quickly Expand "Who Are We Selling To?" Hypotheses
Targeting has traditionally relied on the tacit knowledge of sales managers and executives. AI is a powerful aid for translating that tacit knowledge into "explicit, written conditions."
A Prompt to Generate Many Segment Hypotheses Quickly
Prompt:
"You are a B2B sales-strategy consultant. Our product is 'back-office automation SaaS for small and mid-sized businesses.' For companies with annual revenue between roughly 100M and 3B yen, propose 10 promising customer segments across the dimensions of industry, department, pain points, and decision-making authority. For each segment, list expected pain points, the messaging that should resonate, and a hook for the first contact, in bullet form."
The key is to give your own context (company size, industry, traits of existing customers) as concretely as possible. AI tends to default to "average" answers, so it only becomes useful once you inject your specific context.
Articulating Scoring Criteria
If you have existing customers and lost deals organized as a CSV, asking AI "Based on these win patterns, propose a formula for a lead score" will return a scoring proposal that combines factors like industry, headcount, and recent disclosures. The final weighting is for humans to decide, but you arrive at a working draft far faster than from a blank page.
Phase 2: Company Research — Stick to Public Information and Verify Rigorously
Company research is the most time-consuming step of sales prep. Earnings briefs, corporate sites, press releases, job postings, trade publications — even veteran sales reps "never have enough time" to read all of these and organize the points that matter.
Non-Negotiables
- Public information only. No reposting from social media, no scraping that violates terms of service, no unauthorized use of paid databases.
- AI output is a draft. Always confirm with primary sources. AI can return outdated information or mix in details from a different company.
- Never input confidential or personal data. Compose only with information that is OK to share externally.
Example Prompt for Company Research
Prompt:
"Read the following text I will paste — 'XYZ Corp.'s corporate site, latest earnings brief, and press releases from the past 12 months' — and summarize from these angles: 1) core businesses and recent performance trends, 2) themes of the mid-term plan, 3) recent organizational and personnel changes, 4) hypotheses on which pain points our 'back-office automation SaaS' is most likely to fit, 5) a list of questions to verify in the first meeting. For any inference not supported by an explicit source, prepend 'inference:'."
The trick with this prompt is the instruction to "label inference as inference." AI tends to mix fact and conjecture; forcing the boundary at the output stage saves time on downstream verification.
Understanding Key People
Where a director or department head has given public interviews or spoken at events, summarizing them to extract "topics of interest" and "talking points consistent with their past statements" changes how the first meeting lands. Conversely, mining private social posts for sales purposes is not advisable, both for relationship and legal reasons.
For a more systematic approach to research itself, see the companion article A Complete Guide to AI-Powered Business Efficiency.
Phase 3: Proposal Drafts — A Three-Stage Rocket: Outline, Skeleton, Section Drafts
The biggest tip for using AI on proposals is "don't ask for a finished document on the first try." A flat "write a proposal" will produce only generic platitudes. Instead, direct AI in three stages: outline → per-section skeleton → section drafts.
Step 1: Generate an Outline
Prompt:
"You are an expert in structuring B2B proposals. Based on our service overview and the following customer research notes, propose three different chapter outlines for an A4 proposal of about 15 pages. For each option, add one sentence describing where in the customer's buying decision it lands hardest."
Step 2: Build the Skeleton
Pick the outline you like, then have AI list "the claim, evidence, anticipated objection, and answer to that objection" for each chapter. That alone surfaces the logical skeleton of the entire proposal.
Step 3: Draft Each Section
Now have AI draft prose for each chapter against the skeleton. Specify the tone — for example, "polite, logic-first, avoid hype" — to nudge the writing toward your house style.
One caveat: do not let AI write pricing, scope, contract terms, or your proprietary know-how. The parts that are the lifeline of your sales pitch should always be written and owned by a human. That is the iron rule of AI-assisted proposals.
Phase 4: Email — A Three-Stage Flow of Template, AI Generation, Human Review
Email is probably the phase where most sales reps feel "AI made my life easiest." It is also the area with the highest risk of damaging a relationship.
The Basic Operating Flow
- Build templates: write down internal templates for first outreach, re-engagement, thank-you, follow-up, quote delivery, polite declines, and so on.
- AI personalization: hand AI the recipient's situation and a summary of recent exchanges, and have it generate copy on top of the template.
- Human review: always check proper nouns, numbers, honorifics, and any inadvertent confidentiality leaks before sending.
Example Prompt for a First Email
Prompt:
"Based on the following recipient profile and our value proposition, write the body of a first-touch email. Conditions: 1) subject line within 8 words, 2) body within 150 words, 3) two-line self-introduction, 4) reference one hypothesized pain point of theirs, 5) CTA is 'a 15-minute online intro call,' 6) avoid sales hype; keep it polite and concise."
Specifying conditions all the way down to subject-line word count, body length, and CTA type makes the output dramatically more consistent. Think of it as handing AI your "evaluation rubric" instead of letting it improvise.
A Standard Pre-Send Check
For AI-generated emails, the safest practice is a checklist review covering "(1) salutation and honorifics, (2) proper nouns, (3) numbers, (4) confidentiality, (5) factual accuracy." When operating as a team, keep the checklist itself in a shared document — accidents drop sharply.
Phase 5: Meeting Summaries and Next-Action Extraction
Writing post-meeting notes is unglamorous but reliably eats time. With an AI transcription + summarization flow in place, this step shrinks dramatically.
A Standard Prompt for Meeting Notes
Prompt:
"The text below is a transcript of a sales meeting with XYZ. Produce minutes in this structure: 1) date, time, and attendees, 2) agenda, 3) main points raised by the customer (separating fact from opinion), 4) summary of our proposals, 5) decisions made, 6) action items and next steps (with owner and due date), 7) risks to verify before the next meeting. Keep proper nouns verbatim from the source. Mark any inference as 'inference.'"
The two key elements to bake into the template are "separate fact from opinion" and "label inference as inference." Those two alone significantly reduce later misinterpretation.
For criteria on choosing meeting AI and minutes-assistance tools, see the comparison piece A Practical Guide to Using ChatGPT at Work.
Five Steps to Roll Out AI in Sales — From PoC to Team-Wide Adoption
"Roll it out company-wide on day one" doesn't work. From Mihata's field experience, the highest success rate comes from a five-step staged rollout.
- Pick a pilot: choose just one of the most painful tasks (often meeting notes or first-touch email).
- Two-week PoC with two or three people: articulate what worked and what risks emerged. Build a starter set of about five prompt templates.
- Write operating guidelines: confidentiality handling, personal-data handling, tool selection, review process — all in writing.
- Team rollout: distribute training, prompt library, and checklist; set up one internal Q&A channel.
- Iterate: review KPIs monthly and update prompt templates.
The purpose of the PoC is "finding the walls" rather than "showing impact." Surface real-world issues early — leakage, hallucination, missed reviews — and convert them into rules before scaling. Don't reverse this order.
KPIs and Measurement — Time, Volume, and Win Rate
Make sure the impact of AI adoption is describable in numbers, not just feel. At minimum, track these three axes.
Axis | Example metrics | How to measure |
|---|---|---|
Time | Time to write meeting notes / proposals / emails | Sample timing before and after (e.g., 10-item average) |
Volume | First-touch outreach, proposals submitted, meetings held | SFA / CRM / spreadsheet logging |
Quality | Win rate, average deal size, lead-to-meeting rate | Monthly aggregation of stage transitions in your SFA |
A caution: when you publicize numbers like "X% productivity gain" externally, always derive them from your own primary data. Citing generic numbers without a source destroys credibility instantly. AI buys you efficiency — be honest with the numbers it produces.
Cautions When Adopting AI — Accuracy, Confidentiality, Human Review
Finally, three non-negotiable cautions when bringing AI into sales.
1. Accuracy (Mitigating Hallucinations)
Generative AI can return plausible-sounding output that is factually wrong — known as hallucination. Mitigate it by (a) always asking for sources, (b) having a human re-verify numbers and proper nouns against primary sources, and (c) staying alert that information may be outdated.
2. Confidential and Personal Information
Before sending customer information, deal terms, draft contracts, or personally identifiable data into AI, always check the terms of service and the data-training policy of the service in question. Choose a business plan that allows opting out of training, and put your "what may be input" rules in writing. For personal data, refer to the official guidance from Japan's Personal Information Protection Commission and translate it into your own operating rules.
3. Don't Skip Human Review
Proposals, emails, and minutes generated by AI must be reviewed by a human before they leave the company. Don't skip this. To prevent review from becoming an empty ritual, keep the checklist in a shared doc and log who reviewed when.
How Mihata Helps — From PoC to Steady-State Operation
Mihata works alongside SME sales teams end to end: PoC design, prompt-template libraries, internal guidelines, training sessions, and ongoing improvement. Most of our conversations start at "we don't know where to begin" or "we ran a PoC, but team-wide rollout has stalled." If you want to lift sales productivity step by step in a way that fits your context, please reach out.
Summary — AI Is Not "A Replacement for Sales," but "A Tool That Buys Back Time to Think"
The first decision when bringing AI into sales is "what to delegate to AI versus what humans must own." By breaking sales into the five phases above (targeting, research, proposal, email, reflection) and pairing each with the right prompts and review process, you can reclaim time without compromising quality.
The point is to position AI not as "a replacement for sales", but as "a tool that gives sales back the time to think and face customers." Templates, prompt design, review process, KPIs — none of these come together in a day, but a step-by-step rollout from PoC produces a visibly different team in three months. We hope the prompts in this article help you take the first step.