[Verification] How much can Gemini 3.6 Flash automate sales note organization? Testing BANT and next action extraction in 3 cases

The real "work" begins after the sales meeting ends

Many sales professionals must feel this way. Entering scribbled notes into the CRM later and organizing next actions.

When this follow-up work is put off, the information becomes vague before you know it...

In fact, according to HubSpot's 2026 survey, the time saved by people using generative AI for "entering activity records into CRM/SFA" was an average of 4.0 hours per week, which was the highest among the tasks surveyed. CRM entry is a prime example of a task where sales professionals can easily feel the impact of AI (in the same survey, Gemini was the second most used generative AI by sales professionals after ChatGPT).

If AI could finish this "organizing notes and creating data for CRM entry" in a few seconds, the burden would change significantly.

However, one important thing first. Gemini 3.6 Flash will not automatically register data into your CRM.

What it can do is "extract necessary information from sales notes and organize it into a draft (data for CRM entry) that a person can verify."

In this article, we verified how useful this is in practice using three sales notes.

Conclusion

I will state the conclusion first.

Gemini 3.6 Flash can be entrusted with extracting "BANT, next actions, and concerns" from sales notes and organizing them into a draft for CRM entry.

However, it is based on the premise that registration into the CRM and final verification are performed by a person.

Numbers and proper nouns, in particular, require verification.

Verification Results

Notes with complete information were extracted with zero errors, undecided items were left blank rather than fabricated, and decision-makers were correctly identified even when mixed with small talk.

Creation time was a few dozen seconds (assuming manual work takes about 5 minutes, this is an estimated 90% reduction in draft creation time).

What is Gemini 3.6 Flash? Why is it suitable for "organizing sales notes"?

Gemini 3.6 Flash is a model that emphasizes the balance between speed and processing power. It is suitable for tasks such as extracting necessary information from text like sales notes and organizing it into a fixed format (JSON).

(A brief technical note: Output tokens have been reduced by approximately 17% compared to the previous generation 3.5 Flash. A free tier is also available, and the standard pricing for the Gemini API is $1.50 for input and $7.50 for output per 1 million tokens, which is a lower output unit price than the previous model ($9.00 for output).)

The following three things are mainly what can be entrusted to it.

• Extracting key points: Extracting "BANT, next steps, and concerns" from long meeting notes

• Organizing key points: Summarizing meeting content into an easy-to-review format

• Formatting for CRM fields: Preparing extraction results into easy-to-register JSON (with fields)

What is BANT anyway? (The 4 items to extract)

The "BANT" framework, often used in sales, consists of 4 perspectives for measuring the potential of a deal.

• Budget: Is there a budget available?

• Authority: Is the person a decision-maker?

• Needs: What issues do they want to solve?

• Timeframe: When do they want to implement it?

If you can automatically pick up these 4 items plus next actions and concerns from meeting notes, CRM entry becomes much easier.

Ready-to-use meeting note extraction prompt

I will post the prompt used in the verification as-is so you can copy and use it immediately.

あなたは営業アシスタントです。以下の【商談メモ】から、CRM登録用の情報を抽出してください。

出力は次のJSONのみ（前置き・解説は不要）：
{
  "budget": "予算（不明なら空文字）",
  "authority": "決裁権・担当者（不明なら空文字）",
  "needs": "課題・ニーズ",
  "timeframe": "導入時期（不明なら空文字）",
  "next_action": "次のアクション",
  "concerns": "懸念点（なければ空文字）",
  "confidence": "high | medium | low"
}

・メモに書かれていない項目は、推測で埋めず空文字にする
・金額・日付・固有名詞は、メモの表記どおり正確に写す

【商談メモ】
（ここにメモを貼る）

If you are trying it in the Gemini app, just instruct it to output "JSON only".

If you are integrating it into a business system via API, using Structured Outputs (specifying a JSON Schema) will allow you to receive it in a more stable format.

Verification: Method, Evaluation Criteria, and Results

Forvio's motto is "Verification over atmosphere." Using dummy meeting notes, I compared the same task between "manual work" and "Gemini 3.6 Flash."

Verification Environment

• Verification date: 2026-07-22

• Model used: Gemini 3.6 Flash

• Input data: 3 dummy meeting note cases (1. Information complete / 2. Budget and timeframe missing / 3. Includes small talk, with separate contact person and decision-maker)

• Comparison target: Manual work by staff (assuming approx. 5 minutes per case)

Evaluation criteria (4 items)

1. Extraction comprehensiveness (Were all items in the notes captured without omission?)

2. Extraction accuracy (Were numbers, proper nouns, and dates captured correctly?)

3. Format stability (Does it return the same JSON fields every time?)

4. Judgment (Can it distinguish between the point of contact and the decision-maker, and filter out small talk?)

Results

• Extraction comprehensiveness: For notes with complete information, all items (BANT + Next Action + Concerns) were extracted without omission.

• Accuracy of numbers and proper nouns: Zero errors. It accurately transcribed items like "Initial 5 million + maintenance 800k," "9/5," "March 2027," "General Manager Sato," and "Company B" exactly as written in the notes.

• Handling of missing information (Most important): For notes where budget or timing was undecided, it did not fill them in with guesses but left them as blank. The safety of not producing "plausible-sounding lies" was confirmed.

• Distinguishing between point of contact and decision-maker: Even in notes mixed with small talk, it did not stop at the point of contact (section manager level) but included the decision-maker (General Manager). Small talk like "weather" or "cafes" was not mixed into any fields, and there was no confusion between current costs (600k/month outsourced) and target amounts (300k/month).

• Format stability: In all 3 cases, it output in the JSON format with the 7 specified fields.

• Time: Output was completed in a few dozen seconds. Excluding the time for verification and entry into the CRM, it is estimated that the time to create a draft can be reduced by approximately 90% (assuming manual work takes about 5 minutes).

• Weaknesses (honestly): It occasionally omitted supplementary nuances like "already approved" (this can be mitigated by "splitting fields" as described later).

Note that while the confidence value (high/medium/low) is convenient, it does not guarantee accuracy. Please treat it as a reference value for prioritizing items that humans need to verify.

Actual output (what was returned in the verification)

For example, we provide the following sales meeting note.

Here is the actual output returned (in the specified JSON format).

{
  "budget": "初期500万＋年間保守80万",
  "authority": "株式会社サンライズ商事 購買部 佐藤部長（決裁は佐藤部長で完結）",
  "needs": "在庫管理システムの入れ替え（現行は10年前のもので不満が大きい）",
  "timeframe": "今期中（〜2027年3月）",
  "next_action": "9/5に技術担当同席でデモ",
  "concerns": "競合はB社も検討中",
  "confidence": "high"
}

Amounts, dates, and proper nouns are also accurate. All that remains is for a human to verify and register it.

One practical discovery as well.If you do not specify "JSON only, no preamble"Gemini will return long-form advice such as "Meeting Summary + Next Action Plan."The output format can be properly controlled by specifying it in the promptthat is.

Can it be registered directly from Gemini to a CRM?

No, it cannot. What was verified this time was strictly limited to "creating data (drafts) for CRM registration." For Gemini to write directly into Salesforce or HubSpot, you would separately need CRM API integration, automation tools (like Zapier), or a dedicated SFA integration service.

What Forvio verified washow much can be done with just general-purpose AI and prompts without using dedicated services.

For creating drafts, this is sufficiently practical.

3 tips for increasing accuracy

1. Break down items in detail: For example, splitting authority into "point of contact" and "decision maker." This makes it harder for nuances like "approved" or "undetermined" to be lost.

2. Have unknown items left blank: Do not let it fill them in by guessing (this is already specified in the prompt used here).

3. Have it provide the basis (original text) for the extraction: This makes verification much easier.

For example, this is what happens when you split the items.

{
  "contact_person": "田村さん",
  "decision_maker": "山本部長",
  "budget": "月30万くらい",
  "budget_status": "来期予算は未確定",
  "current_cost": "外注コスト月60万"
}

The weakness this time (supplementary nuances being lost) is not just a model issue, but also partly due tohaving too few JSON fields to act as a receptacle. This can be prevented through item design.

Practice: Steps to introduce this to your sales team

1. Decide which items to extract based on your CRM input fields (BANT + Next Actions + Concerns, etc.)

2. Fix those items in the prompt and have it return in JSON format (use the full prompt above)

3. Paste the meeting notes (or transcript of an online meeting) into thedesignated field in the promptand execute (*Note: If you leave it pasted in the previous conversation, it may not be picked up. It is most reliable to paste it into the designated field every time)

4. Have a human supplement and verify items with low confidence or blank fields

5. Always verify numbers, proper nouns, and dates before registering to the CRMdo

Recommended for / Not recommended for

• Recommended: Sales and inside sales who have many meetings, tend to put off CRM entry or recording, or want to automate the preparation of input data

• Not recommended: Those whose recording format varies by meeting and cannot be defined, those who want to automatically register to CRM without verification, or environments where confidential information cannot be shared with AI

Summary

Gemini 3.6 Flash is not an AI that will automatically register data into your CRM for you.

However, it is fully capable of extracting BANT, concerns, and next actions from meeting notes to create a draft that humans can review.

What you should leave to AI is not "judgment" or "registration," but "extraction, organization, and drafting." Humans should focus on the final review and preparation for the next meeting. I feel this is the most realistic and low-risk way to utilize AI at this moment.

On the other hand, I must be honest about its weaknesses. It occasionally omits nuanced details like "approved," so human review before CRM registration is essential (this can be mitigated by breaking down items into finer detail).

Frequently Asked Questions (FAQ)

Q. Can I use it with the free version?
A. It can be used from the free plan of the Gemini app. However, there are limits on usage frequency and features.

Q. Can it also be used for audio (meeting recordings)?
A. If you paste the transcript of the meeting, you can extract information in the same way (you will need a separate tool for the transcription itself).

Q. Is it okay to input customer information?
A. Please be sure to check your internal regulations, subscription plan, and the terms regarding data storage and training usage (also refer to the notes below).

Q. Can it automatically register data to Salesforce or HubSpot?
A. Not directly. You will need separate CRM API integration or automation tools.

Note from the author (Caution regarding confidential information)

Meeting notes may contain confidential information such as customer names, amounts, contact details, and internal company matters. Instead of inputting them directly into free services for individuals, please check your internal regulations, subscription plan, and the terms regarding data storage and training usage (Google's guidance states that while the free tier uses input content for product improvement, the paid tier does not; conditions vary by plan and service). We recommend anonymizing customer names and contact persons when verifying, if necessary.

Also, the 3 cases of meeting notes verified this time were dummies. Although the accuracy was high, it does not guarantee the same results for every meeting. When actually implementing it, please perform additional verification using your company's own meeting data.

We are also verifying the latest features of Claude

At Forvio, we are also verifying the latest features of Claude while simulating actual business operations. If you are interested in the differences between Gemini and Claude and how to use them for work, please take a look.

Read next (Gemini Work Techniques Series)

• Gemini Work Techniques ① Where does Gemini 3.6 Flash excel in work? (Strengths and weaknesses)

• Gemini Work Techniques ② Customer Support Edition: How much can inquiry handling be automated?

• Gemini Work Techniques ④ Marketing Edition: Automating competitor monitoring into weekly reports

For more systematic learning (Books)

For those who want to connect AI utilization like this to overall sales process improvement rather than ending it as a one-off technique, I have selected these two books.

For those who want to systematically understand CRM, marketing, and customer management systems, I recommend "The HubSpot Encyclopedia [Revised Edition]." For those who want to learn how to utilize AI in sales activities and achieve results, I recommend "Accelerate with AI! The Sales Textbook."

Since this combination covers both "CRM mechanisms" and "practical sales using AI," if you want to streamline not just the organization of meeting notes but your entire sales operation, please take a look.

Sources

• Google Official Blog "Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber": https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/

• Gemini Developer API Pricing and Data Handling: https://ai.google.dev/gemini-api/docs/pricing

• Gemini API Structured outputs (JSON Schema): https://ai.google.dev/gemini-api/docs/structured-output

• HubSpot "Survey on Awareness and Reality of Sales in Japan 2026" (4.0 hours per week saved on CRM/SFA input through generative AI utilization): https://www.hubspot.jp/company-news/stateofsales-20260227

List of articles recommended by the author

If you found this article helpful, please give it a "like" and follow me.

At Forvio, we test AI tools in actual work scenarios and honestly verify what they can and cannot do. We will continue to share information on AI utilization that can be used in business, such as in sales, marketing, and customer support.