Claude Cowork, now part of Claude, can clean a CRM export and draft follow-ups for stale deals. Export open deals and their contacts as CSV files into one folder for Claude Desktop, ask Claude to merge them, flag duplicates and list deals with no customer contact for 21 days, then have it draft one follow-up per deal in Manual mode, approved by you before anything is sent.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowTwo things make this safer than it sounds, and one makes it riskier. Anthropic's Sales plugin works from uploaded files when no CRM is connected, and its README says that when a skill can't write to the CRM it hands you a checklist to apply by hand, with nothing written or sent until you say so. The risk is the default: Anthropic's help centre says Auto is rolling out as the default permission setting on Pro, Max, Ultra and Team plans, and the Gmail connector's tool list includes sending and replying. So before the drafting step, switch the setting to Manual and block every send tool. The export also holds personal data, so keep it to the columns the job needs.
Which CRM columns should you export for Claude?
Export deals and contacts as two CSV files with only the columns the job needs: record IDs, deal name, stage, amount, key dates, owner, and each contact's name, email, phone, unsubscribe and bounce fields.
The illustrative business throughout is a solar installer with an owner and two salespeople, running a HubSpot pipeline with four open stages: Survey booked, Quote sent, Finance pending and Contract sent. Most CRMs export to CSV; HubSpot's knowledge base gives these steps for deals:
- Go to CRM > Deals, choose the pipeline from the dropdown in the upper right, and open the view you want (All deals exports everything in that pipeline).
- Click the view configuration icon to open View settings, then under Sharing select Export.
- Choose the file format and click Customize to pick which properties go in. The default is the properties and associations in your view.
- Click Export. HubSpot emails a download link, which expires after 30 days. You need export permissions to do this.
One catch shapes the whole job: a deals export includes associated contacts as names and record IDs, not their email addresses. So export contacts too, and let Claude join the two files on record ID. The columns the solar installer kept, and nothing else:
deals.csv Record ID | Deal name | Deal stage | Amount | Create date |
Close date | Last activity date | Last contacted |
Next activity date | Deal owner | Associated Contact (IDs)
contacts.csv Record ID | First name | Last name | Email | Phone |
Unsubscribed from all email | Email hard bounce reason
Leave out notes bodies, addresses, finance application details and anything about household income that the sales team may have recorded. They don't help a follow-up, and every extra column is more personal data sitting in a folder. If you need one detail from the notes (say, whether the customer asked about a battery), add it as a short custom property in the CRM first, then export that column. The broader rules for what a sales team should let AI touch are in an AI policy for sales teams' CRM data.
How do you get Claude to clean a CRM export?
To clean a CRM export, ask Claude to join deals to contacts, list likely duplicates without merging them, flag problem deals and save the results to a new workbook, leaving the original files unchanged.
The main worked example. The installer's export held 612 deals and 1,480 contacts across three years. Both files went into a new folder, Sales-2026-11, connected in Claude Desktop with the permission setting on Manual:
Work only in Sales-2026-11. Don't change deals.csv or contacts.csv.
1. Join deals to contacts using the Associated Contact IDs.
2. Find likely duplicate contacts: same email ignoring capital letters;
same phone number once spaces and symbols are removed; same first
and last name with different emails. Never merge: list them.
3. Flag deals with no associated contact, deals with no amount, and
deal stages that don't match the four current stages.
4. Note how many open deals have a close date on the last day of a
month.
5. Save clean-up.xlsx with tabs: Deals (joined), Possible duplicates,
Problems, Summary. Show me the Summary before anything else.
An illustrative Summary:
612 deals, 1,480 contacts. 297 deals are in an open stage.
Possible duplicate contacts: 74 groups (51 same email, different
capitals; 17 same phone; 6 same name, different email)
Deals with no associated contact: 38 (31 of them created before 2025)
Deals in stages that no longer exist ("Quoted", "Hot"): 23
Open deals with a close date on a month's last day: 211 of 297
That last line matters more than it looks. HubSpot's own documentation says that when a deal is created in an open stage without a close date, HubSpot sets the close date to the last day of the month. So 211 close dates are mostly defaults, not anyone's forecast, and a "stale" rule based on close dates would be nonsense. The 74 duplicate groups are a job for a person in the CRM itself, where merging keeps the activity history; Claude's list just makes it quick. For a full CRM tidy-up rather than this one job, see cleaning up a messy CRM with AI.
Working through the duplicates tab
Each duplicate group needs a decision, and the decisions are quicker than they look once they're in a table. Five illustrative groups from the installer's list:
| Group | What Claude saw | Decision |
|---|---|---|
| Same email, different capitals | Two contacts, one with a won deal, one with an open quote | Merge in the CRM; the open quote is for a battery add-on, so not stale |
| Same phone, different names | A couple who each filled in the enquiry form | Keep both; follow up once, addressed to whoever booked the survey |
| Same name, different emails | Work email on one deal, personal email on a newer one | Keep both; use the email on the newer deal |
| Same email, three contacts | A landlord with quotes for three rental properties | Merge contacts; keep three deals; one email covering all three |
| Same name, same street | Father and son at the same address | Not duplicates; leave alone |
The landlord row shows why Claude lists instead of merging. Three separate chasers to one person about three quotes would have looked automated and careless; one email that mentions all three properties reads like someone paying attention.
What makes a CRM deal stale?
A CRM deal is stale when it sits in a stage where deals usually stall, has had no customer-facing contact for a period set by your sales cycle, and has no next activity booked.
A stale-deal list is only as good as its rule. The installer's quotes usually close within three to four weeks, so the owner settled on this:
| Rule | Why | Deals left |
|---|---|---|
| Stage is Quote sent or Finance pending | Where deals stall in this business | 131 |
| Last contacted more than 21 days ago | Customer-facing contact, not internal notes | 94 |
| No Next activity date | Nobody already has a call booked | 86 |
| Not unsubscribed, no hard bounce, no "do not contact" flag | Respect opt-outs; don't email dead addresses | 72 |
| Quote less than 180 days old | Older quotes need re-pricing, not a chaser | 54 |
The 18 older quotes went to a separate "close or revive" tab for the owner to decide: some were marked lost in the CRM, a few got a phone call about a fresh quote, since panel and battery prices had moved since they were written.
The field that nearly sent the wrong list
The first version of the rule used Last activity date, and it quietly excluded 23 deals that should have been chased. HubSpot's definitions explain why: Last activity date includes notes and completed tasks, so a salesperson logging an internal note ("customer still thinking") reset the clock, even though the customer hadn't heard from anyone in two months. Last contacted only counts calls, one-to-one emails, meetings and messages. Switching the rule to Last contacted brought those 23 back. If your CRM isn't HubSpot, check what its "last activity" field actually counts before you trust it.
If the spreadsheet side is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.
How do you get Claude to draft follow-ups that sound like you?
Claude drafts follow-ups that sound like you when it works from three stage-specific templates, a tone guide and only the facts in each deal row, and you edit every draft. Fifty-four follow-ups written from scratch is most of a day; drafted this way, they take an hour of reading and editing. The drafting prompt:
For each deal on the Stale tab, draft one email in a new tab, Drafts.
Columns: Record ID | To (first name, email) | Subject | Body | Notes.
Tone: plain and friendly, from the salesperson named in Deal owner.
Under 110 words. One question the customer can answer in a line.
Use only facts in the row: first name, quote date, system size, stage.
Never mention prices, savings, payback, grants, finance rates or
deadlines. Never say "just checking in".
Quote sent: ask if they had questions about the quote.
Finance pending: ask if anything is holding up the application.
End every email with: "If you've decided not to go ahead, just reply
'no thanks' and I won't chase again."
The three stage templates the installer settled on, filled in with the wording that survived editing:
| Stage | Subject line | The one question | Why it works |
|---|---|---|---|
| Quote sent | Your solar quote from [quote date] | "Did any questions come up when you read it?" | Invites a reply without pressure; questions are the usual blocker |
| Finance pending | Your finance application | "Is anything holding the application up that I can help with?" | Stalled finance often needs one missing document |
| Contract sent | Your installation contract | "Would a quick call help you go through the contract?" | Contracts stall on small worries a call clears up |
An illustrative draft, before editing:
Subject: Your solar quote from 14 October
Hi [first name],
You'll remember we sent over your quote for a 6.2 kW system on
14 October. Customers who go ahead before the end of the month are
seeing payback in under eight years, so I wanted to make sure you
had everything you need. Did any questions come up when you read it?
If you've decided not to go ahead, just reply 'no thanks' and I
won't chase again.
[salesperson's name]
Two fixes before it could go anywhere. The payback claim and the "end of the month" deadline both broke the prompt's rules, appearing because the quote's system size invited a sales line; they came out, leaving "I wanted to make sure you had everything you need". And three other drafts named a panel brand that was in an old note but not in the export, which turned out to be a brand the installer no longer fits. The rule "use only facts in the row" was tightened to "if a detail isn't in a column, leave it out". Pushy wording and invented specifics are the two failure modes to read for in every batch; reading in batches of ten keeps it manageable. For the timing and number of chasers beyond a single follow-up, quote follow-up sequences with AI goes further, and the solar-specific economics of fast follow-up are in solar lead costs and follow-up.
How do you stop Claude sending emails by itself?
To stop Claude sending emails by itself, block the Gmail connector's send, reply and forward tools and keep the permission setting on Manual, because Blocked is the only setting that holds in every mode.
Drafts in a spreadsheet can't send themselves, which is the main reason to start there. If you'd rather have drafts appear in Gmail ready to send, set the connector up first. In Customize > Connectors, select Gmail and open Tool permissions. Anthropic's directory lists 29 Gmail tools, including create_draft, send_message, reply, forward and trash. For this job:
| Tool | Setting | Effect |
|---|---|---|
| search_threads, get_thread, get_message | Always allow | Claude can check whether a customer already replied |
| create_draft | Needs approval | Each draft appears in your Drafts folder after you approve it |
| send_message, reply, forward | Blocked | Nothing can be sent by Claude in any mode |
| trash_message, trash_thread | Blocked | No email can be deleted |
Keep the permission setting in the message box on Manual for the whole session. Anthropic's help centre explains that in Auto mode a tool marked Needs approval runs without asking unless an automated check flags it; Blocked is the only setting that holds in every mode. Checking for replies first is worth the extra step: in the example, 4 of the 54 customers had replied to an older email that never got logged in the CRM, and a chaser would have looked careless.
The Sales plugin route
Anthropic's Sales plugin is listed as Anthropic verified and installs from Customize > Plugins > Discover: select it and click Add (in the older layout, open the Cowork tab first). Plugins are available on all paid plans. Its README lists 36 skills, three of which fit this job: /sales:crm-hygiene-check, a read-only audit for missing fields, stale dates and stage mismatches; /sales:pipeline-review, for ageing and at-risk deals; and /sales:draft-outreach, for personalised emails. With nothing connected, the README says each skill works from an uploaded pipeline export and states plainly what it used and what it couldn't see, and CRM updates come back as a checklist. The plugin does a similar job to the prompts above with less writing on your part; the prompts give you tighter control over the stale rule and the wording. Either way, the permissions table applies, because the plugin uses the same connectors.
If HubSpot is connected through its connector, which Anthropic's directory lists with tools to search and update contacts, companies and deals, set its write tools to Blocked for this job and let Claude read live data instead of an export. The Sales plugin README adds a sensible rule for anyone tempted to schedule any of this: for scheduled runs, set send, post and write tools to ask, so nothing goes out or changes without a person approving it.
What should you check before sending CRM follow-up emails?
Before sending CRM follow-up emails, check that every recipient asked for a quote, that opt-outs and hard bounces are removed, and that each email names the sender and offers an easy way to say no.
A follow-up on a quote someone asked for is a reasonable email to send. Fifty-four of them at once still deserves a check, and the rules differ by where your customers are. Whatever applies to you, these habits keep a stale-deal campaign on the right side of most email and data-protection rules, and of your customers' patience:
- Only people who asked for a quote. Never add bought lists or old newsletter contacts to the stale-deal tab.
- Opt-outs win. Exclude anyone flagged Unsubscribed from all email in HubSpot, anyone with a hard bounce, and anyone a salesperson has noted asked not to be contacted.
- Say who you are. Send from the salesperson's own mailbox with the business's name and contact details in the signature.
- An easy way to say no. The "just reply 'no thanks'" line, honoured by updating the CRM the same day.
- One email per deal per month at most. A single follow-up, not an automated sequence nobody reads.
- Minimal data, deleted afterwards. The export folder is deleted once the CRM has been updated.
If you have customers in the EU, data-protection law such as the GDPR covers the personal data in that export, and e-marketing rules can affect what follow-ups you may send; if you're unsure how they apply to quote follow-ups, ask a qualified adviser before the batch goes out.
Sorting the replies when they come back
Replies arrive over the following week, and they're where the value is. Rather than reading each one against the CRM, the owner pastes them into the same task (or lets Claude read the threads through the Gmail connector's read tools) with this instruction: sort each reply as wants a call, has a question, not now (with the date they mention), went elsewhere, or no thanks, and suggest one next step for each. An illustrative result for the first twelve replies:
| Category | Count | Suggested next step |
|---|---|---|
| Wants a call | 3 | Book a call; salesperson rings within two working days |
| Has a question | 4 | Draft an answer for approval (two about battery sizing, two about the finance form) |
| Not now | 2 | Set a next activity date for the month they mentioned |
| Went elsewhere | 2 | Mark lost with the reason they gave |
| No thanks | 1 | Mark lost; no further emails |
Check the "not now" rows by hand. In the example, one reply said "maybe after the winter, we'll see", and Claude proposed a specific date three months out; the salesperson softened it to a check-in note rather than a firm callback, because the customer hadn't asked to be contacted on a date. Small misreadings like that are why every suggested next step is a suggestion until a person accepts it. And the two battery-sizing answers went to the engineer who did the survey before anything was drafted, since sizing advice depends on the site, not on the quote alone.
How do you put follow-up results back into the CRM?
To put follow-up results back into the CRM, ask Claude for an update checklist from the Drafts tab, then apply it by hand or approve each update one by one. That last step is the one most often skipped: a follow-up that isn't logged means next month's stale list includes the same customers, and the salesperson can't see what was said. Ask for the checklist like this:
For each email I've marked SENT in the Drafts tab, list the CRM update:
log the email on the deal, set Next activity date to 10 days after
today, and leave the stage unchanged. For each reply I paste below,
suggest a stage change and note, but don't decide closed-lost.
The salesperson applies the checklist in HubSpot in about 20 minutes, or approves each update one by one if the HubSpot connector's write tools are set to Needs approval and the session is in Manual mode. Replies saying "no thanks" are moved to closed-lost by a person, with the reason recorded, so the deal never reappears on a stale list.
Two and a half hours a month instead of six
An illustrative sum. Before, the owner spent around six hours a month working through open deals in the CRM and writing chasers, and usually stopped halfway. With this routine: 20 minutes exporting, 30 minutes on the clean-up and stale rule, an hour reading and editing 54 drafts in batches, 30 minutes sending and 20 minutes updating the CRM. About two and a half hours, covering every stale deal instead of half. On cost, Claude's pricing page lists Pro at $20 a month billed monthly or $17 on an annual plan, and Team at $25 a seat a month billed monthly ($20 annually) for two or more people, with no model training on your content by default. Anthropic notes that this kind of multi-step work uses limits faster than chat, so run it as one task rather than several.
If you want help fitting this into how your sales team actually runs, I offer one-to-one AI consultations for small businesses.
Making it a monthly routine without turning it into spam
- First Monday of the month: export deals and contacts with the same columns, into a fresh folder.
- Run the saved clean-up and stale-rule prompts, unchanged, so month-to-month numbers are comparable.
- Check replies, then draft, in Manual mode with send tools blocked.
- Send, log and delete: send approved drafts yourself, apply the CRM checklist, delete the folder.
- Review the rule each quarter. If deals keep going stale at the same stage, that's a process problem in the sales stage itself, not something follow-up emails will fix.
A read-only stale-deal report can be scheduled once the HubSpot connector is set up with write tools blocked, since Anthropic says scheduled tasks run in the cloud with your connectors. The drafting and sending stay supervised. That split, automated looking and human sending, is what keeps a monthly chaser useful to customers rather than another email they ignore.
Claude Cowork and CRM follow-ups: questions owners ask
Do I need to connect my CRM to Claude for this?
No. Anthropic's Sales plugin says every skill works without connectors: upload a pipeline export and it works from that, and where it can't write to a CRM it gives you a checklist to apply by hand. Connecting HubSpot or Salesforce adds live data, but for a monthly clean-up a CSV export is simpler and safer.
Can Claude send the follow-up emails for me?
It can if you let it: Anthropic's directory lists send and reply tools for the Gmail connector, and the Sales plugin can send email when asked. For follow-ups on quotes, keep sending manual. Block the send tools under Customize, Connectors, keep the permission setting on Manual, and send each approved draft yourself.
How many days without contact makes a deal stale?
There's no universal number; it depends on your sales cycle. For quotes that usually close within a month, 21 days without a customer-facing contact is a reasonable trigger. Base it on when you last contacted the customer, not on internal notes, and send old quotes to a close-or-revive review instead of a chaser.
What about the personal data in the export?
Export only the columns the job needs, keep the files in one folder, and delete them when the run is finished. Claude's pricing page marks model training as opt-out on Pro, Max and Ultra, so check your privacy settings; Team plans don't train on your content by default.
Further reads
- Solar Installer Software: What to Buy First and Where AI Helps — Solar CRMs compared, and where AI already sits in them.
- Zoho Zia vs ChatGPT: Which Should Answer Questions About Your CRM? — Built-in CRM AI against a general assistant for CRM questions.
- Solar Battery Quote Template: Options Explained With AI, No Backup Promises — Battery quote wording that avoids promises you can't keep.
- Sales Visit Report Template: Rep Voice Notes Into CRM Entries With AI — A nine-field visit report, the prompt that fills it from a two-minute voice note, and one retailer visit followed from recording to CRM.
- AI for Exporters: Six Jobs From Finding Buyers to Paperwork — Six export jobs AI can take on, from buyer research to shipping documents, each with a tool, a prompt to copy and the check that catches errors.
- AI for Fashion Companies: Line Sheets and Buyer Emails From Product Data — Start with the wholesale line sheet built from product data, then buyer emails. A fashion brand's order, prompts, sample outputs and pre-send checks.
Show 2 more related tutorials
- Claude (Formerly Cowork) vs ChatGPT Work: Which Cuts Owner Admin? — Run the same folder, customer-reply and weekly-report tasks in both tools, then compare checked owner time.
- Claude Scheduled Tasks: Monday Reports and Month-End Runs — Separate the report that runs every Monday from the month-end job you start after the figures are ready and approved.
Browse all AI-for-business tutorials
Sources: claude.com (Cowork is now Claude announcement, Sales plugin page, pricing, Gmail and HubSpot connector pages), Anthropic's Sales plugin README on GitHub, Claude help centre (Claude Cowork and chat are one Claude; Get started with Claude Cowork; Use plugins in Claude; connector tool permissions), HubSpot knowledge base (export records; default deal and contact properties). Checked 17 and 18 November 2026.
