What you need to know
- An AI operator classified all 113 business accounts in one pass: 87 records updated, 3 mis-tags corrected, and every change logged for audit.
- Segmentation went from impossible to a one-line filter, so any vertical (accountants, real estate, trades) is now seconds away.
- Two multi-touch campaigns ran in a single session: around 150 personalised, tracked customer touches, roughly 78 sent in one morning.
- Every send was guarded: dry-run by default, human approval per send, account-aware deduplication, bounce-aware, and fully logged.
- The work was directed in plain English on Zoho CRM and Zoho Books. No spreadsheets, no manual data entry.
An AI operator cleaned Gibson’s own 113-account CRM and ran around 150 personalised, tracked customer touches in a single working session, with a human approving every send. No spreadsheets, no manual data entry, no send-and-hope. This is the real worked example: how to clean a messy CRM with AI, how to segment customers by industry automatically, and how to run AI email outreach without it becoming spam.
It is also a Gibson Sandbox build: AI doing the unglamorous work that usually blocks marketing, with the judgement and the guardrails you would want from a careful operator, all directed in ordinary sentences.
The problem: a CRM too messy to market to
Gibson’s customer base lived in Zoho CRM, but the data was not usable for marketing. Of 113 business accounts, around 85 had no industry or vertical tag at all, and some of the few that did were wrong. A dental practice was filed under “Real Estate”. That made even a basic question (which of our customers are accountants we have done work for?) impossible to answer without eyeballing every record by hand.
So outreach was ad-hoc. No clean segments, no systematic follow-up, and no reliable way to avoid emailing the same client twice or chasing an address that had already bounced.
How do you clean up a messy CRM with AI?
Working in plain English, an AI operator (Claude) was connected directly to Zoho CRM and Zoho Books. The work ran in three steps.
Step 1: classify the entire CRM in one pass
The AI read every account’s name, its contacts’ email domains and its deal history, and for the genuinely ambiguous ones it checked the company’s website. It classified all 113 accounts into a tight set of native Zoho industry values, deliberately not inventing new categories, produced a review sheet for sign-off, then wrote the verticals back to the CRM.
The result of one pass: 87 records updated, 3 wrong tags corrected, and every change logged for audit. Segmentation went from impossible to a one-line filter.
According to Albert Triolo, founder of Gibson Promotions, an AI operator classified all 113 of our business accounts in a single pass, took the records with a usable vertical from roughly 28 to 112 out of 113, and logged every change. The Sydney SMB lesson is simple: clean the data first, and segmentation stops being a project and becomes a one-line filter.
Step 2: stand up reusable, safe outreach motions
Rather than one-off blasts, the AI built repeatable campaign pipelines from components that had already been tested:
- EOFY re-engagement. A value-led, low-pressure note to dormant customers: what have you got going out before end of financial year?
- Google review requests. A soft ask to happy, genuinely-served clients.
- Second-touch follow-ups for both, targeting only the people who had not responded.
Every motion shipped with the same guardrails baked in. If you want the productised, ongoing version of this rather than a one-off, that is our demand reactivation service.
Step 3: run the campaigns with full tracking
Sends were logged per recipient. Replies were matched back automatically and sorted into genuine reply, auto-reply, or bounced, so we knew exactly who engaged. The same enrichment now feeds straight into Gibson’s CRM integrations for every future campaign.
Is AI email outreach just spam? The safeguards
No, and the difference is the guardrails. This is what separates a careful AI outreach motion from a spam blast, and what keeps it on the right side of Australia’s Spam Act guidance:
- Dry-run by default. Every send previews first. A human approves before anything goes out.
- No double-contact. Deduplication is account-aware, so a client emailed at one address is not hit again at another.
- Bounce-aware. Addresses that bounced are held back until corrected.
- Output-guarded. Every email body passes an automated safety and quality check before sending.
- No review-gating. Review asks comply with Google’s review policy. We ask, we never pressure or condition.
- Audit trail. Every classification and every send is logged.
Results: what AI CRM automation delivered in one session
Starting from an unusable CRM, in a single working session:
- Accounts with a usable vertical went from around 28 of 113 (many wrong) to 112 of 113, corrected.
- Time to segment a group like “accountants” went from manual and slow to one filter, in seconds.
- Outreach went from ad-hoc to two multi-touch campaigns, tracked end to end.
Campaign output across the EOFY and review programs: 66 EOFY re-engagement emails (first touch), then 18 review-request follow-ups and 60 EOFY second-touch follow-ups, which is 78 in one morning, all deduped and logged. Responses were auto-categorised, bounced addresses were flagged for fixing, and on-leave auto-replies were separated from real answers.
The AI did not just send emails, says Albert Triolo of Gibson Promotions. It did the unglamorous work that usually blocks marketing (cleaning and structuring the CRM), then turned that into a repeatable outreach engine. One person, one conversation, a fully-segmented CRM and around 150 personalised, tracked customer touches, with a human in control at every send. That is what AI under human control looks like for an Australian SMB.
AI versus manual CRM cleanup and outreach
The honest comparison, task by task:
- CRM tagging. Hours of eyeballing records by hand, versus 113 accounts classified in one pass.
- Segmentation. Rebuilt from scratch every time, versus a one-line filter you reuse.
- Outreach safety. Easy to double-contact or email a bounced address, versus account-aware deduplication and bounce-aware by default.
- Approval. Often send-and-hope, versus a human approving every send on a dry-run default.
- Audit. Patchy, versus every classification and send logged.
Why it matters
The same enrichment now powers every future campaign. Real estate, hospitality, trades, accounting and more are each a single filter away. The work was directed in ordinary sentences, and the system handled classification, drafting, sending and tracking from end to end, with Albert approving the sends. It is the same discipline behind everything Gibson builds: clean, structured data in, useful and accountable output out. If you have not read it yet, the argument for getting the data right first is in AI doesn’t ‘just work’, and the broader sorting exercise is in what to automate first.
Get a once-off AI + Data Assessment
We plan it, structure your data, and run safe, tracked outreach on your own CRM, with a human approving every send.
Frequently asked questions
Can AI really clean up a messy CRM?
Yes. In this case an AI operator classified all 113 business accounts in a single pass, updating 87 records, correcting 3 mis-tagged ones, and logging every change for audit. Segmentation that was previously impossible became a one-line filter.
How does AI segment CRM customers by industry?
It reads each account's name, contact email domains and deal history, and checks the company website for the genuinely ambiguous ones. It then maps each account to a native CRM industry value rather than inventing new categories, so the filters stay clean and reusable.
Is AI email outreach just spam?
Not when it is guarded. Here every send was dry-run and human-approved, deduplicated at the account level, bounce-aware, quality-checked before sending, and fully logged. That is the opposite of an unmonitored blast.
How do you ask customers for Google reviews without breaking policy?
By asking happy, genuinely-served clients with a soft, unconditional request. No incentives, no pressure, and no gating the ask behind a positive rating. That keeps it compliant with Google's review policy.
What tools were used?
Zoho CRM and Zoho Books for the customer and finance data, and Claude as the AI operator, directed in plain English. No spreadsheets and no manual data entry were involved.
How long did it take?
A single working session. The full CRM cleanup plus around 150 personalised, tracked outreach touches across two multi-touch campaigns, roughly 78 of them sent in one morning.
