AI in your CRM needs a customer record you may not have

The AI features being wired into CRM platforms read a structured customer record. If your customer data is a spreadsheet, a mailbox and a WhatsApp thread, there is nothing for them to read, and no subscription fixes that. Choosing a CRM for a small business in South Africa is now a decision about what you can write down, not about which assistant you can talk to.

That is the useful reading of an announcement that will otherwise be forwarded as a reason to feel behind.

What was actually announced

Salesforce and Anthropic have announced Claudeforce, and the first product under it is Salesforce in Claude, a plugin that brings live CRM data and Salesforce workflows into Claude. It carries 37 pre-built sales skills, covering meeting preparation, deal health reviews and pipeline analysis.

Note how far along it is, because the headlines do not. At announcement it sat with selected pilot customers, with an open beta due in September and skills for other business functions after that. Nobody is buying this on Monday.

Google is doing the recognisable version of the same thing from another direction, taking Gemini Enterprise into financial and legal workflows with an emphasis on governance and data controls. Its financial services version carries more than 50 skills and 13 connectors to data providers. The legal one covers contract review, due diligence and regulatory monitoring.

Two different vendors, one pattern. The model is being pointed at business records that were already structured, and every capability on the list assumes those records exist and are current.

What it costs in rands

No rand figure is published, and that is worth saying rather than estimating.

What is published is the shape, and it is two meters rather than one. Pricing is consumption-based against your Salesforce user licence edition, and customers contract separately with Salesforce for API consumption and with Anthropic for Claude inference. Two dollar-denominated bills, both moving with usage, both re-pricing in rands every month.

That is the currency problem every cloud tool carries here, doubled, and it arrives before anybody has agreed what a unit of consumption is.

The bigger number is not the licence anyway. It is the work of getting your customer data into a form the software can read, and that cost is yours regardless of which vendor eventually wins.

What it means for a business with 40 staff

Nothing this quarter, and something quite important underneath.

A 40-person distributor in Germiston is not buying Salesforce, and an announcement about 37 sales skills is not addressed to it. What is addressed to it is the prerequisite, because the same requirement applies to the cheap tools that will carry these features within a year.

The requirement is a customer record. Not a CRM strategy. A record, with the same customer appearing once, identified in a way that does not change, with the things you know about them attached to that identity rather than to a person's memory.

Most businesses at this size have the opposite. The customer exists in the accounting package as an account name, and in a spreadsheet as a slightly different name. They exist in a mailbox as an email address, and in WhatsApp as a phone number with no surname. Four records, one customer, no join.

The minimum viable customer record

This is the part worth taking away, because it is buildable in a week and costs a licence fee of nothing.

One identity per customer, and one place that says which identity is correct. In most South African businesses the accounting package is the honest choice, because it already has to be right for invoicing.

The fields that earn their place are few.

• Registered name and trading name, because they differ and a purchase order will use one of them
• Company registration and VAT number, which are what make a customer unique
• One billing address and one delivery address
• A named contact, with a direct number and an email address that belongs to the company rather than to a person's phone
• Payment terms and current credit status

Then two things almost nobody records. What they bought, in a form you can group, rather than a free-text invoice line. And the last real interaction, with a date, so the business can tell dormant from lost.

That is it. A record like that is legible to a person, legible to a report, and legible to any model that gets pointed at it later.

What blocks it here

POPIA arrives the moment the record becomes useful. A structured customer record with contacts, numbers and buying history is personal information about identifiable people, and the responsibility for that processing stays with your business when a model reads it. Ask any vendor where the data sits, who may read it, and what leaves the country.

The second block is practical. Cleaning duplicates is unglamorous work that nobody has time for, and it is the entire project. Businesses that skip it discover that their pipeline analysis is analysing the same customer three times.

Joining the accounting package, the mailbox and the quoting tool so one identity holds, and so a new customer is created once rather than four times, is exactly the seam Zorah works on. It is worth doing whether or not any AI feature ever gets bought, which is the test of a good integration.

What to do on Monday

Pick your third largest customer and count how many places their details exist in your business. Accounting package, spreadsheet, mailbox, phone, quoting template.

Then check whether the name matches in all of them, and whether anybody could tell you what that customer bought in the last six months without opening the invoices one at a time. If the answer is no, you have found the real prerequisite for every CRM feature being announced this year. A distributor running trade pricing tiers off a customer record that disagrees with itself is already paying for this, in the wrong price, today.