Half your CRM data is wrong: who writes to it and why AI there works only under supervision

On October 20 Salesforce announced it is buying Braze for about $5.66 billion, SaaStr put “half the data in your CRM is wrong” in a headline, and HubSpot and MarTech published guides on choosing a CRM and segmenting a database. For the owner of a service company or a car rental all of this news comes down to one question: who writes to the CRM and whether that data can be trusted. At the PDR company in the US every call lands in the CRM without a salesperson's hands, and below I explain why that matters more than the choice of platform.

The deal and what it means for an ordinary company

MarTech on October 20 looked at why Salesforce is buying Braze: for the speed between a signal from the customer and an action. Salesforce pays $47.50 per share, 58% above the previous close of $30.04. Braze has about 2,770 customers and quarterly revenue of $227.2 million. The author's questions are practical: where will the customer profile live, and will the teams get one profile or one more sync task.

SaaStr the same day worked out the multiples: Salesforce pays 6.8 times trailing 12-month revenue of $834 million, while the day before Braze traded at 3.7. Of the five observations one interests me. Braze's revenue from implementation services grew over the year from $8.3 to $19.6 million: the author explains it by AI scenarios needing manual data setup, and customers pay for it.

Half the data is wrong

SaaStr on October 25 retold an Aurasell session at AI Day. The headline has half the data wrong and 49% executive turnover, and inside there is a story about an aquarium: a prospecting agent wrote an email asking how the recipient's home aquarium was doing. The agent found it in a transcript of a meeting from a year ago in the company's own CRM. What matters to me in this story is the review step: before sending, the salesperson sees every plan for a contact, can correct the agent and confirms the send.

HubSpot on October 19, in a guide to AI CRMs, listed what counts as AI governance: access to AI features by role, review and approval before the result goes to the customer, a log of changes made by AI, and rules for personal data. HubSpot's second text from the same day is about choosing a CRM: build the criteria table before the first demo, otherwise the committee will fit the weights to whoever they liked. The main question for the vendor about the data model and integrations: what breaks after the first ten apps.

MarTech on October 20 on segmentation for the holidays: split the database by recency of purchase into four groups, send the oldest no more than three emails and in small batches. Old addresses bounce, there are spam traps among them, and emails to your best customers can end up in spam. Repeat purchases make up 5 to 30% of a brand's sales.

Example: a car rental and three numbers for one customer

Take a car rental. A customer booked a car in 2024 through the website, in 2025 called from a different number, and this year wrote on WhatsApp from a third. If a manager creates the contact, the CRM will have three cards, and segmentation by recency will show this person as new. If the call and the message land in the CRM through an integration with telephony and the website, the system looks for a match by number and email and merges the cards before the manager sees them. CRM data goes bad where a person enters it, so I change the input method first.

At our company every call goes into the CRM

At the PDR company in the US I have built the server, the telephony and the link to the CRM since 2021. After an incoming call to the IVR, the system creates a lead in GoHighLevel on its own and attaches the recording and an AI breakdown: who called, a lead, an existing customer or an insurer, and how the employee answered. The salesperson types nothing by hand, so the card appears with the number the call came from. The company used to have 20 people, today it has more than 30, every real lead goes through the funnel to the end, and conversion to sale grew by 30%.

My rule about supervised AI comes from the document parsers at the same company: after recognition, the data is checked by AI validators on Anthropic, OpenAI and Gemini services, and an employee confirms the result. About 99.9% of documents go through without manual correction, which frees up about 60 hours a week. I would launch an agent that writes to customers on its own only with the same confirmation step that Aurasell showed.

I connect different CRMs: AmoCRM, Bitrix24, HubSpot, Pipedrive, GoHighLevel and 1C. At the car rental in South Africa it all started in October 2024 with AmoCRM and Google Sheets, and there it became clear how data drifts apart once your own app appears next to the CRM: we had to decide in advance where the rules are executed. Who writes to the platform matters more than which platform it is.

What I would do

Count what share of contacts over the last month was created by hand and what share by integration. Move the manual ones to integration one channel at a time: telephony first, then the website.

Catch duplicates by number and email at the entry point. A quarterly report won't fix them.

Before segmenting, remove addresses with no reply in a year from the mailing, or send to them in small batches, as MarTech advises.

Give any AI in the CRM a role with limited rights and a log: what it changed, when and based on what input.

I wrote about duplicates, missing data and AI's place in the chain in the article how to reliably connect documents, CRM and tasks. Directions: a CRM connected to the team's work and the business as one system. Full projects: a service team raised conversion by 30% and car rental in South Africa with AmoCRM. Neighboring reviews: who called and why: AI call analysis and metrics that teach a team bad decisions. Earlier I wrote about who owns the company context in AI for CRM. If salespeople create contacts by hand and the reports diverge from reality, tell me where your inquiries come from. Discuss your task.