AI-ready CRM data: the minimum for 5 use cases

Before putting AI and agents on top of the CRM, the company needs to secure the data the chosen use case will read, not the whole database. An account summary needs companies without duplicates and activities associated with the right record; a sales forecast needs reliable dates and written stage criteria; a prospecting agent needs current job titles and a record of who does not want to be contacted. Without that minimum, AI amplifies the database that exists and repeats the error in every record it touches, with percentages, charts and the language of analysis. In a Deloitte survey of 2,770 executives in 14 countries, conducted between May and June 2024, 55% of organizations said they had avoided certain generative AI use cases because of data-related issues (Deloitte, 2024).

Gartner found a similar picture among 432 leaders in six countries in the fourth quarter of 2024: data availability and quality rank among the top barriers to AI implementation for 34% of leaders at organizations with low AI maturity and for 29% of leaders at the most mature ones (Gartner, 2025). In a commercial operation, that data lives in the CRM, and the leaders of sales, marketing and CS decide what goes into it.

Why does AI amplify the database you already have?

AI applied to the CRM reads and summarizes records, learns patterns from history or takes actions based on them. In all three cases, the language model, the software that interprets and writes text, knows about your operation only what the database tells it. If the same company appears in three records, the account summary reads one of them and ignores the tickets logged in the other two. If the team pushes the close date every month, the forecast learns that no deal closes on its date.

Scale changes the size of the damage. In a hypothetical example, a follow-up agent receives 400 open proposals. If 15% of them repeat a company already in the queue, 60 customers get two messages on the same morning; if 30% have an overdue close date because nobody updated it, 120 deals enter the queue as stalled because of a logging error, and the rep spends the week fixing drafts.

The discipline floor for the analytical layer is in AI in sales, marketing and CS, and the routine that keeps the database clean is in CRM data quality. This article covers what each use case reads and how leadership decides whether the database is ready to turn it on.

What data does each AI use case require?

The five use cases leadership asks for most often require different things from the database. The last column shows how much each one depends on months of well-kept records, and it guides the rollout order.

Use case	What the AI does	Minimum data	Risk if missing	Dependence on history
Account summary	Summarizes the company’s status, history and next steps	Companies without duplicates, contacts with current titles, activities and notes on the right company, deals with stage and amount	The manager walks into the renewal without knowing about the critical ticket logged on another copy of the company	Low
Lead prioritization	Ranks leads by their chance of becoming an opportunity	Source, segment and size as picklists, the outcome of each lead (qualified, or disqualified with a reason), engagement	The model learns the SDR’s habits instead of the buyer’s profile	High
Sales forecast	Estimates how much the company will close in the period	Reliable creation, stage and close dates, written stage criteria, updated amount, loss reason as a picklist	Leadership decides on hiring and budget based on pushed dates	High
Service agent	Classifies, answers or routes tickets	Updated knowledge base, closed categories, the customer’s product and service level, ticket history	The agent answers with the old policy or treats a strategic customer like a regular account	Medium
Prospecting agent	Researches accounts and writes or sends the first outreach	Written ideal customer profile, contacts without duplicates, current title, customer status, opt-out and legal basis	The agent approaches an active customer as a prospect or writes to someone who asked not to be contacted	Medium

Use cases that read text and use cases that learn from history

The account summary and the service agent work on what is written in the record and in the knowledge base today, and fixing what they read takes weeks: merging duplicates among active accounts, associating activities and updating product documentation. Lead prioritization and forecasting learn from the past, and no cleanup sprint fixes the past. If loss reasons were picked out of habit over the last two years, a reliable history only exists after a few sales cycles under the new rule, and until then lead scoring works better with rules written by leadership.

Use cases that act outside the company

The prospecting agent and the agent that replies to customers write to people outside the company, and their errors reach the market before they reach leadership. For these two, the minimum includes fields that no sales report checks, such as active customer status, account owner and opt-out record. An agent that finds those fields empty treats the gap as permission.

Activity history and notes: the context the AI reads

The activity history gathers the emails, meetings, calls and messages logged in the CRM and associated with a contact, a company or a deal. For the AI, it is worth more than the fields, because it carries what the customer said, what the rep promised and where the conversation stopped. When the negotiation happens in the rep’s messaging app or in an email account with no integration, the CRM shows a silent deal that is in fact at the contract stage. Association matters as much as logging: the meeting saved on the contact disappears from the deal summary, and the note made on the duplicate copy of the company disappears from the account summary.

The quality of the note decides what the AI can do with it. In the teams I work with, the most common meeting note says the customer liked it and will think about it, and no AI can extract the next step or the decision maker from that sentence. I propose a short standard after every relevant meeting: what the customer said about the need, who decides and when, the open objection and the next step with a date. With meetings transcribed, the note records the rep’s reading of the conversation.

Structured fields or free text: what goes where?

A structured field accepts only defined values, such as a picklist, a date or a number; free text is open typing, like a note. AI reads free text much better than traditional reports, and some companies conclude they can drop the picklists. AI needs the structured field to choose which records to read, to trigger an action at the right moment and so someone can measure the result later, and free text provides the context of each case.

Information	Where to store it	Why
Segment, size and source	Picklist	AI and reports filter and compare without interpreting
Loss reason	Picklist and a short note	The list feeds the analysis; the note explains the case
Next step	Date in a field and description in text	The date triggers the agent; the text gives the subject of the message
Stated pain and objections	Note or transcript	AI reads the customer’s words well, and a list created too early impoverishes the information
Opt-out and legal basis	Required structured field	The agent needs a yes or no rule

AI also helps move information from text into fields. An agent can read the transcript and suggest segment, decision maker and next step for the rep to confirm, which improves the database while delivering a result to the team. The definition of each field, with allowed values and owner, belongs in the CRM data model.

Permissions and LGPD: what the agent can read and where it can write

A permission is the rule that defines which records and fields a user, or an agent, can see and change. Many companies connect the agent with an admin user, and it starts seeing finance’s note on overdue payments and the discount negotiated in another business unit. An account summary generated with that access and handed to an SDR shows them what the CRM’s access policy was hiding. Give each agent the scope of its task: read only the objects and fields the use case needs and write in as few places as possible.

Brazil’s data protection law, the LGPD, reinforces the rule. Article 6, item III, of Law 13,709/2018 sets the principle of necessity, the “limitation of processing to the minimum necessary to achieve its purposes”. In a commercial database, this means separating financial or sensitive data into restricted fields, telling the team not to record the buyer’s health or personal life in open notes and keeping the opt-out field filled in. Legal basis, international transfer and review of automated decisions are covered in AI and LGPD; I write as an operations consultant, and every choice goes through the data protection officer and the legal team.

How do you measure whether the database is ready for AI?

Readiness is measured by use case. The checklist gathers the criteria I bring to the decision meeting, with how to check them in the CRM itself and a reference target that leadership adjusts to the size of the database. The chosen use case must pass the criteria that apply to it, and the others can wait.

Criterion	How to check in the CRM	Reference target	Use cases that depend on it	Owner
Companies without duplicates	Companies sharing a tax ID or domain	Duplicates among active accounts resolved	All	Database owner
Contacts with current titles	Sample of active accounts reviewed by the owner	Decision makers checked this quarter	Summary, prospecting	CS and sales leadership
Associated activities	Open deals with no activity in 30 days	Email and calendar integration for the whole team	Summary, follow-up	CRM administrator
Stage and close dates	Open deals with an overdue date	Small share, reviewed every week	Forecast, follow-up	Sales manager
Written stage criteria	Document with entry criteria	Published and enforced in the pipeline meeting	Forecast, prioritization	Sales leadership
Loss reason as a picklist	Fill rate and weight of “other”	List in use for one sales cycle	Prioritization, forecast	Database owner
Knowledge base	Date of each article’s last review	Current products and policies reviewed	Service	Support coordination
Opt-out	Field respected in send lists	Required and integrated with sending	Prospecting, nurturing	Data protection officer and marketing
Agent permissions	Agent profile with objects and fields listed	Scoped access, no admin	All	IT
Test with real accounts	The 20-account test	Errors with a known cause, under control	All	Use case owner

The test closes the checklist because reports show the empty field and hide the field filled in wrong.

Where do you start when the database is not ready?

A messy database delays the predictive layer, and the company can start with uses that work on the text of the moment. Meeting summaries and transcripts use the audio of the conversation. A follow-up draft written from a fresh transcript depends little on history. Classifying new tickets decides based on what the customer wrote, with closed categories and the knowledge base support already maintains. These uses deliver gains within weeks and teach the team to review what the machine writes.

A second group uses AI to fix the database itself: the agent that suggests field values from notes, assisted duplicate search so the database owner can decide on merges and the review of loss reasons saved as “other”. In the diagnostics I run, this group convinces leadership faster than any presentation, because it shows on screen the size of the problem and the pace of the fix.

Start with text-based uses, prepare the fields for the next use case in parallel, turn on the account summary and the service agent when their checklist closes and leave prioritization and forecasting until after a few cycles under the new rule. Whoever reverses this order hears from leadership that AI does not work, when the problem was the data that went in. Who approves each use case and with what data is defined in AI governance.

Common mistakes when putting AI on top of the CRM

Waiting for the whole database to be clean

The company starts a months-long cleanup before testing anything and loses leadership’s sponsorship halfway through. Fix the fields of the chosen use case, turn it on with human approval and leave the rest to the hygiene routine.

Starting with the sales forecast

The forecast is the use case leadership wants most and the one that depends most on history. Turned on over pushed dates, it misses in the first quarter and burns trust in the whole project.

Connecting the agent as an admin

The shortcut exposes data the access policy was hiding and lets the agent write to fields nobody reviews. Create a dedicated profile for each agent and review access when the scope changes.

Treating the note as a dumping ground

The team starts logging everything as free text because it heard that AI understands, and the database loses the fields that make it possible to filter, trigger and measure. Keep picklists for what needs to be counted and the note standard for what needs to be understood.

Blaming the model for the data’s error

The summary comes out wrong, the team switches tools, and the next one gets it wrong in the same place. Before switching, classify the errors by cause, as in the 20-account test.

Who owns what: roles and responsibilities

Without a named owner, the preparation sits between IT, which configures, and the commercial teams, which log. The table summarizes the design I bring to the first meeting.

Role	Accountable for	Approves
Sponsoring leadership	Use case, indicator and deadline	Turning on each use case after the checklist
Area leader	Use case result and the team’s own logging	What the agent does alone and what waits for approval
Database owner or RevOps	Duplicates, picklists, imports and checklist	New picklist values and loads into the database
CRM administrator and IT	Integrations, agent profiles and permissions	Each agent’s technical access
Data protection officer and legal	Legal basis, sensitive fields and opt-out	The personal data each agent reads
Team that logs	Notes in the standard, associated activities and stage fields	The fields suggested by the AI

The approval and logging design for each agent is in AI agents.

Training: who needs to learn what

The rule becomes routine when each role understands what the AI does with what they log. Reps and SDRs learn the note standard, activity association and how to review suggested fields, in the company’s own CRM. Managers learn to read the checklist and to enforce, in the pipeline meeting, the fields that feed the use case, and whoever administers the CRM learns to configure agent profiles. Leadership goes through training first, because it decides the order of use cases and sustains the enforcement of logging.

30-day roadmap for leadership

1. Week 1: choose one use case. Pick one with low or medium dependence on history and an indicator the area already tracks.
2. Week 1: name the owners. Define the use case leader, the database owner and the contact at the data protection office.
3. Week 2: run the checklist. Check the criteria that apply to the use case and record where the database falls below target.
4. Week 2: run the 20-account test. Give the summaries to the account owners and classify errors by cause.
5. Week 3: fix only what the use case reads. Resolve duplicates among the accounts involved, associate activities and publish the note standard.
6. Week 3: scope the permissions. Create the agent profile and validate the personal data with the data protection officer.
7. Week 4: turn it on with human approval. Roll the use case out to a small team and measure corrections, approval without edits and the area’s indicator.
8. Week 4: decide the next use case. Choose the second use case and schedule the preparation of its fields.

This roadmap depends on leadership deciding the order of use cases and on the team logging what the AI will read, and that is the work of the in-company applied AI training. I run the training on the company’s own process and CRM, with groups for leadership, for those who log and for those who administer the system. The company leaves with a prompt library by area and an agent architecture that defines what each agent reads, what it executes and who approves the rest.

Frequently asked questions

Do I need to clean the whole CRM database before using AI?

No. Fix the records and fields the chosen use case will read, such as active accounts for an account summary, and leave the rest to the hygiene routine. A full cleanup before any test tends to use up leadership’s sponsorship before the first result.

What data does AI need to forecast sales?

Reliable creation, stage change and close dates, written entry criteria for each stage, updated amounts and loss reasons as a picklist. It also needs complete sales cycles logged under those rules, which is why forecasting tends to be one of the last use cases to go live.

Can AI use the notes and emails in the CRM?

Yes, and they are the richest context for summaries and drafts. Activities need to be associated with the right deal and company, and the note needs to record need, decision maker, objection and next step instead of a generic impression of the meeting.

Which AI use case works with little CRM data?

Meeting summaries and transcripts, follow-up drafts based on the transcribed conversation and classification of new tickets work with little history. The agent that suggests field values from notes also improves the database while the team uses it.

How do I know if the database is ready for an AI agent?

Run the readiness checklist on the criteria that apply to the use case and do the 20-account test with the account owners. If most errors come from missing or wrong data, the database is not ready for that agent yet.

How do I prepare the team to log the data AI will use?

Train by role, in the company’s own CRM, showing each person what the AI does with the field and the note they log. Leadership goes through training first to enforce the standard in the pipeline meeting, and the evolution of the readiness checklist shows whether behavior changed.