AI for Sales: Use Cases, Agents and the Data Foundation They Require

Nine in 10 sales teams are using AI agents or expect to within two years, according to Salesforce’s 2026 State of Sales research. Yet the data those agents are likely to encounter may not be ready for a larger operational role. In Validity’s 2025 survey of 602 CRM users and stakeholders, 76% said less than half of their organization’s CRM data was accurate and complete, while 37% reported losing revenue as a direct result of poor data quality.

Together, the findings point to a growing mismatch between AI ambition and the information available to support it. Incomplete CRM data may have limited impact if reps and managers supplement manual workflows with their own working knowledge of accounts and deal context. But as AI takes on more sales work, it can’t rely on that unrecorded knowledge. Gaps in the system can then influence which accounts receive attention, how outreach is framed, whether an opportunity is labeled correctly and what revenue leaders can expect to close.

AI can help sales teams interpret more of the context behind an account or deal, including conversations, engagement, product activity and commercial progress. It can also prepare content, update systems and coordinate work across the revenue process. The quality of those results depends on whether the underlying data is connected, current and governed, and whether the system has enough business context to use it appropriately.

AI in sales is the use of predictive AI, generative AI and agentic systems to interpret sales data, support decisions and carry work through revenue workflows.

• Predictive AI estimates what’s likely to happen: It identifies patterns in account, activity and opportunity data to score leads, estimate conversion probability, identify deal risk and forecast revenue.
• Generative AI turns sales context into valuable materials: It produces account research, personalized outreach, conversation summaries, proposals, coaching feedback and other content using the information and instructions supplied to it.
• Agentic AI coordinates multistep work: It retrieves information, selects among approved tools and continues a sales task within the permissions and operating limits set by the application.

These capabilities often work together. A predictive model might identify an account with a high likelihood of converting, while a generative system prepares research and an initial message grounded in the account’s industry, recent activity and prior interactions. An agent could then record the approved outreach, monitor for a response and coordinate the next permitted action.

Watch Anomalo’s Jonathan Karon as he explains how to implement automated data quality within Snowflake:

Prospecting begins with a resource-allocation decision: which accounts and buyers warrant a seller’s attention now. AI can help teams make that decision using a broader and more current set of signals than a manually maintained lead list or static scoring model.

Predictive models evaluate patterns in account attributes, marketing engagement, product activity, previous interactions and historical sales outcomes. As new activity signals become available, the score can change to reflect the account’s current behavior.

AI can support several parts of prospecting and outreach:

• Lead and account scoring: Estimate the likelihood that a lead or account will qualify, convert or progress within a defined period.
• Buying-signal detection: Identify meaningful changes in engagement, product activity, organizational characteristics or other approved data.
• Account research: Collect and summarize relevant information from internal records and permitted external sources.
• Buying-group analysis: Connect contacts, roles and interactions to help reps understand who may influence the purchase.
• Outreach preparation: Draft messages using the account’s situation, prior interactions and the objective of the communication.
• Next-best-action recommendations: Rank possible actions based on the account’s stage, recent activity and applicable sales rules.

Each of these applications depends on current, connected account context. When signals are incomplete, stale or attached to the wrong account, the workflow can prioritize the wrong prospect, misread interest or recommend an action that doesn’t fit the situation. Teams should also define which information is permitted to be used for prospecting.

Once an opportunity is active, sales work produces a continuous stream of information. Buyers ask questions, raise objections, introduce stakeholders, revise requirements and make commitments. Reps respond with product information, pricing, timelines and follow-up actions. Much of that context appears in meetings and messages, and typically doesn’t immediately make it into the CRM system.

AI can turn those interactions into a more complete and immediately usable opportunity history:

• Meeting preparation: Summarize the account, opportunity, recent activity, open questions and previous commitments before a call.
• Real-time assistance: Retrieve approved product, pricing, security or competitive information while a conversation is underway.
• Conversation summaries: Capture the main topics, buyer requirements, objections, decisions and next steps from a call or meeting.
• Follow-up recaps: Prepare customer-facing messages that reflect what was discussed and what each participant agreed to do.
• Proposal and quote preparation: Assemble initial material from approved product, pricing and account information.
• Opportunity updates: Suggest changes to CRM fields based on evidence from recent interactions.
• Sales coaching: Identify patterns across conversations that managers and reps can examine together.

Together, these capabilities reduce the amount of deal context that is lost between conversations, systems and people. Information from a meeting can inform the follow-up, update the opportunity record, prepare the next interaction and give managers a more current view of how the deal is progressing. That continuity is especially valuable in complex sales cycles, where several stakeholders may participate and weeks may pass between interactions.

A pipeline report reflects what’s been recorded about each opportunity: its value, stage, probability, expected close date and forecast category. AI can compare those fields with a wider body of evidence, including recent activity, stakeholder participation, product usage, proposal status, contract progress and the way similar opportunities developed historically.

Common applications include:

• Opportunity scoring: Estimate the probability that a deal will close or reach another defined outcome.
• Deal-risk detection: Identify stalled activity, slipping dates, missing stakeholders or other patterns associated with weaker outcomes.
• Close-date estimation: Predict when an opportunity is likely to close based on its progression and current evidence.
• Pipeline inspection: Surface opportunities whose recorded stage or forecast category appears inconsistent with observed activity.
• Coverage analysis: Compare available pipeline with revenue targets, conversion patterns and the time remaining in the sales period.
• Scenario planning: Model how changes in conversion, deal timing, capacity or pipeline creation could affect expected revenue.
• Revenue forecasting: Estimate future bookings or revenue using pipeline, historical performance and relevant finance data.
• Forecast explanation: Show which opportunities and signals contributed most to the prediction or change.

Forecasting accuracy depends on consistent business definitions across sales, finance and revenue operations. Pipeline value, stage, close date and forecast category may look standardized in the CRM, but teams often apply them differently across regions, products or managers.

Teams therefore need governed definitions for the fields and outcomes used in forecasting, along with a record of when those definitions change. The training data, model output and forecast reporting should all use the same semantic context so the prediction is evaluated against the business outcome it was actually designed to estimate.

Traditional sales automation uses predefined conditions to trigger specific actions. A workflow might assign a lead based on territory, for example, or create a task when an opportunity has no recent activity. These systems are effective for repeatable processes, but they don’t adapt easily when the appropriate action depends on other factors.

AI agents can carry sales work across several systems and steps, adjusting what they do as new account, opportunity and buyer information becomes available. When an opportunity stalls, for example, the appropriate response depends on what has happened. The buyer may be waiting for a security answer, a decision-maker may have stopped participating, the proposal may have expired or the CRM record may simply be out of date. An agent can use that context to determine whether to retrieve missing information, prepare a follow-up, update the opportunity or route the issue to the appropriate person.

Sales agents can support work such as:

• Prospect qualification: Research an account, evaluate it against approved criteria and prepare the supporting evidence for review.
• Meeting preparation: Assemble account history, recent activity, opportunity status and unresolved questions.
• Follow-up coordination: Prepare a recap, create assigned tasks and update approved opportunity fields after a meeting.
• Pipeline maintenance: Identify incomplete or inconsistent records and request or propose corrections.
• Deal inspection: Examine activity, stakeholder engagement, open commitments and commercial progress before a pipeline review.
• Renewal and expansion workflows: Monitor account signals, identify relevant opportunities and coordinate approved outreach.
• Next-best-action workflows: Recommend or carry out a permitted action using the account’s current context.
• Forecast-review preparation: Investigate material changes and assemble the evidence managers need to evaluate them.

Sales teams can use these capabilities at different levels of autonomy. A copilot supports an interactive task in a limited way: A rep might ask it to summarize an account, for example, then manually review the result and decide what happens next.

An agent can take action through more of the workflow, using approved tools to retrieve information, update a system, create a task or route work for approval. As the system takes on more of the work, teams need to define what it’s permitted to do without review and where human approval is required.

The records used by AI in sales applications must stay connected and retain consistent business meanings as they move through workflows. Reliable and responsible AI systems also depend on clear limits around data access, customer-facing claims and the actions an agent may take.

As Baris Gultekin, Snowflake’s VP of Product, AI, points out, “AI agents are a major leap from traditional automation or chatbots, but in order to deploy them at scale, businesses need an AI-ready information ecosystem.”

Connect account, activity and revenue data

Sales decisions draw from CRM records, marketing engagement, sales activity, meeting transcripts, product usage, support interactions, quotes, contracts and finance systems. Each source captures part of the commercial relationship, often using different identifiers and refresh schedules.

A connected account history gives models and applications a consistent way to associate activity with the correct customer, contact, opportunity or buying group. Without that connection, an application may score an account without factoring in an active support issue, or summarize an opportunity without visibility into the latest proposal.

Identity resolution is particularly important when several contacts participate in the purchase. The system needs to distinguish individual people while understanding their relationship with the same account and opportunity. It should also preserve changes over time, such as a contact moving to a new company or an opportunity being divided into several purchasing motions.

Data freshness should match the decision. Annual territory planning may use periodically refreshed account information, for example, while live opportunity assistance requires recent interactions, current pricing and the latest deal status. Teams should monitor missing records, delayed updates and changes in source coverage because each can alter outputs without an obvious application error.

Standardize sales and revenue definitions

Terminology needs clear, consistent definitions across applications and teams. For example, deal value may refer to annual contract value, total contract value or the revenue expected during the current period. Sales, finance and revenue operations teams need to agree on what each term means so models, reports and agents use the right definition for the task.

Semantic context makes those distinctions explicit through governed metrics, views, metadata and relationships among accounts, contacts, opportunities, products and revenue events. It can preserve stage-entry criteria, forecast-category rules, territory assignments, qualification standards and the relationship between bookings and recognized revenue.

Important definitions should have a named owner who approves changes and resolves differences across sales, finance, marketing and revenue operations. When the qualification standard or forecast process changes, teams should also determine whether historical records still represent the same outcome. Otherwise, the model may treat two operational definitions as though they were one consistent label.

Ground generated outputs in approved sales information

Sales applications use both account-specific context and organizational source material. Account context explains the buyer’s situation and the history of the opportunity. Organization-approved source material establishes what can be said about its products, prices, security practices, implementation requirements and contractual terms.

Approved information is especially important in customer-facing content. When the available sources don’t answer a buyer’s question, the output should leave the issue unresolved for a rep or subject-matter expert rather than supply a likely-sounding answer.

Source information needs an owner, version and effective period where appropriate. Retrieval should prioritize current material while retaining enough history to explain commitments already made during the deal.

Enforce access and permitted use

Sales workflows contain personal information, account strategy, call recordings, pricing, product usage and contractual details. Those records remain subject to access and permitted-use requirements when an AI application retrieves them or combines them into a new output.

Role-based access controls (RBAC) limit which users and service identities can view customer-level or commercially sensitive information. Classification and masking protect sensitive fields, while territory, account-team and deal-room permissions may further restrict who can access a specific opportunity.

The application also needs to respect the purpose and terms under which data was collected. Approval to use information for customer support or product analytics does not automatically permit the organization to use it for sales targeting or other purposes. Contracts, regional privacy requirements and internal policies can narrow how activity data and call recordings may be analyzed.

Audit records should show which data the application accessed, which policies applied and how the output entered the sales workflow. This becomes especially important when a recommendation affects account prioritization, a customer receives generated communication or an agent modifies a system of record.

Limit agent permissions and customer-facing actions

Agentic sales workflows need action-level permissions in addition to data access controls.

The workflow should define:

• Which accounts, opportunities and data sources the agent may access
• Which CRM fields it may propose or update
• Which communication channels it may use
• Whether outbound messages require rep approval
• Which product, pricing or contractual statements it may include
• Which commercial changes require manager, finance or legal approval
• When the agent must stop, ask for clarification or escalate the task

Service identities, tool permissions, approval thresholds and stop conditions define the agent’s operating boundary. The application should receive only the data and tools required for the assigned task, rather than broad access to the sales stack. Teams also need traces showing which information the agent retrieved, which tools it called and which actions it completed.

These controls preserve human accountability for customer relationships and consequential commercial decisions while allowing the agent to carry more of the research, coordination and recordkeeping around them.

Evaluate outputs against sales outcomes

Reliable evaluation starts with a data set that reflects the sales work the system will encounter, including common cases, difficult cases and situations where the correct response is to stop or escalate. Expected outputs, permitted actions and acceptance criteria should be defined before testing so the system is measured against a consistent standard rather than reviewer impressions.

The evaluation method then depends on the capability. Predictive models require measures such as calibration, ranking quality and performance across relevant segments. Generative systems need tests for factual accuracy, grounding in approved sources, completeness and adherence to required language. Agents must also be evaluated on the path they take: whether they select the appropriate tools, use the correct data, remain within their permissions, recover from errors and satisfy the task’s stop conditions.

However, component-level results aren’t enough when several models, retrieval steps and tools contribute to one workflow. Teams should also run end-to-end evaluations that capture whether context survives between steps and whether an early error changes a later decision or action.

After deployment, production traces, correction rates, escalations and outcome data help show whether performance is changing as sales data, products and operating rules evolve. Those observations should feed back into the evaluation set so testing continues to represent the work the system performs.

Monitor data quality and bias

Sales data reflects both buyer behavior and the organization’s previous decisions. Historical outcomes are shaped by which accounts entered the CRM, how territories were assigned, which leads received follow-up and how consistently reps recorded their activity. A model trained on those records may learn the organization’s earlier allocation of attention along with patterns of customer demand.

Incomplete activity creates another distortion. A successful rep who records only major milestones may appear to progress deals with fewer interactions than a rep who documents every call and message. Accounts managed through informal channels may look less engaged than accounts whose activity is captured automatically.

Addressing these issues begins with current, governed source data and documented outcome definitions, but it also requires active testing. Teams should compare performance across territories, account groups, acquisition sources and other relevant segments; examine whether missing data changes the prediction; and review how model recommendations affect the future data used for training.

When a system produces uneven results, teams should investigate the records, labels, decision rules and operating process that produced them. Human review at the end of the workflow won’t correct a scoring or prioritization pattern that repeatedly determines which accounts receive attention in the first place.

Snowflake gives sales, revenue operations, data and finance teams a shared environment for working with CRM, activity, product, support and revenue data, and applying predictive, generative and agentic AI without creating a separate data architecture for each sales workflow.

Snowflake ML supports predictive models built and run on governed data in Snowflake. Classification can support workloads such as lead, account or opportunity scoring by learning patterns associated with defined historical outcomes. Forecasting functions can predict future numeric values from time-series data, while broader capabilities enable developing and operating custom models when the workflow requires them.

Because the model runs where the underlying data already lives, teams can update scores and predictions from connected CRM, activity and product data without maintaining a separate scoring copy. The organization still determines the target outcome, training data, evaluation criteria and how the prediction enters the sales process.

Cortex AI Functions apply language and multimodal models to information stored in Snowflake. Sales and data teams can use these functions to summarize meeting transcripts, classify objections, extract requirements or action items and analyze collections of notes or documents through SQL-based data workflows. Those outputs can be joined with structured account and opportunity data for reporting, coaching or downstream applications.

For workflows that need to investigate information and continue across several steps, Cortex Agents can combine Cortex Analyst for structured data, Cortex Search for unstructured information and custom tools connected to an approved business process. A sales application could use those capabilities to examine an opportunity, retrieve relevant interactions and documents, prepare a recommendation and route the resulting work to another system.

Access to Cortex Agents is governed through Snowflake role-based access control, including privileges on the agent and the objects its tools use. The roles, tools and permissions configured for the application determine which information the agent can retrieve and which actions it can perform. Agent monitoring records can provide additional visibility into requests and underlying tool use.

Snowflake Horizon Catalog carries governance across the data and AI workloads behind these applications. Classification, masking, role-based access control and lineage remain connected to the underlying account, conversation and revenue data, helping organizations control who and what may access sensitive sales information.

Together, these capabilities allow organizations to apply AI across sales use cases while keeping the data, business definitions and policies beneath those workflows consistent.