AI in Sales and Marketing

The customer work AI can improve

Marketing connects understanding customer needs with developing, communicating, and delivering valuable offers. Its scope includes research and decisions about what to offer, not just promotion. Sales develops particular buying conversations: understanding requirements, explaining suitability, and working toward an agreement. AI assistance should therefore begin with the decision or customer need being served, rather than the material it can generate. The American Marketing Association's definition provides this broader orientation. [1][2]

Three records recur throughout this work. A lead represents a prospective buyer not yet sufficiently qualified. An account represents a managed customer organization or commercial relationship. An opportunity represents potential business under consideration. These are useful distinctions, not universal database tables: systems differ in how they represent and connect them. [3]

Business-to-business (B2B) selling serves organizations; business-to-consumer (B2C) selling serves individuals buying for personal use. A direct consumer purchase may need a clear explanation and straightforward purchase path. An organizational purchase may involve a buying committee: several people evaluating different aspects of the same decision. A finance participant and an operational user need not want the same explanation. These are contrasting patterns, not rules about every purchase; either relationship can involve repeated contact and future business. [4]

A funnel groups customers into broad categories such as awareness or consideration. A customer journey describes their actual interactions. Someone may return to documentation after a sales meeting or reconsider an earlier option; two people in the same funnel category may have taken different paths. Use the funnel for classification without forcing the journey into a one-way sequence. Likewise, keep faster preparation, a more useful customer interaction, and a stronger commercial result separate: each requires a different observation. [7]

What changed and what endured

AI enters a field already concerned with measuring response, understanding needs, and allocating attention. The following developments addressed different parts of that work. Their sequence does not imply that one replaced another.

Predictive assistance estimates an outcome, such as whether a person will respond. Generative assistance produces material, such as an explanation or draft. Neither settles what the customer needs or whether contacting them is useful. Rules can determine eligibility, predictions can prioritize investigation, generated text can prepare a response, and people can resolve commercial uncertainty within the same workflow. Customer understanding remains the foundation for all four. [10][11]

Research needs without inventing facts

Research should begin with the decision it must inform: which problem to address, what prevents purchase, or what a seller still needs to understand. Voice-of-the-customer research identifies and organizes needs from customers' accounts. Counting recurring themes is useful for navigation, but frequency is not automatically importance. In Griffin and Hauser's study of a portable food-carrying product, highly rated needs were not more likely to appear in interviews. An AI summary that counts mentions does not resolve that distinction. [9]

Win-loss analysis investigates why buyers purchased or declined. It adds a perspective that internal sales notes may miss. Clozd's practitioner discussion distinguishes buyer interviews from call recordings and internal deal-reason fields. When those sources disagree, preserve the disagreement for investigation rather than silently choosing the internal explanation. Interviews reveal expressed motivations; sales conversations reveal discussed requirements; support records expose problems after purchase. None alone represents every prospective customer. [12]

Account research examines a particular organization and its buying situation. Public information can establish an announced initiative or a person's listed role; it does not establish an undisclosed budget or personal purchase intent. A useful brief keeps sourced facts, attributed opinions, interpretations, and missing information visibly separate. It should also identify whose experiences are absent—for example, people who never entered a sales conversation. [13][12]

Retrieval-augmented generation (RAG) composes an answer using selected external information. In research, its role is to connect findings to material a reviewer can inspect; Construct warranted claims explains the support boundary. In Building Alice's Brain, the team describes using lead information to plan retrieval of relevant seller knowledge and synthesize answers for messaging. This illustrates contextual preparation, not a demonstrated increase in sales. Judge such assistance by the usefulness of the resulting brief and total preparation time, including verification. [14]

Synthetic consumers are models prompted to answer as specified customer personas—descriptions of customers with particular characteristics. Their responses can suggest research questions, but are not customer testimony. Maier and colleagues' 2025 preprint converted generated answers into purchase-intent ratings by comparing them with reference statements. Across 57 surveys, this method matched human rating distributions more closely than asking models for numerical ratings directly. However, the reference statements had been manually optimized on those same surveys; performance on new surveys remained unresolved, and some demographic patterns were not reproduced consistently. The study measured stated intent, not purchases. Simulated responses therefore do not replace independent buyer research or establish demand. [15]

Keep people, accounts, and facts distinct

Customer relationship management (CRM) is the practice and software for managing customer interactions and commercial records. A contact represents a person; an interaction event records something that happened. Record ownership identifies who maintains or acts on the record. Keeping these separate from accounts and opportunities prevents a sent email from being mistaken for buying progress. [3]

A person can have several organizational relationships, and those relationships change. Salesforce documents multiple-account contact relationships, including current and past associations. A useful design records the relationship relevant to each interaction, rather than interpreting every old event through the person's latest employer. This historical association is a design obligation, not a guarantee supplied by a contact-to-account link. [16]

First-party data comes from people's direct interactions with the organization's products or services. Externally acquired information has a different origin. Enrichment adds or updates information from additional sources, but does not certify it. The buyer-intelligence workflow combines providers and deduplicates their results; its speaker also identifies inaccurate enrichment as a cause of inaccurate communication. Duplicate people, shared addresses, and uncertain affiliations need explicit handling. Duplicates depend on identity develops that work. [17][13]

Designate a system of record for each important fact: the application treated as authoritative for that field. Preserve its source, relevant dates, and unresolved conflicts; store generated interpretations separately from accepted facts. Name the authoritative records explains the ownership decision. Implementation paths matter: HubSpot's enrichment documentation describes fill and overwrite settings, yet conversational and workflow enrichment can overwrite properties despite the mapping's overwrite rule. A setting on one importer therefore does not establish protection across every writer. Refresh schedules likewise do not guarantee accuracy. [18]

Connected records make account-wide engagement visible without requiring a particular graph database. Their value comes from preserving which person, organization, interaction, and potential deal each assertion concerns. Those connections supply context for the next decision; they do not themselves grant permission to act. [19]

Establish current contact eligibility

Permission to read a record is different from permission to analyze it, personalize with it, or contact its subject. Consent is agreement to a specified use; preferences express communication choices. Suppression retains the information needed to exclude a recipient from particular contact. Frequency limits restrict repeated contact over an interval. Begin with the intended purpose, recipient, channel, source, and applicable time. Specify the intended use explains why availability alone is insufficient. [20][21][22]

Requirements differ by jurisdiction and message. Under the UK's Privacy and Electronic Communications Regulations (PECR), unsolicited electronic marketing to individual subscribers generally requires consent or a qualifying soft opt-in—an exception subject to specific conditions. Corporate subscribers have different rules; sole traders can count as individual subscribers. Public contact details do not establish consent, and the soft opt-in does not apply to bought-in lists. The ICO's guidance also explains why suppression records should be checked to prevent further unwanted contact. [21]

The US CAN-SPAM guidance instead describes commercial-email obligations including accurate sender information and subject lines, advertising identification, a postal address, and an opt-out mechanism. Outsourcing sending does not remove the promoted business's responsibility. An existing customer relationship does not automatically make a message transactional: its primary purpose matters. These email rules are not permission for calls, texts, or every use of personal data. [23]

Platform policy adds another boundary. For the personalized-ad features covered by Google's data-use policy, first-party audience creation is allowed while creating targeting audiences from third-party data is prohibited. First-party origin still does not settle consent or other applicable obligations. [17]

Selection can become stale while a message waits. Klaviyo documents that suppressed contacts may remain list members; frequency controls and scheduled-audience snapshots are separate mechanisms. Check current authoritative restrictions when sending, not just when exporting the audience. If a contact opts out after selection, block the queued marketing message while retaining the suppression information needed to prevent later sends. Respond to changed authority distinguishes stopping contact from correcting or deleting records. [22]

Repeated contact is also a relationship decision. Nanda Piersma and Jedid-Jah Jonker's 2000 mailing-frequency research distinguished selecting responders for one mailing from managing contact across successive periods. Its model-based simulations do not prescribe a modern sending frequency. The enduring distinction is useful: maximizing the next response and managing a continuing relationship are different objectives. [24]

Choose relevance before personalization

An audience is the set of people or accounts considered for communication. Segmentation groups records by relevant differences; targeting chooses which groups to pursue. Start with inspectable conditions rather than an unexplained generated list. HubSpot's filter documentation illustrates AND/OR conditions over records and their associations. A match establishes that stored values meet the rule, not that the values are correct. Keep missing values visible so an unknown industry is not silently treated as a known unsuitable industry. [25]

An ideal customer profile (ICP) describes account types the offering is expected to serve well. Treat it as a revisable hypothesis, not a fact about every person in a company. Account-based marketing coordinates attention around selected organizations and their buying participants. Account suitability and an individual's information needs remain distinct. [4]

Keep four dimensions separate: fit with the offering, observed engagement, readiness to buy, and current contact eligibility. HubSpot's scoring model distinguishes fit properties from engagement events. Neither is automatically a purchase probability. Repeated visits can indicate research without an active purchase; an excellent-fit account may have no present need. Combining everything into one score can hide the reason for an action. [26][27]

A relevant message can still make someone uncomfortable about how their information was obtained. In Aguirre and colleagues' 2015 study, participants imagined a Facebook interaction about a car loan and viewed an advertisement. Greater personalization increased stated click intention when participants were aware of information collection, but not significantly when collection was concealed. Under concealed collection, greater personalization also increased reported vulnerability. These were reactions to a scenario, not observed clicks or purchases. Assess both the relevance of the message and the recipient's experience of data use; additional personal detail does not necessarily help. [28]

Use only the personal information needed for the purpose, rather than inferring sensitive circumstances to decorate a message. Revisit segments as facts change, and examine who the selection rules exclude: feedback from contacted people alone cannot describe everyone omitted. Choosing product items within an experience is a related but distinct problem, developed in Recommendation Systems. [20][29]

Prepare content for a commercial purpose

A campaign coordinates marketing actions and assets for an audience and objective. It is an organizing unit, not a customer's journey or sales stage. Salesforce's campaign model groups related efforts, targeted people, and responses. Positioning states the offer's relevant distinction for that audience. A call to action (CTA) names the next step requested, such as examining documentation or arranging a demonstration. [30]

Write a brief before generating variants. It should identify the customer need, audience, offer, supported benefit, channel, destination, requested response, and constraints. This is a practical design contract: it separates choosing what to communicate from producing ways to express it. A model can draft an outline, shorten copy, or adapt material for a seller's conversation without being allowed to invent a capability or a customer need. [31][32]

Localization adapts material for a language or market; sales enablement prepares material that helps sellers explain and discuss an offer. Both require preserving meaning, not just fluent wording. Adobe's email workflow combines generated copy with approved content fragments, including product descriptions and regional disclaimers. Stable claims need not be regenerated every time their surrounding presentation changes. [32]

For this fictional reporting product, the approved statement is: “Pro supports CSV export of monthly reports.” Adapt the opening and presentation while preserving that plan restriction, format and report scope.

The destination must fulfill the message's promise: a format-guide CTA should reach the relevant guide, not a page offering a different capability. Review translations and variants against the same claim boundaries. Count their checking, updating, and retirement as production work when comparing assistance with the incumbent process. More variants are useful only when they serve a meaningful audience or communication difference. Image, audio, and video production methods belong in Generative Media. [33][34]

Expand ⤢

Review claims and authorize distribution

Brand voice is the organization's relatively stable communication style; tone adapts that style to the situation. Mailchimp's guide distinguishes them and prioritizes clarity over forced entertainment. Neither verifies truth. A message can sound entirely appropriate while making an unsupported product promise. [35]

Claim substantiation means having appropriate support for objective assertions. The US FTC policy requires a reasonable basis before dissemination; later evidence does not replace that obligation. Review comparisons, guarantees, prices, and discounts against their actual conditions. Testimonials require special care: FTC guidance explains that an AI avatar does not make a fabricated customer experience permissible. [33][36]

Reviewers need the proposed artifact, its support, the intended audience and destination, and the consequence of approval. Make review a consequential decision develops this requirement. Adobe's review interface supports comments, requested changes, and version comparison. Adopt an explicit rule that material changes return to the relevant reviewer; do not assume the presence of version history automatically enforces it. [34]

Approval of a reusable fragment does not settle whether a new combination is appropriate. Content approval also does not override suppression or grant budget authority. Enforce resource- and operation-specific permissions where the action occurs, rather than treating a model's proposed action as authorization. [32][37]

Review must remain smaller than the work it replaces. The buyer-intelligence talk describes how drafts requiring extensive editing can make approval queues a chore that sellers stop using. Track editing effort and queue age, and improve the source material or drafting task when reviewers repeatedly rewrite outputs. Assign an owner who can pause future distribution and coordinate corrections to material already published. Stopping scheduled work cannot retract an email already sent. [39][40]

Qualify the next useful conversation

Qualification decides whether and how to pursue a potential relationship. It should identify a useful next action, not merely assign an impressive number. A marketing-qualified lead (MQL) is one marketing considers ready for sales; a sales-qualified lead (SQL) is one sales considers a potential customer. These are locally configured categories. Recording either status does not independently establish purchasing authority, a committed budget, or an accepted deal. [41]

BANT organizes discovery around budget, authority, need, and timeline: affordability, purchasing participants, the problem to solve, and timing. Salesforce's account also cautions against rigid checklist use. Early conversations may not have settled budgets or timelines, and committee purchases may have no single decision maker. An unknown budget is missing information, not evidence of no budget. Use the framework to guide discovery rather than manufacture certainty. [42]

AI can assemble these observations and prepare discovery topics. Using agents to build an agent company describes combining lead responses, customer records, and industry research into use cases and meeting preparation. Microsoft's qualification-agent documentation likewise makes target profiles, knowledge sources, and qualification criteria administrator responsibilities. Its disclosed research and generated-email evaluations do not establish incremental revenue. [5][43]

Evaluate both directions of error: weakly supported leads consume seller attention, while rejected relevant buyers disappear from the pursued pipeline. Historical sales outcomes are shaped by whom sellers previously contacted and how they treated them. A score predicting those outcomes does not establish who would benefit from a new intervention. Association, action, and causal effect explains that distinction. Preserve the observations behind qualification so a seller can accept, revise, or challenge it. [26][44]

Carry conversations into owned work

Inbound engagement starts with the customer; outbound engagement starts with the business. Nurturing is relevant follow-up while a decision develops. A sales development representative (SDR) develops and qualifies early conversations before further sales work. Assistance can support these activities by answering product questions, preparing meetings, summarizing commitments, and drafting follow-ups. It should not turn every interaction into another automated message. [43][6]

Source visibility makes review more useful. Microsoft's sales-email workflow distinguishes customer-record information in suggestions, lets sellers change the associated opportunity, and supports meeting-based follow-ups. The seller reviews, inserts, edits, and sends. This is a concrete separation between preparing a response and committing it externally. Booking a meeting, changing a record, or offering a discount requires its own authority; a good draft grants none of those permissions. [45][37]

Follow-up needs stopping conditions. HubSpot documents configurable stops after replies or meetings, including an option to stop other contacts at the same company; out-of-office replies generally behave differently. A reply trigger does not classify sentiment. Refusals and opt-outs should stop the relevant contact, while contradictory facts or requests beyond authority should route to someone able to resolve them. Deferral can remain a legitimate disposition: configurable follow-up leads illustrate that disqualification need not mean permanent exclusion, without proving renewed interest or permission. [40][46]

A handoff transfers responsibility and the context needed to continue. One practical design is a work item containing the customer's need, supporting product facts, qualification observations, preferences, unresolved commitment, receiving seller, and acceptance deadline. Record acceptance separately from notification, and completion separately from acceptance. For example, a request to confirm an integration requirement remains open when the seller receives it. Acceptance establishes who will investigate; completion requires a confirmed answer and communication back to the customer. Transfer responsibility explicitly explains how to manage these separate obligations. [47][6]

People may respond after the originating session ends. Preserve the work's progress and unresolved obligations across that wait; Direct work across long waits covers execution mechanics. Faster replies are useful only alongside accurate answers, useful meetings, and completed commitments. These distinctions determine what to measure next. [48]

Measure what the work accomplished

A conversion is completion of a specified desired action. It might be a form submission or a purchase; always name which. The sales pipeline contains active potential business, whose recorded value is not realized revenue. An evaluation assesses work against its intended purpose. Define worthwhile improvement explains how to choose the baseline and criteria before treating a rising number as success. [49][2][50]

Attach a population and period to every rate. A useful-meeting rate among booked meetings diagnoses meeting quality; useful meetings per eligible account describes a different result. Preserve nonresponders and rejected leads in the appropriate population rather than reporting only successful paths. Leading indicators occur earlier, such as responses; lagging outcomes arrive later, such as purchases. Their order does not prove that improving the earlier measure improves the later one. [50]

Measurement machinery can change the apparent result. Mailchimp explains that security scanning, privacy processing, and previews can generate opens or clicks without a recipient acting. Filtering is incomplete, and a filtering change can alter reported engagement without changing people. Inspect complaints, unsubscribes, bounces, and downstream actions alongside engagement. [51]

Delivery has continuing consequences too. Gmail's sender guidance explains that spam reports can reduce domain reputation and affect subsequent delivery. It calls for authentication, monitoring, and volume adjustments when messages bounce or are deferred. These provider requirements neither establish contact permission nor guarantee inbox placement. [52]

An outcome that has not yet arrived is not necessarily a failure. Delayed-conversion research distinguishes nonconversion from conversion that will occur after the current observation cutoff. Define a maturity window, retain unresolved cases, and distinguish observed results from modeled projections. Google's lift reporting explicitly includes modeled linkage and, for certain studies, projected delayed conversions. Account for incomplete feedback develops the broader problem. Measure preparation, checking, correction, and seller effort over the same scope as the outcomes. [53][54]

Attribute credit without claiming causation

A touchpoint is a recorded interaction. Attribution assigns outcome credit to touchpoints or channels under a chosen rule or model. An attribution window limits which interactions are considered. First-touch assigns credit to the first interaction, last-touch to the last, and linear multi-touch divides it equally. HubSpot's reporting documentation illustrates these different allocation rules. [55][54]

Reconstruction comes before allocation. Match interactions to the right people and accounts, join available online and offline records, and avoid counting one purchase several times. Several committee members can produce different interaction paths to the same deal. Missing calls, incomplete integrations, and uncertain identities leave gaps; widening the attribution window does not recover an event that was never recorded. [19][55]

AI can help organize competing interpretations of a recorded journey. In How Juries and Librarians Can Solve GTM's AI Trust Problem, Alex Bauer describes independent analysts producing cited attribution opinions and a judge weighing their reasoning. This structures the judgment about credit, but agreement cannot establish what would have happened without the communication—the counterfactual. A purchase can receive a well-supported attribution even if it would have happened anyway. Attributed revenue is not necessarily additional revenue. Estimating the additional effect requires a comparison with a credible alternative. [56]

Estimate the additional business effect

Incrementality is the difference an intervention makes relative to a credible alternative. A control group receives that alternative; a holdout is withheld from the particular intervention under study. Conversion lift measures additional defined conversions through such a comparison. Google's conversion-lift overview describes advertising eligibility versus a holdout. It does not automatically isolate AI's contribution. Choose the live experiment covers the general methods. [49]

Targeting complicates observational comparisons because people likely to buy may also be more likely to receive advertising. Gordon and colleagues' 2019 study, A Comparison of Approaches to Advertising Measurement, compared observational estimates with 15 randomized Facebook advertising experiments. Observational approaches often failed to recover the experimental effects despite extensive customer information. This does not prove that every observational method fails; it shows why comparing exposed buyers with unexposed people can be misleading. Assignment in those experiments made treatment users eligible for the focal advertising, rather than guaranteeing exposure. Preserve that distinction when measuring actual reach. [57][58]

Contamination occurs when an intervention reaches a control group. In B2B work, assigning a whole account together can be appropriate when participants share purchasing information. Account membership alone does not prove this dependence. Shared sellers may reuse assisted material across accounts, requiring consideration of seller or team boundaries. Cluster-randomization guidance recommends the lowest assignment level that adequately addresses contamination while retaining enough independent groups; analysis must account for the grouping. [59]

Capacity creates another connection. If both groups share constrained sellers, extra demand from assisted accounts can reduce service available to controls. Research on operational dosage explains how intervention effects depend on available service capacity. Applied to sales, trial results describe the tested staffing and queue conditions, not automatically a larger rollout. [60]

A bounded AI result illustrates why outcome definitions matter. Lu Fang and colleagues' 2025 preprint, Generative AI and Firm Productivity, reports field experiments at a cross-border retailer. In its push-message experiment, generated messages were compared with incumbent content. Conversion—placing at least one order—increased by 3.0%, while the estimated 1.6% expenditure increase was not statistically significant. These are relative changes, not percentage-point changes. The analysis covered the first day, and generated messages reached roughly 40% of the treatment group. The result neither establishes zero expenditure effect nor proves profit gains, long-term retention, or B2B sales improvement. [61]

Specify whether the target is near-term orders, mature deals, repeat purchases, or contribution after costs. Keep observation periods comparable and distinguish unfinished follow-up from completed negative outcomes. Also consider cannibalization: an intervention may shift purchases from another channel or time rather than add them. A discount may increase orders while reducing the contribution of each. These are reasons to define the business outcome broadly enough for the decision, not assumptions that such effects occurred in a particular experiment. [53][62][57]

Expand the work that creates value

Return to the actual bottleneck. If sellers lack verified preparation, better research may help. If accurate drafts wait for review, more generation may not. Compare the incumbent process, a simpler template or rule, and AI assistance on the same eligible work. An overloaded queue may already be handled by rules rather than individual expert attention; that real alternative belongs in the comparison. [63][39]

Contribution margin is revenue less variable costs; the remainder covers fixed expenses before contributing to profit. It is not total rollout value. OpenStax's explanation makes that boundary explicit. For an operating decision, also account for integration, data services, media, review, seller effort, correction, and ongoing operation. Do not subtract a cost twice if it is already included in the contribution calculation. [62]

Customer acquisition cost (CAC) divides stated acquisition spending by new customers acquired over a stated boundary. Shopify's convention includes acquisition software and allocated labor as well as advertising, while separating retention spending. State your own allocation and timing rules. Observed CAC differs from cost per incremental acquisition: the latter needs an estimate of additional customers, not merely attributed customers. Lower CAC alone does not establish that those customers are profitable. [64][49]

Time released for other work is capacity, not automatically cash savings. Acquisition and retention are also different benefits: helping an existing customer does not create a new customer. AI Cost and Performance Engineering develops these accounting distinctions. In Using agents to build an agent company, automated marketing and qualification were followed by a substantial personal meeting workload. The anecdote illustrates a downstream obligation, not a causal estimate of profitable growth. [65]

Revenue operations (RevOps) coordinates processes and connected systems across marketing, sales, customer success, and finance. Its contribution is continuity, not simply passing more records between teams. Assign owners for customer records, qualification criteria, claim support, distribution, accepted handoffs, and outcome reporting. Use wins, losses, and deferrals to reconsider audience hypotheses and research knowledge. Updating a rule or knowledge base is not the same as training model parameters; preserve which change was made and assess its consequences. [47][29]

The strongest expansion case connects a real customer need to supported communication, permitted action, completed responsibility, and a credible improvement over the alternative. AI earns a larger role when that complete chain works—not when one stage produces more material.

Open questions

1. 01

   The long-term B2B value of combined research, qualification, and outreach assistance remains difficult to establish because deals mature slowly and seller capacity changes the delivered intervention. Progress would mean controlled comparisons that follow useful meetings through accepted opportunities and contribution, including review and downstream labor. [60][65]

2. 02

   Synthetic customer research still needs independent validation on new populations and decisions. Agreement with previously optimized survey targets does not establish demand forecasting; progress would require untouched surveys and behavioral outcomes, including groups whose responses are poorly reproduced. [15]

3. 03

   Adaptive qualification can repeatedly learn from the customers it already selects, leaving excluded buyers poorly understood. Progress would preserve selection history, investigate missed opportunities, and distinguish a better prediction of existing sales behavior from a better policy for allocating attention. [44][29]

4. 04

   Reliable sales handoffs require more than accurate summaries: someone must accept responsibility and fulfill the remaining commitment. Progress would connect customer context, acceptance, completion, and customer-visible resolution in operating evidence, rather than treating a notification as the endpoint. [47][48]