Customer-facing organizations are facing a widening capacity gap. Demand is increasing across channels while human-only workflows and fragmented data make it increasingly difficult for teams to respond quickly and consistently.
The U.S. Bureau of Labor Statistics projects that employment of customer service representatives will decline by 5% through 2034, even as organizations continue to manage growing volumes of customer inquiries. The trend suggests that automation, rather than continued expansion of customer service headcount, will increasingly play a role in closing the capacity gap.
Sales organizations face a similar challenge, particularly when it comes to responding quickly to new leads. A landmark Harvard Business Review study examining 2,241 U.S. companies found that the average organization took 42 hours to respond to a new sales lead. Companies that waited 24 hours or longer were more than 60 times less likely to qualify a lead than those that responded within the first hour. Although the study was published in 2011, its findings remain one of the most frequently cited demonstrations of the importance of speed-to-lead.
At the same time, consumers are becoming increasingly familiar with AI while remaining cautious about its reliability. Pew Research Center reports that roughly half of U.S. adults now regularly use AI chatbots, yet only 29% of those users say they trust the information these tools provide.
For small and midsize businesses, these trends create a difficult balancing act. Customers expect faster responses and more personalized experiences, but businesses cannot simply continue adding people to every workflow. AI agents offer a potential solution, but successful deployment requires more than connecting a model to a company’s systems.
Recent conversations hosted by Emerj with leaders from Salesforce provide a practical framework for understanding what SMBs need before introducing AI agents into customer-facing operations. Sharif Karmally, VP of SMB Product Marketing; Matt Kravitz, Head of Customer Transformation for Service Cloud; and Vanessa Tabbert, VP of Agentic Transformation and Sales Development, each approach the opportunity from a different perspective, but their insights converge on four critical principles: reliable data, carefully selected use cases, clear governance, and seamless workflow integration.
Building a Structured Data Foundation for Reliable Agent Behavior
For many SMBs, the biggest obstacle to successful AI adoption is not the AI model itself. It is the quality and organization of the information the model is expected to use.
Sharif Karmally frames data readiness as an operational foundation problem rather than simply a technical one. Businesses often rely on fragmented spreadsheets, stale CRM fields, disconnected inboxes, and inconsistent customer histories. When an AI agent is asked to operate across those sources, it may have access to a large amount of information without actually having reliable context.
That distinction becomes increasingly important as organizations move from basic AI assistance toward autonomous execution.
Matt Kravitz describes agent maturity as a progression. Early-stage agents can answer questions, while more advanced systems can access contextual business data and eventually take actions on behalf of employees. As autonomy increases, the importance of reliable context increases with it.
An agent that can answer a generic question does not necessarily need deep access to business systems. An agent that can modify a customer record, initiate a workflow, communicate with a prospect, or resolve a service issue does.
This creates a common problem for SMBs: attempting to deploy highly autonomous agents while their underlying data remains fragmented.
Kravitz’s maturity model highlights why access and actionability should develop alongside data maturity. Organizations cannot expect an agent to make reliable decisions when the information it receives is incomplete, inconsistent, or outdated.
Vanessa Tabbert provides another perspective based on the scale of inbound sales demand. Even with structured lead-management processes, her SDR organization could only prioritize a fraction of the opportunities coming into the business. Roughly three out of four inbound leads never reached a human representative because the volume of opportunities exceeded what the team could realistically engage.
The problem was not necessarily a lack of leads. It was a lack of capacity to process and act on them.
For SMBs, this distinction is critical. As customer interactions, leads, conversations, and service requests accumulate, businesses need systems capable of maintaining consistent context at scale.
Karmally describes the challenge this way:
“It’s not a single use case — it’s a map of your entire business. The business processes, the stages a customer goes through, that context is so much more than just the data sitting in a spreadsheet. I’ve seen this firsthand working with chief data officers at large enterprises: counterintuitively, the more data and context you add without structure, the worse the results get over time. It’s a phenomenon called context rot. A CRM is the best pre-built infrastructure for agents because it keeps everything unified and up to date, so agents don’t do things wrong.”
The underlying lesson is straightforward: AI agents amplify the quality of the environment in which they operate. If the data is structured, current, and governed, agents have a stronger foundation for reliable behavior. If the data is fragmented, AI can accelerate the problems already present in the organization.
Starting With High-Volume, Low-Risk Use Cases
Once the data foundation is in place, the next challenge is deciding where an AI agent should actually be deployed.
The strongest starting point is often not the most sophisticated workflow. Instead, SMBs can gain more from targeting repetitive, predictable processes that already consume significant amounts of employee time.
Kravitz argues that customer demand tends to cluster around a relatively small number of predictable questions and interactions. Many of these requests do not require complex business reasoning. They are repetitive, high-volume tasks that can place unnecessary pressure on human teams.
These workflows can provide an accessible starting point for AI adoption because they are easier to model, easier to monitor, and generally lower risk than complex decision-making processes.
Kravitz also cautions organizations against adopting AI simply because a particular capability is available. For example, summarization and generative responses may be useful, but they are not necessarily the highest-value starting point for every SMB. Businesses should instead identify the specific interactions that create the greatest operational burden or customer friction.
Tabbert’s experience demonstrates the potential of this approach in sales.
Her team began by addressing a portion of the sales funnel that had effectively been deprioritized: leads that arrived faster than SDRs could contact them. Rather than disrupting the team’s core sales motions, the agent was applied to opportunities that humans were already unable to cover.
The results provided a measurable proof point.
“When we launched, we took leads we previously would have done nothing with and booked 150 meetings in the first month alone. Once we tuned the agent based on what we were seeing, we went from booking 150 meetings in a month to booking 150 meetings in a single week, with the same quality and quantity of leads. That’s when I knew we were onto something.”
— Vanessa Tabbert, VP of Agent Transformation & Sales Development, Salesforce
The significance of this example goes beyond the number of meetings booked. The workflow was highly measurable and relatively straightforward to evaluate. The organization could compare opportunities that previously received little or no human attention with the results generated after introducing the agent.
This creates a useful model for SMB adoption. The best initial AI use case may be the workflow humans have already been forced to deprioritize because of volume.
For one organization, that could be unanswered sales leads. For another, it could be order-status requests, appointment scheduling, after-hours inquiries, repetitive customer questions, or follow-up communications.
The important factor is not simply whether AI can perform the task. The question is whether automating the task can reduce customer friction while producing measurable operational value without introducing significant risk.
Starting small also creates an opportunity for organizations to learn how their agents behave before expanding them into more consequential workflows.
Establishing Clear Boundaries for Autonomous Execution
As AI agents become capable of taking action, governance becomes increasingly important.
An agent that answers a question presents a different risk profile from one that changes customer records, adjusts pricing, approves requests, sends communications, or makes decisions that affect revenue.
Karmally emphasizes that SMBs cannot rely on informal processes or tribal knowledge once agents begin operating within real workflows. Organizations need clear boundaries defining what an agent can do, what it cannot do, and when a human must take over.
These boundaries should be treated as part of the agent’s operating environment rather than as restrictions added after deployment.
Routine updates and repetitive workflow actions may be appropriate for autonomous execution. Pricing decisions, approvals, sensitive account changes, and other business-critical actions may require explicit authorization. Decisions involving negotiation, exceptions, compliance, or significant financial consequences can also require human review.
Governance becomes even more important when multiple people or multiple agents are working with the same customer information.
Karmally explains:
“As soon as you have multiple people — or multiple agents — you need governance. You need guardrails for what an agent can do, and permissions for what agents or humans have access to and can edit. You need decision processes for when there’s disagreement, and human-in-the-loop checks for the most critical things.”
This approach positions agents more like junior digital teammates with clearly defined responsibilities. They can execute repetitive work within established boundaries while humans retain authority over decisions requiring judgment.
Tabbert reinforces this concept by comparing agent management to managing human employees. Agents are not simply deployed and forgotten. Their performance needs to be monitored, measured, tuned, and improved based on real-world outcomes.
That feedback loop is essential because even well-designed agents will encounter edge cases that were not apparent during initial development.
Kravitz adds another dimension through channel strategy. Businesses need to determine which customer interactions are appropriate for self-service, which digital agents should handle, and which still require direct human involvement.
The objective is not maximum automation. It is appropriate automation.
Embedding Agents Into Existing Workflows
Even a technically capable AI agent can fail to deliver value if employees have to change how they work to use it.
The Salesforce leaders repeatedly emphasize the importance of embedding AI into the environments teams already use. Instead of creating another destination that employees must remember to visit, agents should operate directly inside existing workflows.
For customer service teams, this could mean embedding agents into the service console. For sales teams, it could mean integrating them into CRM, email, voice, or other communication environments. For internal collaboration, agents can operate within platforms where employees already share information and coordinate work.
Karmally points to Slack as an example. When customer conversations, product discussions, and internal coordination already occur within a communication platform, embedding an agent into that environment can allow it to observe conversations, identify relevant CRM updates, and help maintain records without forcing employees to switch between applications.
Tabbert describes a similar dynamic within sales development. SDRs are less likely to embrace tools that introduce additional clicks or new behaviors. The agent succeeded because it extended existing workflows rather than requiring the team to create an entirely new process.
This allowed the agent to absorb after-hours calls, long-tail nurturing, and repetitive outreach while leaving the team’s existing operating rhythm largely intact.
The principle is simple: AI should fit the workflow rather than force the workflow to fit the AI.
Matt Kravitz describes the most advanced version of this concept as an augmented workspace in which the employee does not need to leave the primary console to obtain information or initiate an action.
“The third level is when you never leave the console. It’s not just that I can provide a contextual experience — the console itself is saying, ‘Hey, can I help, and can I execute actions on your behalf?’ It’s eavesdropping on the work and asking, ‘Can I check that order status? Can I cancel that for you?’ That’s really the maturity model: moving from a transactional console, to a contextual one, to one that’s augmented and can act on your behalf without you switching tools.”
— Matt Kravitz, Head of Customer Transformation, Service Cloud, Salesforce
This represents a broader shift in how enterprise software may evolve. Instead of employees constantly moving between applications to retrieve information and perform tasks, the workspace itself becomes capable of understanding context and assisting with execution.
The Path to Practical Agentic AI for SMBs
For SMBs, the transition to agentic AI does not have to begin with a complex enterprise-wide transformation.
The conversations with Salesforce leaders point toward a more practical sequence. Organizations should first establish reliable and unified customer data. They can then identify high-volume workflows where automation can reduce friction without introducing excessive risk. From there, they can establish clear permissions, decision boundaries, and human checkpoints before gradually expanding autonomous capabilities.
Finally, agents should be embedded into the tools and workflows employees already use.
The common thread across all four stages is operational discipline. AI agents are most effective when they are connected to trustworthy information, assigned clearly defined responsibilities, measured against meaningful outcomes, and integrated into the environments where work already happens.
For SMBs facing increasing customer and sales demand, this approach can provide a path to greater capacity without simply adding more people or more disconnected software.
The opportunity is not to automate everything. It is to give teams the ability to respond to more customers, engage more opportunities, and complete more work without sacrificing the judgment and human relationships that remain central to customer experience.
As AI agents move from experimentation into everyday business operations, the organizations most likely to succeed will be those that treat agentic AI not as a standalone technology purchase, but as an extension of their data, workflows, governance, and customer strategy.