How to Use Salesforce AI: A Practical Walkthrough

The 71% Problem: Why Sales Teams Lose Time to Admin Work

Imagine hiring a salesperson and then telling them they can only sell for about a third of their day. That's effectively what's happening across the industry: according to Salesforce's research, salespeople spend 71% of their time on non-selling tasks — drafting emails, scheduling meetings, updating records, and chasing data entry instead of closing deals.

The math is uncomfortable. If your top rep works a ten-hour day, fewer than three of those hours involve actual selling conversations. The rest disappears into the administrative scaffolding that surrounds every deal: follow-up emails, CRM hygiene, meeting coordination, and pipeline documentation.

This is precisely the gap that AI agents are built to close. Rather than asking reps to work harder, teams are delegating the repetitive scaffolding to autonomous systems that can draft outreach, create to-dos, and send meeting requests without human involvement. The highest-frequency sales actions are exactly the ones machines handle best.

The results are hard to ignore. Salesforce's data shows that 83% of sales teams using AI saw revenue growth in the past year, compared to just 66% of teams without it. Adoption is accelerating accordingly:

• 9 in 10 sales teams already use AI agents or expect to within two years
• 88% of reps working with agents say AI increases their odds of hitting sales targets
• Agent creation among first-mover companies grew 119% between January and June 2025

The shift is also changing how companies measure success. Salesforce has moved from counting logins to tracking Agentic Work Units — actual completed tasks — which reframes AI investment around work done rather than seats sold.

For teams evaluating which agent capabilities to adopt first, independent benchmarks matter. SkillzBench is building a free directory that security-scans, validates, and benchmarks agent skills in a sandbox environment, so you can see which ones genuinely lift performance before deploying them against your CRM workflow.

The 71% problem isn't a productivity lecture — it's an opportunity. Every administrative hour a well-designed agent absorbs is an hour returned to pipeline. The teams that reclaim those hours first are the ones already outgrowing the field.

Einstein vs. Agentforce: Knowing Which Salesforce AI Does What

Most teams don't have an AI problem in Salesforce — they have a matching problem. Einstein and Agentforce sound interchangeable in a demo, but they do fundamentally different jobs, and mixing them up wastes money and momentum.

Einstein is the predictive layer: a "silent advisor" that analyzes historical data to score leads, forecast outcomes, and surface patterns, as Synebo's comparison of the two systems explains. It doesn't send emails or update records — it quietly makes your data smarter so humans can act on it.

Agentforce is the autonomous actor. Powered by the Atlas Reasoning Engine and grounded in unified Data Cloud context, it executes live tasks end to end: drafting and sending outreach, creating to-dos, booking meetings, and updating CRM records. That action layer matters because salespeople still spend 71% of their time on non-selling tasks, and autonomous execution is what reclaims those hours.

Yana Chekan, Head of Delivery at Synebo, offers the cleanest rule of thumb: use Einstein to improve how your teams work with data, and use Agentforce to keep license usage lean, accurate, and cost-efficient. Match the tool to the job:

• Lead scoring and forecasting → Einstein, which learns from historical patterns
• Sending emails, creating to-dos, booking meetings → Agentforce, which executes in real time
• Summarizing call transcripts or account history → Einstein, as an analytical layer
• Updating records or routing escalations autonomously → Agentforce

The split is already proving itself at scale. Agent creation among early adopters grew 119% between January and June 2025, and agent-led service conversations grew 22x in the same period, according to Salesforce's Agentic Enterprise Index. Teams that deploy the right tool for the right task — predictive insight in one lane, autonomous action in the other — get compounding returns instead of overlapping spend.

If you're building out agent workflows beyond Salesforce, SkillzBench is building a free directory and MCP connector for testing agent skills before you trust them in production.

The Practical Walkthrough: Five Steps to Deploy Salesforce AI for Sales

Deploying Salesforce AI for sales follows a deceptively simple logic: fix your data first, automate the repetitive stuff second, and keep humans close to the deals that matter. Skip any of these steps and the agents you build will flounder in exactly the "last mile" problem Salesforce's own leadership warns about.

Step 1: Unify your data in Data Cloud before you activate anything. As Salesforce President Miguel Milano notes, the last mile is hard because companies need context — data, metadata, and deterministic workflows. Agentforce's Atlas Reasoning Engine is powerful, but it can only reason over what it can see. Pull structured CRM records and unstructured artifacts (call transcripts, email threads, PDFs) into Data Cloud first, or your agents will hallucinate their way through deals.

Step 2: Automate the top three high-frequency sales actions. Salespeople spend 71% of their time on non-selling tasks, which makes administrative automation the highest-leverage starting point. Configure agents to handle:

• Drafting and sending follow-up emails based on meeting context
• Creating and prioritizing to-dos from conversation summaries
• Sending meeting requests and handling scheduling logistics

These three actions recur constantly across every pipeline stage, so automating them frees reps to spend their actual selling time selling.

Step 3: Design human-in-the-loop escalation triggers. The most successful deployments use a hybrid model where agents handle initial contact and routine queries, then route complex negotiations and high-value deals to humans. Notably, escalation rates rose from 22% to 32% between Q1 and Q2 2025 — not because agents failed, but because they got better at identifying which conversations need a human closer. Define explicit triggers: deal size thresholds, negotiation language, competitive mentions.

Step 4: Measure results in Agentic Work Units, not logins. The old seat-based metric tells you nothing about whether AI is working. Agentic Work Units (AWUs) count completed tasks — emails sent, records updated, meetings booked — which is the metric that actually justifies AI spend to a CFO. Track AWUs weekly and correlate them with pipeline velocity.

Step 5: Plan for iterative post-launch refinement. Demetri Salvaggio of Engine noted that after their agent launched in 12 days, "the real work was just starting." Budget for continuous testing and tailoring of agent actions post-launch; the gap between a working agent and a great one is almost entirely iteration.

If you're building custom agent skills to extend these workflows beyond Salesforce's native actions, SkillzBench's free skills directory lets you browse, benchmark, and equip verified SKILL.md packages to any agent — with trust scores and measured lift, free.

What Good Results Look Like: Real Outcomes and the Hybrid Model

So what does success actually look like once the agents are live? The honest answer, based on Salesforce's own reporting, is a mix of impressive gains and one counterintuitive trend that surprises most teams.

Start with the vendor-reported outcomes, with the caveat that these come from Salesforce's newsroom and partner coverage rather than independent audits. Publishing company Wiley reported a 213% ROI and a 40%+ increase in case resolution within the first weeks of deployment. Travel platform Engine cut average handle time by 15% while autonomously resolving more than 500,000 inquiries annually. And 1-800Accountant says its agent now resolves up to 60% of incoming requests, like tax return status questions, without human involvement.

• Volume absorbed: Williams-Sonoma's "Olive" agent handles roughly 60% of website conversations, freeing staff for higher-value work.
• Speed gained: The IRS Office of Chief Counsel cut tax court case openings from 10 days to 30 minutes, saving up to 98% of manual activities.
• Pipeline built: Salesforce itself used Agentforce to contact 130,000 leads and create 3,200 pipeline opportunities in four months.

Now the counterintuitive part. Between Q1 and Q2 of 2025, escalation rates to human agents rose from 22% to 32%. That sounds like failure, but it is the opposite. Agents got better at recognizing when a customer needed a human expert and routed them faster. The lesson for sales leaders: a rising escalation rate during early deployment is often a signal that routing is improving, not that the AI is struggling.

This is why the hybrid human-agent model has become the standard rather than full automation. Agents handle initial contact, routine queries, and data summarization; humans own complex negotiations and high-value deals. Design your escalation triggers deliberately instead of treating them as an afterthought.

Two practical notes temper the hype. First, Engine's agent was built in 12 days, but as Senior Director Demetri Salvaggio put it, "the real work was just starting" — post-launch testing and tailoring drive the results above. Second, measure completed work, not logins: the industry's shift to Agentic Work Units reframes ROI conversations from "how many licenses did we buy?" to "how much work did AI actually do?" — a distinction that matters when justifying spend to a CFO.

If you go on to extend these agents with third-party skills and tools, vetting matters — published research has found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building free benchmarking, security scanning, and trust scores for agent skills. Treat results as directional, plan for iteration, and keep humans in the loop where judgment counts.

Going Further: Extending Your Salesforce AI Workflow with Agent Skills

Once your Salesforce agents handle email drafting, meeting requests, and record updates, the obvious question is what to automate next. The work surrounding your CRM — enriching lead lists, summarizing call transcripts, prepping outreach sequences — rarely lives inside Salesforce itself, and that's where agent skills come in.

Agent skills are SKILL.md instruction packages that teach coding agents like Claude Code, Codex, Cursor, and Gemini CLI (plus chat-app agents) how to handle specific tasks. Instead of prompting from scratch each time, you install a skill once and the agent applies it consistently. The productivity case is strong: Salesforce research shows salespeople spend 71% of their time on non-selling tasks, so anything that automates adjacent busywork compounds the gains you get from Agentforce.

The catch is trust. Published research found that roughly 26% of published skills carry at least one vulnerability, and most directories do nothing to verify what a skill actually does before you point it at customer data. Installing an unvetted skill on an agent with CRM access is a real security decision, not a convenience.

That's the problem SkillzBench is being built to solve, with an evidence-based pipeline for evaluating skills before you equip an agent:

• Discover — continuously crawl and dedupe skills from GitHub so you're not guessing which package is canonical
• Verify — pin every version to a commit SHA, then license-check, security-scan, and quality-judge it
• Benchmark — test each skill in a sandbox with the same model, with and without the skill, and publish the measured Lift
• Equip — connect any agent via a free MCP server, toggling skills and tools per agent so it only sees what you chose
• Stay current — track upstream changes, re-scan, re-benchmark, and alert you when something shifts

The benchmarking approach mirrors what the best Salesforce deployments already do. As Engine's Demetri Salvaggio noted after launching their agent, the build was fast but "the real work was just starting" — continuous testing is what produced results like a 15% reduction in handle time. The same discipline applies to skills: measure the lift, verify the security, and keep watching after launch.

SkillzBench is currently in pre-launch, and everything — the directory, the MCP connector, benchmarks, and alerts — will be free, with trust scores kept strictly separate from any commercial influence. If you're ready to extend your Salesforce AI workflow beyond the CRM walls, browse the skills directory and join the waitlist at skillzbench.com.