How do I use AI in Salesforce?

The AI Adoption Gap in Salesforce: High Hopes, Uneven Results

Here's the tension every Salesforce admin and business leader is feeling right now: the AI hype is real, but so is the disappointment. On one side, Salesforce's SMB survey found that 91% of small and medium businesses using AI report revenue gains, and Agentforce reached 12,000 customers after its October 2025 general availability. On the other, IBM's independent survey of 1,200+ Salesforce customers tells a much harder story.

Only 33% of AI initiatives meet their ROI targets. A full 72% fail to scale across business units, and 20% stall, fail, or get abandoned entirely. As IBM's report puts it bluntly, for many Salesforce customers, AI is "delivering more invoices than insights."

So which picture is true? Both. Adding AI to Salesforce is genuinely easy to start — Salesforce markets a setup that takes "minutes," and the free entry point through Salesforce Foundations is real. What's hard is making it pay off. The gap between starting and succeeding comes down to factors the marketing pages don't emphasize: data quality, an integrated tech stack, governance, and realistic cost planning.

The confusion starts with the product names themselves. Salesforce ships two overlapping AI layers — Einstein, the predictive and generative AI embedded in the CRM interface since 2016, and Agentforce, the autonomous agent platform that can reason through multi-step tasks without human intervention. Many users struggle to tell them apart or figure out which one to use, and the honest answer is usually "both, for different jobs."

The cost picture is equally muddled. Salesforce says every customer can start with Agentforce for free, while practitioner estimates put a real production deployment — with the Data Cloud prerequisite and per-conversation fees — at $200K–$400K in the first year for a 50,000-conversation workload. Even Klarna publicly pulled back agent deployments after admitting cost weighed too heavily in its original decision.

The failure pattern is consistent enough to name:

• Weak data foundations — agents can only act on what they can read, and fragmented data limits them to isolated tasks.
• Missing governance — only 21% of surveyed customers strongly agree they have the governance they need.
• Unclear economics — 64% say unclear total cost of ownership makes it hard to judge whether agents save money at all.
• No measurement discipline — agents deployed without baselines, outcomes, or review cycles quietly drain budget.

None of this means AI in Salesforce isn't worth doing. Top performers in IBM's survey report 60% greater efficiency and more than double the pipeline expansion of their peers. The difference between them and the 72% isn't luck — it's approach. And the same discipline applies beyond Salesforce's native tools: whether you're extending agents with custom skills or connecting them to external systems, verifying what actually works before you deploy matters. That's the problem SkillzBench is building for — a free, independent directory that benchmarks agent skills for security, quality, and measurable lift, so teams don't have to guess which components will help and which will hurt.

Einstein vs. Agentforce: Choosing the Right AI for the Job

Ask a Salesforce admin which AI to use — Einstein or Agentforce — and the honest answer is usually "both, for different jobs." These two products are complementary, not competitors, and understanding the division of labor is the first step to getting real value from either.

Einstein is the AI embedded directly in the CRM interface, delivering predictive and generative features since 2016 — think AI-based lead scoring, next-best-action recommendations for reps, and auto-drafted prospecting emails. It's toggle-on assistance inside the tools your team already uses, with lower governance overhead. Per practitioner analysis, new 2026 deployments default to Einstein for embedded predictive AI while reserving Agentforce for autonomous workflows.

Agentforce, by contrast, is a layer on the Salesforce Platform for building and deploying autonomous AI agents. Its Atlas Reasoning Engine breaks a prompt into smaller tasks, evaluates each step, and proposes a plan until the action is complete — no pre-determined conversation trees, no human intervention required. According to Salesforce documentation, agents execute work by searching for data, analyzing it, and actioning a plan across systems via Flows, Apex, and MuleSoft connectors.

Where each fits:

• Einstein: lead scoring, embedded UI assistance, content generation, natural language search — fast wins with minimal setup.
• Agentforce out-of-box agents: Service Agent, SDR, Sales Coach, Campaign Optimizer, Merchandiser, and Buyer Agent cover common service, sales, and marketing jobs.
• Agentforce custom agents: multi-step workflows that chain across systems — one Fortune 500 finance team cut reporting time from 15 days to 35 minutes using specialized chained agents.
• Cross-system actions: agents integrated via MuleSoft, SAP, or Shopify augment rather than replace existing systems.

The results justify the split. Reddit's Agentforce deployment deflected 46% of support cases and cut resolution times by 84%, while a retail deployment reduced inventory redistribution from 10 days to one hour. One caveat: agents can only do what their allowlist includes, so scope Data Cloud and governance before scaling.

If you're extending agents beyond Salesforce, validated third-party agent skills can automate the rest of the workflow — SkillzBench is building a free, independently verified directory and MCP connector so you can browse, test, and equip skills with confidence.

How to Add AI to Your Salesforce Workflows: The Implementation Path

Here's the good news buried in the hype: you can start adding AI to Salesforce for free, and you may not need to build much at all. The bad news, per IBM's survey of 1,200+ Salesforce customers, is that only 33% of AI initiatives meet ROI targets and 72% fail to scale — so the path you take matters more than the tool you buy.

Start with the free on-ramp. Salesforce Foundations, a $0 add-on for Enterprise Edition and higher, includes credits for your first 1,000 Agentforce for Service conversations, letting you experiment before committing budget. Just don't mistake free entry for free production: practitioners estimate a 50,000-conversation first year can run $200K–$400K once conversation fees and prerequisites are counted.

Fix the data foundation first. This is the most underestimated step. As one practitioner analysis puts it bluntly, without Data Cloud, Agentforce has nothing useful to read — the single biggest scoping surprise for buyers. Agents need a shared source of truth across clouds and workflows; without one, they're limited to isolated tasks that never deliver enterprise value.

Reuse what you already have. Agentforce sits on the Salesforce Platform, so existing Flows, prompt templates, Apex code, and MuleSoft API connections can be converted into agent actions rather than rebuilt. Agent Builder lets admins define an agent's job, instructions, and guardrails in natural language. Start with out-of-the-box agents — Service Agent, SDR, Sales Coach, Campaign Optimizer — before building anything custom.

Install the three guardrails before go-live. Research on stalled Agentforce programs found that missing any two of these predicts failure inside the first compliance review:

• Identity-aware permissions — the agent runs with the same record-level access as its named human owner, never more.
• Audit trails that capture step-by-step reasoning, so every action is reviewable.
• A kill switch with human-in-the-loop approval on customer-facing actions — able to pause all agents within 60 seconds.

Salesforce provides some of this out of the box: the Einstein Trust Layer adds dynamic grounding, zero data retention, and an audit trail, while Agentforce Guardrails block off-topic conversations and mitigate hallucination. Governance is still your job — only 21% of customers strongly agree they have the governance they need.

Finally, treat each agent like an employee: give it a job description tied to measurable outcomes, review performance at 30, 90, and 180 days, and sunset what doesn't deliver. And if you're extending agents with third-party skills and tools, verify them first — SkillzBench is building a free directory that security-scans, benchmarks, and trust-scores agent skills so you know what you're plugging in before your Salesforce agent touches it.

Making It Pay Off: Measuring ROI and Avoiding Cost Traps

AI agents can transform a Salesforce org — or quietly drain your budget. In an IBM Institute for Business Value survey of 1,200+ Salesforce customers, only 33% of AI initiatives met their ROI targets, and 72% failed to scale across business units. The difference between those outcomes usually comes down to cost planning and measurement discipline.

Start with the free on-ramp, then scope real costs. Salesforce offers free entry through Foundations, including credits for the first 1,000 Agentforce for Service conversations. But production pricing is variable: practitioner estimates put the standard tier at $2 per conversation, with Data Cloud as a frequently underestimated prerequisite — a 50,000-conversation first year can run $200K–$400K all-in. Budget Data Cloud as part of the agent engagement, not a separate project.

Then measure agents the way you'd measure people. IBM recommends treating each agent like an employee:

• Write a job description tied to risk-adjusted ROI before deployment.
• Set performance baselines so improvement is measurable.
• Review results at 30, 90, and 180 days.
• Sunset agents that don't deliver — as IBM puts it, "give agents an outcome — or take them off the payroll."

The payoff can justify the rigor. Reddit's Agentforce deployment deflected 46% of support cases and cut resolution times by 84%, according to practitioner analysis. A Fortune 500 finance team reduced reporting cycles from 15 days to 35 minutes, dropping per-report cost from $2,200 to roughly $9. Top performers in the IBM survey report 60% greater efficiency and more than double the pipeline expansion of peers.

Guard against the cost traps, too. 62% of customers worry about unpredictable AI costs, and 64% say unclear total cost of ownership makes it hard to judge whether agents actually save money. Klarna's public pullback from agent deployments — after admitting cost was weighted too heavily — is a useful cautionary tale.

Finally, extend agents carefully. When adding third-party agent skills to any AI workflow, verify what you install: research suggests roughly 26% of published skills carry at least one vulnerability. SkillzBench is building a free directory that security-scans and benchmarks skills so you can see each one's measured lift before it touches your agents — worth watching as your Salesforce agent footprint grows.

Going Further: Extending Salesforce AI with Agent Skills

Agentforce can automate service, sales, and marketing inside Salesforce — but the workflows around your CRM don't stop at the org boundary. Data prep, report generation, and cross-system integrations often live in spreadsheets, scripts, and third-party tools that Salesforce's built-in agents never touch.

That's where agent skills come in. Skills are SKILL.md instruction packages that teach coding agents like Claude Code, Codex, Cursor, and Gemini CLI how to perform specific jobs — cleaning a CSV export before it loads into Salesforce, building a pipeline report, or wiring an API integration. They slot into the same "reuse what you have" philosophy Salesforce recommends: existing Flows, Apex, and APIs can become agent actions, and skills extend that pattern to the adjacent work an agent can't reach from inside the platform. Given that growing SMBs are twice as likely to have an integrated tech stack (66% vs. 32%), closing those gaps matters.

The catch is quality. Skills are everywhere, but nobody can tell which ones actually help and which are risky — published research found roughly 26% of skills carry at least one vulnerability. Installing an unvetted skill into an agent that touches customer data is a governance problem, not a convenience. IBM's survey of 1,200+ Salesforce customers found only 21% strongly agree they have the governance they need, so adding unverified instructions to your agent stack works against you.

Before you install anything, it's worth checking what's actually in a skill. A few things to look for:

• A pinned version tied to a specific commit, so you know exactly what code you're running
• A security scan for common vulnerability patterns
• A license check and evidence of ongoing maintenance
• Benchmark results showing whether the skill measurably improves agent performance

That's the checklist SkillzBench was built around. It's a free, independent directory of agent skills that continuously crawls and dedupes skills from GitHub, pins every version to a commit SHA, runs security scans and license checks, and benchmarks each skill in a sandbox — testing with and without the skill, using the same model, to publish a "Lift" score in percentage points. A trust score from 0 to 100 combines security, quality, license, maintenance, and community signals, and money never touches the scores. SkillzBench is also building a free MCP connector with per-agent loadouts, so your agent only sees the skills and tools you explicitly choose — the same allowlist principle Salesforce practitioners apply inside the platform.

The platform is currently in pre-launch with a live waitlist, so you can join now and be first in line when the directory and MCP connector go public. If your Salesforce AI program is one of the 72% that struggles to scale, the skills around it — data prep, reporting, integrations — are often where the leverage hides. Vetting them properly is the difference between an agent stack that compounds and one that stalls.