How to Use Salesforce AI Tools: A Practical Walkthrough

The Salesforce AI Promise vs. the Implementation Reality

Salesforce wants you to believe the age of autonomous AI is already here — and to be fair, the technology genuinely does what they describe. Agentforce agents work by searching for relevant data, analyzing it to formulate a plan, and then executing that plan without human intervention, according to Salesforce's own research. Marc Benioff frames this shift as comparable to the cloud and mobile revolutions. The pitch is compelling.

The vendor statistics back it up, at least on the surface. In a survey of 3,350 SMB leaders, 91% of those using AI say it boosts revenue, and 92% of service teams with AI report cost reductions. Wiley, an early Agentforce customer, saw a 40%+ increase in case resolution. These numbers explain why adoption momentum feels unstoppable.

But independent research tells a more sobering story. IBM's Institute for Business Value surveyed more than 1,200 Salesforce customers and found that only 33% of AI initiatives meet their ROI targets. Worse, 72% have failed to scale AI across business units, and 20% of initiatives have stalled, failed, or been abandoned outright. The gap between vendor promises and customer reality is wide.

The friction points are consistent and predictable:

• Cost opacity — 62% of customers worry about unpredictable AI costs, and 64% cite unclear total cost of ownership.
• Governance gaps — only 21% strongly agree they have the governance their AI needs, and 56% haven't issued comprehensive AI guidance to employees.
• Trust deficits — only 10% of people fully trust AI to help them make informed decisions, per Salesforce's Trends in AI research.

The deployment path matters enormously here. Consultancy analysis suggests pre-built Agentforce deployments take 4–6 months versus 18–24 months for custom builds, with failure rates of 0–10% versus 73% for custom AI projects. In other words, how you implement matters more than whether you implement.

That's exactly why a disciplined walkthrough beats a feature tour. The same principle applies beyond Salesforce: when you assemble any agent workflow, the individual skills and tools you bolt on carry real risk — published research found roughly 26% of agent skills contain at least one vulnerability. Verification before deployment is the discipline that separates the 33% hitting ROI from everyone else. SkillzBench is being built to make that verification routine: a free directory, security scanning, and sandbox benchmarking for the agent skills that increasingly power these workflows.

So as we walk through Salesforce's AI tools practically, keep one question front of mind: not "what can this do?" but "what will we measure, and how will we know it worked?"

Start with Data Foundations and Pick Pre-Built Over Custom

Before you touch a single AI tool in Salesforce, decide how you'll deploy it — because that choice, more than the technology itself, determines whether you join the winners or the majority who stall. Independent research from IBM's Institute for Business Value found that only 33% of AI initiatives meet ROI targets, and 72% of Salesforce customers have failed to scale AI across business units.

The good news: Salesforce's own research identifies three success patterns that separate the adopters who succeed from those who don't. Treat them as your deployment checklist before evaluating any agent or prompt.

• Data foundation first. Data quality directly determines AI results. An agent that searches, analyzes, and acts on messy CRM data will simply make confident mistakes faster.
• Integrated stack. AI delivers value when it can act on data across the whole business, not just one cloud or silo. Salesforce-native agents require no middleware and scale from 100 to 100,000 users.
• Security and trust. Security ranks as the top technology challenge, and 77% of business leaders cite trusted data and ethics issues that could halt AI deployments entirely.

Once those foundations are in place, the deployment evidence points clearly in one direction: start with pre-built, not custom. According to deployment comparisons, pre-built Agentforce implementations take 4–6 months with failure rates of 0–10%, while custom builds stretch to 18–24 months, fail 73% of the time, and cost $500K–$2M in development alone. As the consultancy authors put it, the best AI solution isn't the most customized one — it's the one that gets adopted and starts delivering value fast.

Practically, that means starting with what's already built. Salesforce ships dozens of customizable, turnkey prompts and actions across its clouds, and partner-built agent packs — like A5's AgentIQ suite with OrderIQ, SalesIQ, and AssistIQ — distribute through AppExchange and AgentExchange, making deployment as simple as installing an app. The momentum backs this up: over 10,000 AI agents were reportedly created in three days at Dreamforce 2024 using Agentforce.

Verification still matters, even with pre-built options. Published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building a free directory that security-scans, license-checks, and sandbox-benchmarks agent skills before you connect them through its free MCP tool. The same verify-before-deploy discipline applies whether you're installing a Salesforce agent pack or wiring skills into your own agents.

Finally, give every agent a job description. IBM's recommendation is blunt: give agents an outcome or take them off the payroll. Set baselines, measure at 30, 90, and 180 days, and sunset underperformers — a discipline that matters when only a third of initiatives hit their ROI targets.

The Walkthrough: Give Every Agent a Job Description and Measure It

Deploying an AI agent without a defined outcome is like hiring an employee with no job description — and it's exactly why IBM's research bluntly advises: "Give agents an outcome—or take them off the payroll." That same study found only 33% of AI initiatives meet ROI targets, so treating agents like accountable members of the workforce isn't optional. It's the difference between a pilot that impresses a demo audience and one that survives budget review.

Start by mapping each agent to the way Salesforce agents actually work: they search for relevant data, analyze it to formulate a plan, and then execute that plan without human intervention, according to Salesforce's own description of the agent workflow. Your job description should therefore cover three things: the data the agent can access, the plan it's authorized to form, and the actions it's allowed to take.

Before flipping the switch, write down the numbers you'll be judged against. IBM recommends setting a baseline before deployment and measuring at 30, 90, and 180 days, tying each agent to a risk-adjusted ROI. If an agent can't beat its baseline by the 180-day mark, sunset it. This discipline directly counters the 62% of Salesforce customers who told IBM they're concerned about unpredictable AI costs.

A well-written agent job description answers four questions:

• What outcome does this agent own, expressed as a single measurable metric?
• What data sources and tools is it permitted to touch, and what's off-limits?
• When must it escalate to a human instead of acting autonomously?
• What baseline does it need to beat, and by when?

The proof this works is in the results. Salesforce reports that Wiley achieved a 40%+ increase in case resolution with Agentforce, with 70% of cases resolved autonomously — numbers that only mean something because someone defined the baseline first. That's the pattern to copy: narrow scope, clear metric, hard checkpoint.

The same accountability principle applies to the agent skills you attach to any AI workflow. Published research found roughly 26% of agent skills carry at least one vulnerability, so every skill an agent uses deserves the same scrutiny as the agent itself. SkillzBench is building a free directory and MCP connector that verifies, security-scans, and sandbox-benchmarks agent skills before they reach your stack — the same "measure before you trust" mindset IBM prescribes for the agents themselves.

Governance, Costs, and Trust Controls That Keep Deployments Alive

Most Salesforce AI deployments don't fail on technology — they fail on trust, cost, and control. The numbers back this up: IBM's survey of 1,200+ Salesforce customers found that only 33% of AI initiatives meet ROI targets, and 72% of customers have failed to scale AI across business units.

Cost anxiety is a major driver. That same research found 62% of customers worry about unpredictable AI costs, and 64% cite unclear total cost of ownership. When an agent's consumption is opaque, budget owners hesitate to expand pilots — and hesitation kills deployments faster than any technical bug.

Governance gaps make it worse. Only 21% of surveyed customers strongly agree they have the AI governance they need, and 56% have not issued comprehensive AI guidance to employees. The result is predictable: according to Salesforce's own research on AI trends in CRM, 55% of employees use unapproved AI tools — and 40% use tools that are explicitly banned. Shadow AI flourishes exactly where official programs feel unsafe or unclear.

Clara Shih, CEO of Salesforce AI, has articulated what enterprise deployments actually require: "Success requires much more than an LLM." Her checklist gives you concrete evaluation criteria you can apply to any Salesforce AI rollout — pre-built agent pack or custom build:

• User access control — every agent inherits strict permissions so it only sees the data its role justifies
• Audit trails — a complete, reviewable record of what each agent did and why
• Data masking — sensitive fields hidden from prompts and outputs by default
• Citations — answers grounded in verifiable sources, not model guesses
• Trusted data foundations — agents act on governed, quality-checked records

IBM's recommendation is blunt: "Governance must evolve from a compliance function to become the operating system for enterprise AI." Practically, that means treating each agent like a hire — give it an outcome, measure it at 30, 90, and 180 days, and take underperformers off the payroll.

These controls matter beyond Salesforce, too. The same verification discipline applies to the agent skills that extend AI workflows — deployment research shows adoption and trust determine success more than customization does. That's the problem SkillzBench is building to solve: a free directory and MCP connector that security-scans, license-checks, and sandbox-benchmarks agent skills before they touch your stack — because published research found roughly 26% of skills carry at least one vulnerability.

Run every AI deployment — Salesforce or otherwise — through Shih's checklist before it goes live. Trust controls aren't overhead; they're what keeps deployments alive.

Extending Your Salesforce AI Stack with Verified Agent Skills

Your Agentforce agents are live, the data foundation is solid, and the first ROI numbers are trickling in. But an agent that resolves cases inside Salesforce is only the beginning — the rest of your CRM workflow still runs on manual busywork, and salespeople spend 71% of their time on non-selling tasks.

That's where agent skills come in. Skills are instruction packages (SKILL.md files) that teach coding and chat agents like Claude Code, Codex, Cursor, and Gemini CLI how to handle specific tasks — lead enrichment, report drafting, pipeline hygiene, anything that surrounds your Salesforce deployment. Think of them as the connective tissue between your agents and the work your team still does by hand.

The catch: skills are risky. Published research found roughly 26% of agent skills carry at least one vulnerability, and that number should worry anyone who just spent months hardening their CRM data foundation. It echoes what IBM's survey of 1,200+ Salesforce customers found: only 21% strongly agree they have the AI governance they need, and 72% have failed to scale AI across business units. Verification isn't optional — it's the difference between an agent that helps and one that quietly introduces risk.

This is the problem SkillzBench is being built to solve. Currently in pre-launch, it's a free, independent directory of agent skills designed around a simple idea: before you equip an agent, know exactly what you're installing.

• Discover: continuous crawling and deduplication of skills from GitHub, so you're not guessing what exists
• Verify: every version pinned to a commit SHA, with license checks, security scanning, and quality validation
• Benchmark: each skill tested in a sandbox with the same model, with and without the skill, publishing the Lift in percentage points
• Equip: a free MCP connector plus per-agent loadouts — your agent only sees the skills and tools you've toggled on

The design choices matter for governance-minded teams. Lift is always shown separately from ratings, so a skill can't buy its way to a good score — money never touches scores. A 0–100 trust score combines security, quality, license, maintenance, and community signals. And when upstream skills change, SkillzBench re-scans, re-benchmarks, and notifies you, so a skill that was safe at installation doesn't silently become a liability.

Everything on the platform — the directory, the MCP connector, benchmarks, and alerts — is free. If you're extending a Salesforce agent stack and want proof a skill actually helps before it touches your CRM, join the SkillzBench waitlist and see what verified, benchmarked skills can do for the rest of your workflow.