Agentforce Optimization: How to Improve Agent Performance After Launch

Launch is the beginning, not the end

The most common mistake after an Agentforce go-live is treating it as a completed project. The team celebrates, the project manager closes the Jira board, and the agent runs unsupervised until someone notices it's performing poorly six months later.

The best-performing Agentforce deployments we manage treat optimization as an ongoing program. Agents that started at 40% deflection reach 65–70% within 12 months through systematic tuning. Agents left unoptimized plateau or regress as your product changes, your customer base evolves, and your knowledge base drifts out of date.

Here's the optimization framework we use.

The four levers of Agentforce performance

Every performance improvement in Agentforce comes from one of four places:

1. Role and topic instructions — how clearly the agent understands its job and how to do it
2. Action quality — how accurately actions retrieve, process, and write data
3. Knowledge base quality — how accurate and complete the agent's reference information is
4. Data quality — how complete and accurate the underlying CRM records are

When optimizing, always diagnose which lever is causing the performance gap before making changes. Random prompt tweaks without a diagnosis are slow and often counterproductive.

Week 1–4: Transcript review and baseline measurement

The first month after launch is a data collection phase. Your goal is to build a clear picture of where the agent is failing and why.

Set up your monitoring dashboard

Create a Salesforce report (or CRM Analytics dashboard) that tracks these metrics daily:

• Deflection rate: % of conversations resolved without human escalation
• Escalation rate: % escalated, broken down by escalation reason
• Resolution time: Average turns to resolution for fully-deflected conversations
• CSAT: Post-conversation rating (configure Einstein Feedback to collect this automatically)
• Topic distribution: Which topics are being triggered and in what proportion
• Action error rate: How often actions fail or return no results

Weekly transcript review process

Set aside 90 minutes per week for transcript review. Sample 20 conversations: 5 fully resolved, 5 escalated, 5 with low CSAT, and 5 random. For each conversation, ask:

• Was the topic classification correct?
• Did the agent invoke the right actions in the right order?
• Was the response accurate based on the data returned?
• Was the response tone and length appropriate for the channel?
• If escalated, was the escalation decision correct?

Document every failure. After 4 weeks, you'll have a clear pattern of the top 5–10 failure modes. These become your optimization backlog.

Month 2: Instruction optimization

Instruction optimization is the highest-leverage improvement you can make. Small changes to role descriptions and topic instructions frequently produce 5–15 percentage point improvements in deflection rate.

Role description refinement

The most common role description failures:

• Scope too broad: The agent is trying to handle everything and doing nothing well. Add explicit scope boundaries: "You handle X, Y, and Z. For all other requests, transfer to a human agent."
• Missing limits: The agent is making commitments it shouldn't. Add explicit prohibitions: "You cannot commit to specific timelines, pricing, or outcomes. Refer these questions to the discovery call."
• Unclear escalation conditions: The agent isn't escalating when it should. Add precise escalation triggers with examples: "Escalate immediately if the customer mentions legal action, uses profanity three or more times, or explicitly asks for a manager."

Topic instruction refinement

Each topic's instructions should be reviewed independently. For each failing topic:

1. Read the topic description — would a new employee understand exactly what this topic covers?
2. Check for overlap with other topics — could the same request plausibly trigger a different topic?
3. Review the topic instructions — are they specific enough to guide the agent through edge cases?
4. Check the action library — does every action the topic needs exist and have the right description?

Month 2–3: Knowledge base optimization

Knowledge base quality is the most underinvested area in most Agentforce deployments. The agent's answers are only as good as the articles it can retrieve.

Knowledge base audit process

For every failure where the agent gave incorrect information, trace back to the knowledge article that should have contained the right answer. Then classify the failure:

• Article missing: The information didn't exist in the knowledge base. Create it.
• Article inaccurate: The article contained wrong information. Update it.
• Article ambiguous: The article contained correct information but the agent misinterpreted it. Rewrite for clarity — shorter sentences, explicit statements, no ambiguous pronouns.
• Article not retrieved: The right article exists but the agent didn't surface it. Check the article's data category tags and update them to match the terms the agent uses when searching.
• Conflicting articles: Two articles gave different answers to the same question. Consolidate or explicitly reference one from the other.

Article formatting for AI agents

Articles written for human readers often perform poorly as agent grounding. Optimize for agent retrieval:

• Start each article with a one-sentence answer to the question it addresses
• Use headers that match the exact phrases customers use ("How do I cancel?" not "Cancellation Policy")
• Keep articles to a single topic — split multi-topic articles
• Replace tables with bullet lists where possible — LLMs parse lists more reliably than tables
• Date-stamp articles and set a review cadence — stale articles are a constant source of agent errors

Month 3+: Action and data optimization

Action performance analysis

Pull your action invocation logs and calculate a success rate for each action: how often does the action return a useful result vs. an error, empty result, or partial data? For any action below 90% success rate:

• Check the action's SOQL query or API call for errors in edge cases
• Review the action description — is the agent invoking it correctly?
• Check the permission set — does the agent have access to all records the action needs?

Data quality improvements

Identify the 5–10 fields most frequently queried by your agent's actions. For each field, calculate the completion rate across the records the agent typically touches. Any field below 80% completion is a reliability risk.

Fix data quality issues at the source: add required field validation rules, update your data entry workflows, or implement an enrichment action that populates missing fields from external data sources when the agent encounters a record with gaps.

The quarterly optimization review

Once you're past the initial stabilization period, shift to a quarterly optimization cycle:

• Performance review: Compare current metrics to the prior quarter and to your initial baseline. Calculate the trend on each core metric.
• Use case expansion: Identify the next highest-volume, highest-impact use case to add to the agent's scope. Every 3–6 months, expand what the agent can handle.
• Competitive benchmarking: Review Salesforce release notes for new Agentforce capabilities. Each Salesforce release typically includes new actions, improved grounding, or new channel support that you can leverage.
• Stakeholder review: Present results to your executive sponsor. The quarterly review is where you build the business case for the next phase of investment.

Performance benchmarks by industry

Based on our deployment data, here are the deflection rate and CSAT benchmarks we see at 12 months post-launch:

Industry	Deflection Rate	CSAT Improvement	Handle Time Reduction
B2B SaaS	60–72%	+0.4–0.6 pts	35–45%
Financial Services	45–60%	+0.3–0.5 pts	25–35%
Healthcare / Life Sciences	40–55%	+0.2–0.4 pts	20–30%
Manufacturing / Distribution	50–65%	+0.3–0.5 pts	30–40%
Professional Services	35–50%	+0.2–0.4 pts	20–30%

If you're below the low end of your industry benchmark at 6 months post-launch, a structured optimization engagement is almost always the right next step. The gap between a well-optimized and a neglected deployment widens every month.