Agentforce is Salesforce’s platform for AI agents that reason over your CRM data and act through Flows, Apex and prompt templates. Implementations succeed when the org is ready first: clean, deduplicated records, documented Flows, field-level security the agent will inherit, a grounded knowledge base and one narrow use case with tested utterances.
Agentforce is the agent layer on top of the Salesforce platform. The agent reasons; your Flows, Apex, data and permissions do the work. That’s why readiness is an org problem, not an AI problem.
Where agents are defined: topics (jobs), instructions, and the actions each topic can call. Agent Script adds code-first definitions you can version and deploy through the Metadata API.
BuildClassifies the request into a topic, plans which actions to call, checks results and re-plans. Overlapping topic descriptions are the most common reason it picks the wrong path.
ReasonAutolaunched Flows, Apex, prompt templates and external APIs (often via MuleSoft). Actions run with the invoking user’s permissions, so sharing and field-level security decide what an agent can see and change.
Where it breaksThe grounding layer for anything beyond simple CRM lookups: unified profiles, unstructured content and real-time context. It’s metered separately from Agentforce usage.
GroundPermission-aware retrieval, PII masking, toxicity detection and zero data retention by the model provider. It enforces your security model; it doesn’t fix one that’s too permissive.
ProtectTest utterances per topic in Builder, reasoning traces that show which topic and action were chosen, and observability dashboards for token spend and outcomes after launch.
Prove| Pricing model | List price (published) | Best fit | Watch out for |
|---|---|---|---|
| Flex Credits | $500 per 100,000 credits; a standard action uses 20 credits ($0.10) | Mixed internal and external use where you can count actions | Multi-step agents call several actions per request; unused credits don’t carry into the next term |
| Per conversation | $2 per customer conversation | High-volume service chat with simple journeys | Can’t be combined with Flex Credits in the same org |
| Per-user add-ons | $125 per user/month (Sales, Service, Field Service); $150 for Industries | Employee-facing agents with predictable seat counts | Only licensed users get unmetered use; customer-facing traffic is priced separately |
| Agentforce 1 Editions | From $550 per user/month, with an org-level credit pool | Orgs standardizing on Salesforce for sales, service and data | Credit pool sizes differ across Salesforce pages; confirm in writing |
List prices as published by Salesforce and summarized by Supered (“Agentforce Pricing 2026,” checked October 1, 2026). Data 360 consumption is metered separately and isn’t included above. Negotiated pricing varies; confirm with your Salesforce account executive. Platform naming per Salesforce’s 2025–2026 releases: Data Cloud became Data 360, and Agentforce Builder replaced the earlier Agent Builder.
Every agent type inherits the quality of the org underneath it. The right-hand column is the work most implementations skip.
| Agent type | Does well | Doesn’t do | Data & org prerequisites |
|---|---|---|---|
| Service agent | Answers customer questions on web chat, messaging, email or voice; looks up orders and cases; creates and updates cases; escalates by rule. | Can’t answer from articles you haven’t published or that contradict each other; can’t act on objects its user can’t access. | Current, deduplicated Knowledge; case and entitlement data that’s reliable; escalation queues and routing in place; contact matching that works. |
| SDR / sales agent | Engages inbound leads, answers product questions, books meetings, nurtures on rules you set. | Doesn’t know your ICP or territories unless they’re fields; won’t fix lead routing that’s broken today. | Lead-to-account matching, clean lead status and source, ICP fields as picklists, working assignment rules and calendars. |
| Employee agent | Summarizes accounts and opportunities, drafts emails, updates records, answers “how do I” from internal docs in Slack or the app. | Can’t summarize activity nobody logged; can’t fix stages that don’t match how you sell. | Activity capture, stage definitions with exit criteria, required fields per stage, permission sets that mirror job roles. |
| Custom agent | Org-specific, multi-step jobs (quote checks, renewal prep, data stewardship) built from your Flows and Apex. | Not a replacement for missing automation; if the Flow doesn’t exist or fails silently, the agent fails too. | Documented, tested Flows and Apex with fault paths; a written input/output spec; a dedicated agent user and permission set. |
Thresholds are JayOh implementation standards drawn from our audit frameworks, not Salesforce requirements. The 20–50 utterance range matches common practice for topic-level testing in Agentforce Builder.
Twelve questions, two per domain. Answer for the objects your first agent would touch, not the whole org.
Unanswered questions count as zero.
Fix firstBands (JayOh guidance): under 50% not ready, fix the foundation first; 50–79% pilot-ready for one internal or tightly scoped use case; 80%+ ready to scale with test gates.
Score the six domains for the objects in scope, pick one use case with one owner and one KPI, and estimate actions per request so the pricing model fits.
Why: the use case decides which data, Flows and permissions must be ready. Auditing the whole org first wastes weeks.
Output Scored readiness map + one-page use-case specDedupe and backfill agent-facing fields, convert free text to picklists, retire legacy automation, add fault paths, build the agent user and permission set, and clean the Knowledge articles in scope.
Why: every one of these turns into a wrong answer or a failed action once an agent runs on top of it.
Gate Duplicates < 3%, inputs > 85% filled, Flows documentedWrite distinct topic descriptions, declarative instructions (what to check, when to escalate), and wire actions to tested Flows and Apex. Set turn limits and escalation rules.
Why: overlapping topics are the most common routing failure, and actions are where agents touch real records.
Output Agent in a sandbox with every action unit-testedRun 20–50 real utterances per topic, read the reasoning trace for each miss, and fix instructions, topics, data or actions. Re-run the full set after every change.
Why: a test set turns “it seems fine” into a number leadership can approve.
Gate ≥ 80% correct before any customer or rep sees itLaunch to one channel, segment or team, with human review where the agent writes data or talks to customers. Review outcomes, escalations and cost weekly.
Why: production traffic finds edge cases no test set does, and a narrow pilot keeps them cheap.
Output Weekly outcome report: accuracy, escalations, cost per outcomeExpand volume or autonomy only after two consecutive weeks above threshold. Add the next use case through the same readiness gate, and keep change control on Flows, permissions and Knowledge.
Why: agents drift when the org underneath changes. Operating is a standing job, not a launch task.
Output Agent register, cost dashboard, quarterly re-scoreA RevOps consultant plus Salesforce developers who audit, fix and build in the same sprints. Read-only first, sandbox before production.
We engineer growth systems. JayOh builds the RevOps architecture, GTM execution and data infrastructure that turn messy revenue operations into predictable pipeline for B2B SaaS and PE-backed teams.
