Agentforce makes it genuinely easy to build an AI agent on top of Salesforce. It does not make it easy to build one you would actually put in front of a customer or trust to update a record unsupervised. The distance between an impressive demo and a production agent is enormous, and it is almost entirely about grounding, tool use, evaluation, and governance, not about the prompt. That distance is exactly where our practice lives.
We design and ship production Agentforce agents on Sales Cloud, Service Cloud, Marketing Cloud, and custom apps. That means agent design and the tool use that lets an agent actually take action, evaluation so you know whether it is right before it goes live, human-in-the-loop workflows for the moments that need judgment, and governance so it stays inside the rules. The integration work with Data Cloud, MuleSoft, and external systems is what lets an agent move the work forward instead of just describing it.
The common failures: an agent grounded in messy data that answers confidently and wrongly, no evaluation so nobody can tell good output from bad, automation handed authority it should not have, and a pilot that demos beautifully and then quietly never ships because the production hardening was never scoped. We build the grounding, the guardrails, and the evaluation from the start, because those are the parts that decide whether an agent survives contact with real users.
Our Agentforce work is grounded in real production deployments, not slide decks, and we are both a Salesforce Consulting Partner since 2017 and a Select Services Partner in Anthropic’s Claude Partner Network. That combination matters: shipping an agent that works requires deep knowledge of the platform, the data underneath it, and the model on top. We build agents that do the work, with the controls that let you trust them to.
The engagements we run most often on Salesforce Agentforce, from first implementation through optimization.
Workflow analysis, agent opportunity mapping, and a sequenced roadmap with measurable outcomes, not a thousand small experiments.
AI agents for case classification, deflection, summarization, knowledge retrieval, and recommended response generation.
AI agents for prospecting research, meeting prep, follow-up drafting, opportunity hygiene, and CPQ-assisted quoting.
Apex actions, prompt templates, evaluation frameworks, and tool use integrations beyond the out-of-box patterns.
Audit logging, approval workflows, accuracy monitoring, drift detection, and human-in-the-loop escalation patterns.
Grounding agents in real-time customer data via Data Cloud, with relevance and freshness controls.
Engagements are measured by movement on the numbers that matter. These are the directions of travel we commit to.
Predictable phases. Clear deliverables. No surprises.
One to two working sessions to map your current state, business goals, and gaps. We come out with a written scope and recommendation.
Documented architecture, realistic timeline, and transparent commercial proposal. No surprises and no hidden scope.
Configuration, development, integrations, data migration, and QA, with weekly demos and on-the-fly adjustments.
Training, change management, hypercare, and ongoing optimization. We do not disappear at go-live.
Consultant-level analysis from the consultants delivering the work.
What each of Salesforce's seven named Agentforce agents does, which are available now, the long-horizon runtime and multi-agent orchestration behind them, and how to decide which one to pilot first.
Read the full post →Subagents, called topics until the April 2026 rename, group the jobs an agent may do, and natural-language instructions guide behavior inside each one. The Agent Router picks which subagent handles a request. Tight scoping keeps an agent away from questions it should decline, and a business owner can review the rules in plain English.
Actions are the concrete operations an agent can perform, built from flows, Apex, prompt templates or external API calls. This is where an agent stops talking and changes something, such as rescheduling an appointment or creating an order. Each action therefore needs its own input validation, permissions and testing before real customers can reach it.
Agentforce Builder in Agentforce Studio is now the recommended place to build agents, with a visual canvas and a script view. Agent Script lets you hard-code steps that must always happen, such as verifying identity first, while leaving judgment calls to the model. The Atlas Reasoning Engine then executes that mix of fixed logic and reasoning.
Reusable, versioned prompts that merge Salesforce record data into generation tasks such as summaries, emails or recommendations. Templates let administrators control wording and grounding centrally, so a change to tone or policy is made once rather than scattered across many agents. They also support testing against sample records before activation.
Testing Center, now part of Agentforce Studio, takes batches of sample utterances and checks the chosen subagent, action and response against what you expected. Conversation-level tests also simulate multi-turn exchanges, where drift tends to appear several turns in. A regression suite here catches problems before users do and gives release managers evidence to approve a change.
Agentforce Observability, which Salesforce first introduced as Command Center, monitors live agents: conversation volumes, escalations, errors and session traces in one view. Operations teams use it to find where an agent hands off too often, where an action keeps failing or where customers abandon. Tuning then becomes a routine review rather than a reaction to complaints.
Define the agent’s role the way you would for a new hire: which requests it handles, which it refuses, what it may change and when it must escalate. Include examples of good and bad responses. This document becomes the source for subagents, instructions and test cases, and it gives legal and compliance something concrete to approve.
Create each action as a flow or Apex method with explicit inputs, validation and error messages the agent can interpret. Put mandatory sequences, such as identity checks before account changes, in Agent Script rather than instructions. Run the agent under a dedicated user holding only the permissions its actions need, and test each action independently first.
Collect real phrasing from past cases, chats and emails, including off-topic requests, ambiguous wording and attempts to push the agent outside its scope. Record the expected subagent, action and outcome for each. Run the suite in Testing Center after every change so regressions surface before anyone outside the project team sees them.
Put the agent in front of employees first, such as service representatives who can see its suggestions or answer alongside it. Their corrections reveal gaps in knowledge, confusing instructions and missing actions without any customer exposure. Treat this phase as data collection, and feed what you learn back into the test suite.
Expose the agent to a limited channel, region or customer segment, with escalation to people clearly visible. Review transcripts and handoff reasons regularly, and widen access only when escalations reflect genuine complexity rather than agent mistakes. Keep a documented way to disable the agent quickly if an action or answer misbehaves.
Agentforce can be bought several ways, and they do not all mix. Consumption runs on Flex Credits, charged per action, or on per-conversation pricing, and Salesforce states the two cannot run in the same org. Per-user options include the Agentforce User License for employees, which draws on Flex Credits, plus add-ons for specific clouds and industries and Agentforce 1 Editions. Salesforce Foundations gives eligible orgs a starter credit allotment. Flex Credits are shared with Data 360, and Digital Wallet tracks consumption by agent and action. Ask your account executive which model fits customer-facing versus employee-facing agents, and confirm current terms with Salesforce. Size usage from your real case and chat volumes.
Teams often pick the most visible, complex process to prove the technology, then stall on edge cases and approvals. A narrower first agent that handles a frequent, verifiable request builds testing habits, governance and internal confidence. Harder work can follow later, without betting the whole program on a single launch.
Running an agent under an administrator or integration user with wide access means any reasoning error can touch records it never needed. Give the agent a dedicated profile limited to the objects and fields its actions use. Narrow permissions contain mistakes and simplify security review, especially once agents face customers.
The reasoning engine chooses actions largely from their names and descriptions. Labels like Update Record or Process Request leave it guessing, which produces inconsistent choices that look random during testing. Describe what each action does, when to use it and what inputs it needs, as if briefing a colleague on their first day.
Counting conversations handled says little about whether customers got what they needed. Track resolution without follow-up contact, escalation reasons, action error rates and reviewer scores on sampled transcripts. Those measures show whether the agent is actually reducing work or just moving it to another queue where people clean up after it.
A Salesforce Health Check reviews your org and gives you a ranked list of what to fix first, with effort estimates. It is the right starting point before an optimization project, a new cloud or an AI initiative.
Choosing a use case with clear value, preparing the data and permissions the agent will use, designing subagents and actions, testing against real cases, setting up evaluation and governance, and rolling out with monitoring.
Most orgs need some preparation: duplicates, stale records, loose permissions and conflicting automation all get amplified by an agent. A Salesforce health check is a good first step.
A biotech company’s AI agent produced send-ready draft responses for 30% of question-type cases, and an infrastructure contractor used Document AI and Agentforce to read forms its existing tools could not.
Yes. In March 2026 we launched an Agentforce Center of Excellence that brings our production Agentforce and Claude work into one delivery framework with evaluation and governance.
Agentforce runs on models that Salesforce selects and manages, and the lineup changes over time. Under Salesforce’s expanded agreement with Anthropic, Claude models run inside the Salesforce trust boundary and are positioned as a preferred option for regulated industries. Some configurations also let you bring your own model for specific prompt templates. We confirm the current options for your org and edition during design, because model choice affects behavior, testing and data reviews.
Traditional bots follow scripted decision trees and break when a customer phrases something unexpectedly. Agentforce interprets the request, picks from the subagents and actions you define and can finish multi-step work, such as checking eligibility and then updating a booking. Agent Script lets you lock down the steps that must never vary. That flexibility is why testing and permissions matter: the agent chooses its path within the boundaries you set.
Yes. Escalation is configured as part of the agent’s design, routing to a queue or a specific team through Omni-Channel with the conversation history and a summary attached. Decide when to escalate, such as after repeated failed attempts, detected frustration or sensitive topics. Also decide what the receiving person sees, so customers never repeat themselves. Test handoffs as carefully as answers, including outside staffed hours.
Yes. Agents can be deployed across messaging channels, Experience Cloud sites, Slack and inside the Salesforce interface for employees. Agentforce Voice adds spoken conversations, with connections to several contact center platforms; confirm which telephony setups are supported for your org. Each channel has different expectations for response length, formatting and escalation, so we tune instructions per channel and test each with its own sample conversations before launch.
Treat each agent as a product with an owner. Review sampled transcripts on a regular schedule, rerun the regression suite whenever instructions, actions, knowledge or models change, and track escalation reasons for patterns. Platform releases can also shift behavior, so put agent regression runs on the same calendar as your other Salesforce change management. Retire instructions that no longer apply instead of letting them pile up.
The CRM where sales pipeline lives, when it is built right.
The operating system for modern customer service, case, knowledge, voice, and AI agent assist together.
Scheduling, dispatch, mobile work orders and parts for teams that work on site.
Cross-channel marketing, email, mobile, journey, and data-cloud-powered personalization at scale.
Industry CRM for wealth management, banking, insurance, and lending, built for compliance from the start.
Configure-price-quote, billing, subscription management, and quote-to-cash on Salesforce.
Tech Talk
A monthly brief for the people who own Salesforce, AI and revenue technology
What changed in Salesforce and AI this month, and what to do about it.
One email a month. Written by the consultants who deliver the work, not by a marketing team, for the leaders who make the technology decisions.
Consultant analysis, not vendor recaps. One click to leave.
