Salesforce AI Workflows: What They Are and How to Start.

A Salesforce AI workflow is a job handed to an AI agent instead of a fixed set of rules. You give the agent a goal. It works out the steps. It uses your Salesforce data to get there.

Here is how it works, what agents can do, and how to start without breaking anything.

TL;DR

• Old Salesforce automation follows fixed "if this, then that" rules. AI agents aim at a goal and pick their own path.
• Agents suit messy, repeated work: service questions, lead checks, campaign tweaks.
• Start small. Pick one process, measure it first, then build.
• Test the agent like a new hire. Give it narrow permissions and keep a human nearby.
• Track results against a baseline (your numbers from before the agent).

What a Salesforce AI workflow actually is

Think of a kitchen. Old automation is a recipe card. Follow step one, then step two, then step three. It works well when the ingredients never change.

An AI workflow is a cook. You say, "Make dinner for four." The cook checks the fridge and picks the steps.

In Salesforce, that cook is an agent. Salesforce's product for this is called Agentforce. Its architecture guidance says agents reason towards a goal. Traditional automation follows rigid rules that break when data is unexpected or unstructured.

That is the key difference. Rules need you to plan every path. An agent chooses a path based on the situation in front of it.

Old automation versus AI agents

Neither is better in every case. They do different jobs, and most good setups use both.

	Rule-based automation	AI agent
How it decides	Fixed if/then rules	Reasons towards a goal
Path	You define every step	Agent picks the steps
Messy input (emails, free text)	Often breaks	Built to cope
Predictability	Same result every time	Can vary, so needs testing
Best for	Clear, repeatable steps	Judgement-heavy, varied tasks

A simple test helps. If you can write the steps on one page and they never change, use rules. If every case looks a little different, an agent may fit.

How agents work inside Salesforce

The reasoning engine

Agentforce runs on what Salesforce calls the Atlas Reasoning Engine. Per Salesforce's launch announcement, it evaluates a request, retrieves relevant data, builds a plan and refines it. That lets agents handle tasks with many steps.

Think of a new staff member. They read the request. They look up the customer. They plan the reply. Then they act.

Teams of agents

One agent can do one job well. Many jobs need a team. Salesforce's architecture guide lists the benefits of using several agents: better performance, modular parts, fault tolerance (one failure does not sink everything) and shared governance.

The same guide says open standards help agents talk to each other. These are the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) protocol. A protocol is just a shared set of rules for talking, like a common language.

If you want the bigger picture of agents working as a team, see our guide to what AI orchestration is.

Back-office work

Salesforce also has Agentforce Operations, formerly called Regrello. Its process playbook describes it as a platform for orchestrating middle and back-office processes with AI-first patterns. That covers work that sits between teams, like approvals and hand-offs.

What jobs can Salesforce AI agents do?

Salesforce's implementation guide says agents can answer customer service questions, qualify sales leads and optimise marketing campaigns.

Here is what that can look like day to day:

• Service: An agent reads a customer question, checks the case history and drafts or sends a reply.
• Sales: An agent looks at new leads, scores them and flags the ones worth a call.
• Marketing: An agent reviews campaign results and suggests changes.
• Operations: An agent chases missing information before a hand-off.

Lead work is a good example. Agents are only as good as the leads they pass on. If you also run outreach outside Salesforce, AI lead generation for B2B shows why checking replies before a sales hand-off matters.

Want the same idea in another CRM? We wrote about HubSpot AI workflows too.

Get ready before you build

Most failed AI projects fail before the build. The agent is fine. The process behind it is a mess.

Salesforce's playbook advises you to do three things first:

1. Define the outcome you want.
2. Pick high-priority processes and record how they perform today (the baseline).
3. Set up a Center of Excellence (CoE). That is a small group that owns the rules and takes in new requests.

The playbook has eight steps in total. They run from strategic alignment and readiness, through process selection and quick value, to roadmap refinement, scaling, measurement, value reinforcement and operating-model evolution.

A simple readiness check

Ask yourself:

• Can I describe this process in plain words?
• Is the data in Salesforce clean enough to trust?
• Who decides when the agent is wrong?
• What number will tell me it worked?

If you cannot answer these, pause. Fix the process first. An agent on a broken process just breaks it faster.

Pick one small job

Do not start with "automate sales." Start with one narrow task. For example: sort inbound leads into hot, warm and cold. Small jobs give quick wins and teach you how agents behave.

Our post on AI workflow design walks through picking and shaping that first job.

Build, test and launch

Treat an agent like a new hire. You would not give a new hire full access on day one. You would train them, check their work and widen their duties over time.

The five phases

Salesforce describes an Agent Development Lifecycle with five phases:

1. Ideation and design
2. Development
3. Testing and validation
4. Deployment
5. Continuous monitoring and tuning

The last phase matters most. An agent is never "done." You keep watching and adjusting.

Testing is different for agents

Regular software gives the same output for the same input. Agents can vary. So the same guide says testing must go further. It names unit tests, end-to-end tests, adversarial testing (trying to trick the agent), human-in-the-loop review and scale simulation.

In kitchen terms: taste the dish, cook it ten times, try to make it fail, and have a chef check it.

Technical set-up notes

If your team deploys with the Salesforce command line, check the prerequisites. Salesforce's help article says you need to target API version 64.0 or higher. It also says features such as Einstein, Einstein Bots, Agents and Data Cloud must be switched on in the target environment.

Your admin or developer should confirm the current requirements before launch. They change over time.

Keep it safe

Agents act. They do not just report. So safety belongs in the design from day one.

Salesforce's admin guidance says agents have memory, chain tools on their own and work beyond a single user's session. Standard settings are not enough. It recommends:

• Give each agent a clear identity.
• Grant narrow, explicit permissions.
• Log decisions and actions.
• Plan targeted human review.

Salesforce also describes the Einstein Trust Layer, with guardrails such as zero data retention, toxicity detection and secure data retrieval, in its launch announcement.

One more point from the playbook: build governance and compliance in from the start. Adding them later costs far more.

If your business handles customer data, ask your legal or security lead to review the agent's permissions before launch. For a deeper look, read our AI agent governance playbook and our piece on AI policy development.

Measure whether it works

No number, no proof. Before you launch, write down today's figures. Then compare.

Salesforce's playbook points to real-time ROI dashboards that track agent accuracy, adoption and business impact against baselines. It also says teams that scale well keep asking one thing: does this process create more value than it costs?

Good numbers to track:

• Time to first reply
• Cases solved without a human
• Leads handled per person
• Hours your senior people spend on repeat decisions
• Errors caught in review

Pick three. Not twelve. Check them weekly for the first month.

Our view: the agent is only as good as the expertise inside it

We work with founders running $1M+ education and consulting businesses. The same problem shows up again and again. The founder is the bottleneck. Every tricky call lands on their desk.

Most AI projects miss this. They automate tasks. They do not capture how the founder decides.

A Salesforce agent is a strong tool, but it still needs instructions. If the instructions are generic, the output is generic. Your edge is your method: how you judge a lead, how you handle an upset client, what "good" looks like.

So we start there. We pull that judgement out of your head and write it down. Then we build a team of agents that applies it across content, operations, delivery and support. We build them with Claude Code, as one AI operating system. Salesforce can be one of the places they work. (See custom AI delivery systems with Claude Code.) The goal is a team that delivers to your standard without you in every loop.

We care about the boring parts too: clear ownership, no vendor lock-in, a system you own.

If you want to see how the pieces fit, read how to connect AI agents across business operations.

James Killick

Want help turning your method into agents?

We offer two ways in.

• Foundation Month, $3,500 one-off. A focused start. We map your processes and build a first prototype.
• 90-day program, $18,500 one-off. A done-with-you build of a network of AI agents based on your own expertise. We aim to have a working system live in 30 days.

Prefer to build on your own? Our platform costs $90 per month or $900 per year.

See the details on our AI consulting services page, or try the Platform Access plans.

These are built for founder-led consulting and education businesses earning $1M+ a year. If that is you, we should talk.

Your next step

Pick one repeated decision that eats your week. Write down how you make it. Then book a conversation with us and we will show you how to turn it into an agent.

Sources

• Determining Agentic and Traditional Workflow Automation, Salesforce Architects
• Agentforce general availability announcement, Salesforce
• Enterprise Agentic Architecture and Design Patterns, Salesforce Architects
• AI Process Automation Playbook, Salesforce
• Agentforce implementation guide, Salesforce
• The Agent Development Lifecycle, Salesforce Architects
• Agentforce SFDX Deployment Prerequisites, Salesforce Help
• From Flows to Decisions, Salesforce Admins
• AI Lead Generation for B2B