Start with the task, not the agent
A queue of repetitive work is a reason to investigate. It is not proof that an AI agent is the right answer. First identify what makes the task difficult: missing information, a broken handoff, a rule that is not automated, or language that needs interpretation.
Our recommendation is to choose the least complicated approach that can meet the requirement. Anthropic’s engineering guidance similarly distinguishes predefined workflows from model-directed agents and recommends adding complexity only when the task needs it.
Anthropic: workflows, agents and when to use each
| Task | Starting point | What to test |
|---|---|---|
| Assign a record from agreed fields | A deterministic rule or Salesforce Flow. | Missing fields, overlaps and the exception owner. |
| Draft a summary from a long note | An AI-assisted step with a fixed input and output. | Accuracy against the source and reviewer effort. |
| Find an answer in approved documents | Permission-aware retrieval with a sourced response. | Missing, conflicting and outdated sources. |
| Investigate a request using several tools | A bounded agent, if simpler steps are insufficient. | Allowed actions, stopping conditions and recovery. |
An example inside a lending workflow
Illustrative workflow: an operations analyst receives a broker’s submission with documents and an email explaining the request. Fixed rules check whether required files are present. An AI step drafts a summary of the supplied material with references for the analyst to inspect.
The analyst confirms the summary and missing items before a response is sent. The workflow does not approve credit, set terms or infer facts absent from the documents. If the files cannot be read or disagree, the request stays in a review queue.
See the application-intake workflow example
Where Agentforce may fit
Agentforce is a Salesforce option for building agents connected to the platform. Salesforce documents a lifecycle covering design, configuration, testing, deployment and monitoring. Assess the relevant agent type, data needs and action capabilities in your actual org before choosing the architecture.
Salesforce: designing and implementing agents
A custom AI service connected through approved APIs is another option when the workflow spans systems or needs a different interface. Compare the two against the same requirements: identity, access, testability, deployment, operating cost and who will support the result. Buying a platform does not remove the need for that design work.
Access and actions need explicit boundaries
Agree which records the system can read, which actions it may request and which actions need a person’s approval. Enforce those rules in the application and integration permissions. A sentence in a prompt is not an authorization mechanism.
Salesforce describes Agentforce security as a shared responsibility: the platform supplies foundations, while administrators configure permissions and agent guardrails. Validate the effective access for the actual agent and user context.
Salesforce: Agentforce security responsibilities
Retrieval helps provide context; it does not prove an answer
A retrieved document may be old, incomplete or unrelated to the current customer. An apparently well-formed summary may still contain a mistake. Show the relevant source and its context to the reviewer, and define what happens when there is insufficient evidence.
Do not send every field or document to a model just because it is available. Select the information required for the task and agree how inputs, outputs and logs are handled.
Make the pilot answer a business question
Choose one workflow, a representative test set and an owner who can judge the output. Measure completed tasks, corrections, escalations, review effort and operating cost. Compare against the current process, including the difficult cases.
Release only within the agreed scope and review conditions. Expand after the team has evidence that the workflow is useful and manageable. A fixed promise of production in a few weeks cannot account for unknown data, access and integration work.
Check the foundations with the readiness scorecard
Scope the next step
Bring one workflow, the systems it touches and an example of where the current process breaks down. We can discuss whether to improve the process, use conventional automation or scope an AI pilot. A delivery estimate follows the agreed requirements.