How to Plan the Transition to Agentic Operations

An agentic target architecture may look convincing while the live operation still depends on older products, integrations and people. How do you introduce agentic capabilities into a live workflow without losing control? Forrester sees a more coherent direction for Salesforce, but asks how customers will get there from the products they already use. Some acquired products will remain standalone, other transitions are unclear, and buyers need to test permissions and establish what agent actions will cost. [1]

That gap is the job of a transition architecture: to show how today's operation can change, who remains in control, and what must be proven before the next step.

CrmWare advises on Salesforce transformation. We would apply these transition tests to our own proposals too.

Start by selecting one live workflow

Choose one workflow within your current operation before drawing the future platform. A marketing journey, for example, involves customer data, a decision about the next communication, an action and people who handle exceptions. Record the systems and roles involved. Set an outcome baseline, such as journey completion, alongside unwanted contacts or manual corrections.

The governed-workflow brief shows what to record before introducing an agent. Add two questions: who might be affected by the agent's decisions, and what is the first action it may take? The answers set the impact and authority boundary for the first phase.

Ask the workflow owner to approve this boundary. Otherwise, a pilot may appear successful because it has not inherited the difficult parts of the live operation.

Decide what to keep, integrate, replace or defer

For each data source, automation, interface and product, decide: keep, integrate, replace or defer. Record its dependency and owner. Coexistence means the old and new paths have explicit divisions of work and authority; they need not process the same cases.

Salesforce's marketing products show what coexistence can look like. Selected data from Marketing Cloud Engagement can become available in Marketing Cloud Next while existing integrations continue to use it. [2] Salesforce also describes an agent whose journey decisions can be used in Journey Builder. [3] The transition question is which parts of the journey stay as they are, what decisions the agent takes on, and who owns the handoff. Check whether the prerequisites, editions and integrations fit your estate.

Draw the current, interim and next states for that journey. In each, identify the data, decision and action path, human role and control owner. In the interim state, name the authoritative system at each step, including where the agent's result enters the journey. The platform-role question helps decide which products stay; the map shows what cannot move yet.

Salesforce's architecture guidance distinguishes the requesting user's permissions from custom actions that may run in system mode. [4] Mark the effective identity at that handoff, since the interface does not reveal whose access governs the action.

Approve the interim state when every dependency and decision handoff has an owner. An unlabeled arrow between products is not a transition decision.

Test authority at every handoff

When a decision or action shifts from a person or existing automation to an agent, identify who grants that authority. Trace what the agent can read, decide and change across connected systems. At each handoff, test the identity and permissions the action actually uses.

Test an allowed and a denied role. Follow one consequential action to its effect on a person, then simulate failure and human override. Record who authorised the agent's scope, who can narrow or suspend it, and what evidence they would see. NIST's voluntary AI Risk Management Framework describes clear responsibilities, human oversight and ways to override or deactivate an AI system. [5]

The AIGP body of knowledge covers development, deployment, use and monitoring. Apply that lifecycle lens whenever an agent's authority changes. [6] The World Economic Forum and Capgemini's Agent Capability and Authorization Profile addresses authorisation at the agent level. [7] The transition map shown above instead shows when authority enters the workflow, who approves it, and what must be true before it expands.

Set the evidence and cost gate for each phase

Before a pilot expands, compare its outcome with the baseline. Review exceptions, overrides and permission tests, including denied or failed actions. Name who can continue, pause, narrow or reverse the phase. Approval to launch the first phase should also name who will monitor outcomes, respond to problems and decide whether the agent's authority needs to change. NIST calls for monitoring and response after deployment. [5]

Cost belongs in that decision. Salesforce documents metering during some Agentforce testing and sandbox activity, with actions or prompts mapped to usage and credits. Billable units depend on the feature and agreement; this guidance gives no price for the journey. [8] Forrester advises clarifying usage definitions and seeking price protection. Check both against the agreement. [1]

At the steering review, request the measured outcome, control exceptions, expected usage at the next volume and the decision-maker. If these are missing, hold the phase. A working pilot does not answer the go-live acceptance question for a wider rollout.

A transition architecture should make the next decision possible. For one live workflow, show what will change, who may act, who can intervene, and what evidence is needed to proceed. Those questions apply whether the systems are Salesforce, another platform or a mix of both.

References

[1] Forrester. From CRM To The Agentic Enterprise: Salesforce’s Biggest Dreamforce 2026 Reveals. 2026. https://www.forrester.com/blogs/from-crm-to-the-agentic-enterprise-salesforces-biggest-dreamforce-2026-reveals/

[2] Salesforce Help. Working with Data Extensions. 2026. https://help.salesforce.com/s/articleView?id=mktg.mke_data_ext.htm&language=en_US&type=5

[3] Salesforce Help. Create and Use the Journey Decisioning Agent. 2026. https://help.salesforce.com/s/articleView?id=mktg.mceplus_journey_decisioning.htm&language=en_US&type=5

[4] Salesforce Architects. Trust for the Agentic Enterprise. 2026. https://architect.salesforce.com/docs/architect/well-architected/guide/agentic-enterprise-trust.html

[5] National Institute of Standards and Technology. AI Risk Management Framework Core. 2023. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/

[6] International Association of Privacy Professionals. AIGP Study Guide 2026. 2026. https://pages.iapp.org/rs/138-EZM-042/images/AIGP%20FSG.pdf?version=4

[7] World Economic Forum and Capgemini. AI Agents in Action: A Playbook for Trusted Adoption, Authorization and Scaling. 2026. https://www.weforum.org/publications/ai-agents-in-action-a-playbook-for-trusted-adoption-authorization-and-scaling/

[8] Salesforce Help. Agentforce and Generative AI Usage and Billing. 2026. https://help.salesforce.com/s/articleView?id=ai.generative_ai_usage.htm&language=en_US&type=5