From Copilot to Ambient Agent: A Practical Agentforce Maturity Model

I spent three days at Dreamforce 2026 moving between agent sessions and conversations with customers. Organizations are moving beyond asking whether AI can work and are now asking how to make agents useful, trusted, governed, and scalable in real business workflows.

The question that particularly stood out was the shift from conversational AI towards quicker availability of more proactive and eventually ambient agents. But the underlying message was that greater autonomy has to be earned through reliable data, appropriate guardrails, human oversight, and measurable outcomes.

Industry data reinforces why this matters. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. In my experience, these challenges often stem from decisions made too early in the implementation process, particularly around how much autonomy each agent should have.

Here is the position I will defend in this piece. Autonomy is earned one workflow at a time, and most workflows should never become fully ambient. Below is the four-indicator model I use to decide where each workflow belongs, and how to tell when it has earned the next step.

Why do most agent programs stall after the first copilot?

They stall because the agent pattern is chosen before the use case is fully understood. I see three versions of this mistake repeatedly.

1. Autonomy treated as a setting. Teams assume the next step after a copilot is simply to turn on more autonomy. But autonomy is not a toggle. It rests on reliable data, appropriate permissions, and a proven track record. None of these come automatically with an AI license.
2. Use cases chosen from the demo, not the data. The most impressive keynote demo is rarely the workflow your data can support today. Gartner warns that without clear business context, agents are far more likely to hallucinate. This is often what undermines user trust once agents move beyond controlled demonstrations and into everyday workflows.
3. Governance written after the second agent. One agent can be monitored by the person who built it. But when multiple agents start interacting with the same records and business processes, teams need clear rules around ownership, permissions, accountability, and oversight. Too often, these rules are established only after something goes wrong.

These patterns broadly align with what I hear in customer conversations. A common scenario is a customer identifying a compelling AI use case, only to discover, when working backward from the desired outcome, that data quality, permissions, or process ownership still need attention before the agent can be given greater autonomy.

The fix is to classify each use case before deciding how the agent should behave. That is what the four indicators are designed to help teams determine.

What are the four indicators for classifying an agent use case?

Every agent use case can be described by four indicators. Score a workflow on all four and the right agent pattern usually becomes obvious.

Indicator	The question it answers	Low end	High end
Data autonomy	How independently does the agent access, interpret, and act on data?	Reads data a user points it to	Pulls from multiple systems and acts on what it finds
Trigger point	What starts the agent?	A user request	A system event, a schedule, or continuous monitoring
User engagement	How much interaction, clarification, approval, or oversight does it need?	A human reviews every step	A human reviews exceptions only
Value driver	What business outcome does it deliver?	Fewer clicks, better responses	Proactive resolution, real-time assistance, end-to-end task automation

The indicators are useful because they separate two things teams often blur together: what the agent can do and what it should be allowed to do today.

What is the difference between conversational, proactive, and ambient agents?

The three patterns sit on a progression. Each step raises data autonomy, moves the trigger further from the user, and reduces how often a person has to engage.

	Conversational agent	Proactive agent	Ambient agent
Data autonomy	Retrieves and summarizes on request	Monitors defined data and recommends action	Interprets signals across systems and acts within set limits
Trigger point	User asks a question	System event or schedule	Continuous monitoring
User engagement	User drives every turn	User approves or rejects a recommendation	User handles escalations and exceptions
Value driver	Fewer clicks, faster answers	Issues caught before they escalate	Work completed without a hand-off
Typical first use	Account research, case summaries	Renewal risk alerts, approval nudges	Order exception handling, data hygiene

Each step to the right is earned with evidence, and many workflows should stop at proactive.

1. Why should most workflows never go ambient?

This is where I part ways with much of the Dreamforce excitement. Ambient agents are the most impressive pattern to demo, but they are the right answer for only a small share of work.

The progression is not a ladder every use case must climb. It is a set of levels, and each workflow has a natural ceiling. A workflow should stop climbing when any of these conditions apply:

A mistake is expensive or hard to reverse. Pricing, credit, contract terms, and anything a regulator reads should retain human approval, however capable the agent becomes.

The judgment is the value. If a customer is paying for an expert’s assessment, automating that decision removes the very thing they value.

The data underneath is not owned. If no team is accountable for the quality of the data an agent acts on, the agent should not act on it independently.

That is not a lack of ambition. A well-run proactive agent that a manager approves in two seconds can deliver most of the value of an ambient one, at a fraction of the risk.

This is an important distinction. The objective is not to make every workflow autonomous. It is to identify where autonomy creates value without compromising accountability, trust, or control.

2. How do you score a real workflow?

Score each workflow against the four indicators, then establish its ceiling using the three stop rules above.

The following table illustrates how the framework applies to six workflows across manufacturing and financial services. These are representative scenarios, not examples attributed to specific customer implementations.

Workflow	Data autonomy	Trigger point	User engagement	Value driver	Right pattern today	Ceiling
Mfg: Pre-call distributor briefing (warranty, open cases, last shipment)	Reads CRM + ERP	Rep asks	Rep reads and decides	Minutes saved per call	Conversational	Proactive (auto-briefing before each meeting)
Mfg: Case spike on one product line	Monitors case data	Event threshold	Manager approves the customer notice	Issue caught before escalation	Proactive	Proactive: a customer notice is hard to reverse
Mfg: ERP shipment delay updates the opportunity and alerts the owner in Slack	Acts across ERP, CRM, Slack	Continuous	Escalation past contract threshold only	Work done with no hand-off	Proactive, moving to ambient	Ambient
FS: Pre-review client summary for an advisor	Reads holdings, interactions, requests	Advisor asks	Advisor reads and decides	Better-prepared meeting	Conversational	Conversational: the advisor’s judgment is the value
FS: Expiring documents or life-event signals	Monitors records	Schedule	Advisor reviews drafted outreach	Retention, compliance	Proactive	Proactive: client contact stays human
FS: Incomplete onboarding files routed back to operations	Checks fields, routes records	Continuous	Ops handles exceptions	Faster onboarding	Ambient (internal only)	Ambient, never client-facing

Two things stand out. The same organization can run all three patterns simultaneously, depending on the workflow’s risk, data readiness, and business value. The manufacturing examples also show why cross-system integration matters: Salesforce needs to work alongside ERP data, service information, automation, and Slack.

In financial services, ambient work is best suited to controlled internal processes such as document checks and operational routing, rather than client-facing decisions or communication.

3. How do you build enough trust to give an agent more autonomy?

Trust is built in layers, and each layer has to hold before the next one is added.

Reliable data. An agent is only as good as the information it is grounded in. Data quality, event integrity, masking, and secure test data come first. Gartner predicts that organizations prioritizing business context and semantics in AI-ready data will increase agentic AI accuracy by up to 80% and reduce costs by up to 60% by 2027.

Guardrails. Define what the agent may read, what it may write, and what it must never touch through permissions, not just prompts. A prompt is a request. A permission is a rule.

Human oversight. Start with approvals on every action. As accuracy is demonstrated, move towards approvals for exceptions rather than every routine action.

Gradually increasing autonomy. Widen the scope one workflow at a time, based on measured accuracy and adoption, not enthusiasm.

Once several agents are running, a fifth layer becomes necessary: orchestration. Multi-agent environments need clear rules for which agent owns which task, observability into what each agent did, auditability, and defined escalation paths to a named person.

4. Where does Slack fit in the maturity model?

Slack is increasingly where the progression becomes visible to users. It can serve as the primary work interface for reading Salesforce data, starting automation, creating content, and updating records without opening a CRM screen.

That matters for adoption. A proactive alert that lands in the channel where a team already works is more likely to be acted on. The same alert buried in a dashboard may go unnoticed.

Agents can also connect Salesforce, Slack, external systems, and customer-facing digital experiences, shortening the path from insight to action. But the more systems and surfaces an agent touches, the more important the governance layers become.

5. If I were starting tomorrow: the first 90 days

I usually meet a program after the launch excitement has faded. That perspective makes one thing clear: the programs that keep moving are the ones that do the unglamorous work first.

Rather than treating the first 90 days as a fixed implementation schedule, I’d approach them as four phases. The pace will depend on data quality, integration complexity, security requirements, use-case scope, and existing governance maturity.

Phase 1: Identify and prioritize the right use cases.

Gather candidate workflows, score each against the four indicators, and apply the three stop rules. Prioritize based on business value and readiness, not how impressive the demo looks.

Phase 2: Validate data and governance readiness.

Before building, assess data quality, event integrity, masking, grounding, access controls, and sandbox readiness. Confirm that the right teams own the data and that permissions and escalation paths are clearly defined.

Phase 3: Pilot one measurable workflow.

Start with one workflow that has a clear business outcome and appropriate human oversight. Agree on success measures upfront, including accuracy, time saved, and user adoption. Use the pilot to establish whether the agent has earned greater autonomy.

Phase 4: Expand with governance in place.

Before introducing additional agents, establish permissions, orchestration, monitoring, auditability, escalation paths, and a named owner for each agent. Expand autonomy only when accuracy, adoption, and controls have been demonstrated.

Ongoing: Meet users where they work.

Deliver approvals, pipeline reviews, account research, and service triage through Slack rather than adding another dashboard. Make the agent part of existing workflows, not another destination users have to remember to visit.

The 90-day mark should be a point to evaluate progress and readiness for expansion, not a deadline to make every agent autonomous.

How do you know a workflow has earned the next level?

Move a workflow up only when all four are true:

• Accuracy has held at its current level for an agreed period
• Users act on its output without rechecking every result
• Permissions and monitoring already cover the new actions it would take
• A named person owns its escalations

If any box is empty, the answer is “not yet”, and that is a healthy answer.

The bottom line

The organizations that win with agents will not be the ones with the most autonomous agents. They will be the ones that know exactly how much autonomy each workflow has earned, and can prove it to their business and their regulators.

Classify first. Pilot narrowly. Govern before agent number two. Let each workflow find its own ceiling. That is how an agent program keeps moving after the keynote glow fades.

Planning your next Agentforce initiative? Talk to the LevelShift team about identifying the right use cases, assessing your data and governance readiness, and building agents that deliver measurable business value within the right guardrails.

Contact LevelShift to get started

Frequently asked questions

What is an ambient agent in Salesforce?

An ambient agent runs continuously in the background, monitors signals across systems, and acts within defined limits. People step in only for escalations and exceptions.

Do we need to start with a conversational agent?

Not always, but it is the lowest-risk way to prove data quality and build user trust. Some use cases, such as event-driven alerts, can start at the proactive level with human approval on every action.

How do we know when a workflow is ready for more autonomy?

When four things are true: accuracy has held at its current level for an agreed period, users act on its output without rechecking every result, permissions and monitoring already cover the new actions it would take, and a named person owns its escalations. If any of the four is missing, wait.

Is ambient AI realistic in regulated industries?

Yes, within bounds. Start with internal tasks such as data hygiene and routing, and keep client-facing or financial actions under human approval.

What role does Slack play in an Agentforce rollout?

Slack can act as the user’s main interface for Salesforce data and actions, which raises adoption of proactive alerts and approvals. It also widens the surface the agent touches, so permissions and observability need to cover it.