DECIDE WHAT THE PILOT MUST PROVE
End the pilot with a defensible production decision.
The pilot should leave the team able to explain why the use case should move forward, stop or change direction.
A credible decision combines business value, observed behavior, control boundaries and the effort required to operate the use case after go-live.
DEFINE THE DECISION FIRST
Design the pilot around the decision you need to make.
Do not start with “Can Agentforce do this?” Start with “What would we need to see before we would trust this use case in production?”
- What business improvement must the pilot demonstrate?
- Which behavior would make the use case unacceptable?
- Which permissions and actions must remain bounded?
- Where must a person review or take over?
- What operating effort would make the use case not worth scaling?
These criteria turn the pilot from an experiment into a decision process.
TEST HOW IT BEHAVES
Do not test only whether the feature works. Test how it behaves.
A pilot can be functionally correct and still be unsuitable for production. The more important questions are often about consistency, boundary handling and what happens when the input is incomplete or unusual.
- Does the output remain useful across realistic variations?
- Does the system stay inside its permission and action boundaries?
- Does it stop or hand over when the task becomes ambiguous?
- Can a reviewer understand which information influenced the result?
- Are failure and edge-case behaviors predictable enough to support?
Production readiness is about observable behavior, not only successful execution.
A PILOT CAN SUCCEED BY STOPPING
A no-go can be the most useful result.
A pilot has done its job if it reveals that the use case is too unreliable, too expensive to operate, too dependent on weak data or too difficult to control.
Stopping early is cheaper than turning a weak pilot into a production dependency.
- The expected value is too small for the operating effort
- The required data or knowledge is not trustworthy enough
- Permission or action boundaries cannot be made sufficiently narrow
- Edge cases require too much manual intervention
Further steps (1)
- Monitoring and support would cost more than the benefit justifies
- Bounded task
- Realistic tests
- Evidence
- Go
- Revise
- No-go
- Controlled operations
Frequently asked questions about Agentforce pilots
How long should an Agentforce pilot run?
Long enough to cover a representative set of real cases and edge conditions. Calendar duration matters less than whether the pilot has produced enough evidence to judge value, behavior and operating fit.
How many use cases should we include in one pilot?
Usually one bounded use case is enough. Adding several tasks at once makes it harder to understand which data, permissions or behaviors are responsible for the result.
Does a successful pilot automatically mean go to production?
No. A prototype can work functionally and still fail the production test because the value, controls, handoff, monitoring or operating effort are not good enough.
CLOSE THE PILOT PROPERLY
Close the pilot with a defensible go/no-go decision.
Do not let the pilot drift into production because “it mostly worked”. End it with a documented decision against criteria that were agreed before the final review.
Go — Proceed with the agreed production scope. Revise — Change the task, data, permissions or operating model and retest. No-go — Stop because the evidence does not justify production.
- Business value
Did the use case improve the work enough to justify continuing?
- Behavior
Was the output consistent enough across normal and edge cases?
- Controls
Did permissions, actions and human handoff remain inside the agreed boundaries?
- Operations
Can the team monitor, support and change the use case with reasonable effort?
Further steps (1)
- Data readiness
Is the required context trustworthy enough for production use?
PRODUCTION GATE
Minimum criteria for production approval
Approval should require more than a good demo. At minimum, the team should be able to answer the following questions with evidence.
Value
Has the pilot demonstrated a useful business improvement?
Data
Is the required data and knowledge trustworthy enough for the task?
Permissions
Are read, write and action permissions restricted to what the use case actually needs?
Handoff
Is the point of human review or takeover explicit and tested?
More approval questions (4)
Behavior
Have realistic, incomplete and edge-case inputs been evaluated?
Monitoring
Can the team detect quality, failure and operating issues after go-live?
Ownership
Is it clear who owns the use case, configuration, review and support?
Change
Is there a controlled process for changing prompts, actions, data or scope?
If the evidence is missing, the production decision is still incomplete.
If it is still unclear whether the task needs an agent at all, start with AI in Salesforce: Which tasks are worth it?.
PILOT FLOW
Six steps from use case to production decision.
Keep the pilot narrow enough that every step produces evidence for the final decision.
01
Define the scope
Choose one task, one outcome and the limits of what the pilot is meant to prove.
02
Connect the context
Provide only the data and knowledge the use case genuinely needs.
03
Control the actions
Define which actions Agentforce may take, which require approval and where it must stop.
04
Evaluate behavior
Test normal cases, edge cases, incomplete inputs and known failure conditions.
05
Design the handoff
Make sure a person receives the original request, relevant context and open decision when taking over.
06
Establish operational evidence
Document monitoring, ownership, support and the criteria for the final go/no-go decision.
The pilot should make the production decision easier — not create another round of discovery.
LEARN SOMETHING USEFUL
Make the pilot informative even if you decide not to launch.
TEST CASES
Test beyond ideal conditions
Include incomplete information, ambiguous requests, permission limits and realistic edge cases — not only examples designed to work.
IMPACT
Assess business impact and AI quality separately
A response can look good while the use case still creates too little value or too much operating effort.
SCOPE
Make scope changes explicit
If the pilot starts requiring more data, permissions or actions, document the change instead of silently expanding what is being tested.
A pilot is valuable when it reduces uncertainty — even when the answer is no.
Sources and further reading
Primary Salesforce documentation and reference material for Agentforce, AI controls and production-readiness considerations.
- Salesforce – How Agentforce worksChecked on
Supports Agentforce piloting, testing, trust, and operational requirements; recheck product and security details before publication.
- Salesforce – How We Build Trusted AIChecked on
Supports Agentforce piloting, testing, trust, and operational requirements; recheck product and security details before publication.
- Salesforce Well-Architected – ResilientChecked on
Supports Agentforce piloting, testing, trust, and operational requirements; recheck product and security details before publication.
Topics in this article
THREE WORKING VIEWS
Three views that make an Agentforce pilot easier to trust.
Three illustrative configured workspaces show scope, handoff and decision history. All records and test results are fictional.
Realistic roles and action boundaries
Illustrative configured UI · Fictional data.
Swipe with your finger or drag with the mouse button held down. You can also select an image or use the arrow keys.
Image 1 of 3: Realistic roles and action boundaries
Continue reading
AI in Salesforce: Which tasks are worth it?
Not every Salesforce task gets better with AI. The useful starting point is a specific piece of work where interpretation, summarization or drafting creates a real advantage over a simple workflow or automation. The best first use case is narrow enough to test, useful enough to matter and controlled enough to operate.
Read article
The Data Foundation for AI: What Agentforce and Data 360 Actually Need
AI and Data 360 do not need every field to be perfect. They do need the important data to mean what teams think it means — with enough ownership, evidence and context to use it safely. The practical goal is not “clean all the data.” It is to make the critical data trustworthy enough for the process you want to improve.
Read article
Escaping the Monolith: Designing Scalable Multi-Agent Ecosystems in Agentforce
A practical guide to replacing one overloaded Agentforce agent with a lean router, bounded specialists, explicit handoffs, and evidence that the architecture scales.
Read article