1 of 20

AI AGENT: PROMISE VS. REALITY

ROI and Lessons Learned

CzechDreamin

CzechDreamin

2 of 20

  • Jim Stukas

Yosra Saidani

Man of Many Titles,

since we embrace change at Salesforce

(Currently Deployment Strategist at Salesforce, but wait a few weeks)

Salesforce MVP,

Paris WIT trailblazer group leader

FTD co-Organizer

Head of Innovation at Cloudity

About Us

CzechDreamin

3 of 20

The Promise of AI with Agentforce

CzechDreamin

4 of 20

The Promise of AI

​

Intelligent Automation

Improved Customer Experience

Increased Productivity

Reduced Human Involvement

24/7 Availability

CzechDreamin

5 of 20

Behind the Scenes: Between Expectations and Reality

​

Misunderstanding of the tool

Needs that could be addressed differently

Data quality and availability: a key factor that can often be improved.

Technical limitations

Maturity of business processes

Organizational limitations

CzechDreamin

6 of 20

6 Steps to Implement �an Agent for Maximum Impact

CzechDreamin

7 of 20

Success factors of a successful implementation

Workstreams

Incremental and iterative implementation that unlocks value

Start small with a simple use case and continue iterating

Foundational technology

Data 360 & MuleSoft (data & integration)

Organizational readiness and governance

Agentic CoE, operating model, governance, change management

Strategic roadmap and strategy for AI/Agents & Data

End-to-end roadmap for AI, agents, data, integration

1

2

3

4

CzechDreamin

8 of 20

A few headwinds

While we have clearly defined success factors, we still have a few hurdles to overcome

FOMU

Skill gap

People, process, data challenges

Unrealistic expectation

FOMO

Pace of innovation

Compliance and regulatory concerns

CzechDreamin

9 of 20

6 steps to implement an agent�for maximum impact

6

5

4

3

2

1

Executive alignment & AI strategy

​

Lay the foundation for success

Go-live & beyond

Configure & test your agent

Plan for scale

Define your

top use cases/agents

CzechDreamin

10 of 20

Step 1: Executive alignment & strategy

  • Don’t build agents for agents sake, use “right technology” to address business problems
  • Build agents that will drive�business value
  • Use crawl, walk, run approach
  • Define business objectives and identify blockers
  • Set short and long-term goals
  • Align on company values and ethics
  • Assess organizational readiness�and define success metrics-

What to do

Lessons learned

CzechDreamin

11 of 20

Step 2: Define your top use cases/agents

  • Identify use cases that align with business value and agentic technology
  • Build a quality data foundation
  • Address tech readiness with foundational tech
  • Design agents with jobs�to be done and experience in mind
  • Start small, scale wisely
  • Have clear business requirements around the agents jobs to be done
  • Define agent-to-human interaction points and auditing
  • Align on value

What to do

Lessons learned

CzechDreamin

12 of 20

Step 3: Lay the foundation for success

  • Define data and integration strategy and approach
  • Apply risk mitigation framework (people, business, technology & data)
  • Implement guardrails to mitigate these risks
  • Agent Instructions
  • Limiting Agent Access
  • Operating Model

What to do

Lessons learned

  • High-quality data is essential, as well as real-time data availability - Data Quality is Never “Done”
  • Involve stakeholders early to align on ethical, compliance, and operational guardrails
  • Pilot guardrails and data controls with a limited scope before scaling, to validate assumptions and avoid over-engineering

CzechDreamin

13 of 20

Step 4: Testing - More important than ever

​

  • Keep instructions concise ('Always'/'Never'). Avoid over-constraining with 'if...then' rules.
  • Architecture: Use modular actions in limited topics. Keep action outputs small and focused.
  • Safeguards: Define data access rules early. Build clear, testable human escalation paths.

What to do

*Beta Nov

​

  • Avoid too many topics
  • Focus on coverage of critical topics, user personas, negative test cases, edge scenarios, and **guardrail/toxicity testing.
  • Tooling: Use Testing Center to automate and accelerate test cycles (1000 cases / 10 concurrent).
  • Process: Always version your agent. Use realistic test data. Learn to read Event Logs for debugging.

Lessons learned

CzechDreamin

14 of 20

Step 5: Go live & beyond

  • Establish roles that will be responsible for agent management/monitoring
  • Applied defined measurement framework and compare agent’s performance against success criteria
  • Drive adoption and awareness
  • Deploying your agent into your production environment
  • Monitor the initial rollout closely for issues, using available reports / event logs (Observability - GA Nov’25)
  • Fine-tune the agent and expand its capabilities to meet future needs

What to do

Lessons learned

CzechDreamin

15 of 20

Step 6: Plan for scale

  • Execute your agentic AI and data strategy and roadmap
  • Identify your next set of use cases
  • Consider standing up Agentic CoE and Agentic Lab
  • Start your journey towards becoming an Agentic Enterprise

What to do

  • Ensure Business & IT collaboration and alignment
  • Foster culture of innovation and experimentation

Lessons learned

Trailhead: Build an AI Agent with Agentforce

CzechDreamin

16 of 20

On-the-Ground Experience with Agentforce

CzechDreamin

17 of 20

My customer project: a customer-facing agent on the B2B portal

One agent, six jobs to be done — orchestrating Salesforce-native AI and an external LLM behind a single conversation.

THE BACKLOG — 6 jobs to be done

1

Knowledge & troubleshooting�Answer technical questions for internal & external users from the knowledge base.

2

Case management�Create cases and add comments directly in the conversation.

3

Order information�Retrieve order details and live order status.

4

Ticket tracking�Follow up on existing support tickets.

5

RMA handling�Initiate and manage return requests.

6

Account self-service�Modify delivery address and account details; cancel an order.

MULTI-LLM ARCHITECTURE — one orchestrator prompt

Customer · B2B portal (Experience Cloud)

ORCHESTRATOR PROMPT�Single system prompt · detects intent · routes dynamically · shared session memory

routes by detected intent

Salesforce Einstein�Transactional CRM actions

orders · cases · RMA · address

Azure OpenAI · GPT-4o�Conversational & RAG

knowledge base · troubleshooting

Salesforce data — Orders · Cases · Accounts

Knowledge base — RAG grounding

CzechDreamin

18 of 20

Promise vs. reality in production: what held, what stayed fragile

The demo always works. Production is where you learn the truth — here is what we had to re-adjust on the real agent.

✓

Strict grounding killed most hallucinations�Answers grounded only on the certified knowledge base — no free invention. Sources cited so users can verify.

✓

Right model for the right job�Einstein for transactional acts, Azure GPT-4o for conversation. Each model does what it is actually good at.

✓

Native, familiar experience�Living inside Salesforce meant a short learning curve and fast adoption once trust was earned.

✓

Deterministic guardrails on risky actions�Cancel, RMA, address change all go through a confirmation step. The agent never acts blindly.

!

Reliability drops past ~8 instructions�Over-loading one agent with rules degrades it. We split topics and kept instructions short — Always / Never, not long if-then.

!

Prompt stability is hard-won�20+ iteration cycles on the master prompt. A small wording change could shift behaviour — every version had to be tested.

!

Only as good as the content�Poorly structured knowledge articles still produced weak or wrong answers. Data quality is never “done”.

!

Unrecognised intents & edge cases�New phrasings the agent never saw still slip through. We monitor them and enrich the prompt continuously.

The fix was less magic, more structure: grounding, short instructions, deterministic guardrails and human escalation are what turn an impressive demo into an agent you can trust in production.

​

WHAT WORKED WELL

​

! WHAT STAYED FRAGILE

CzechDreamin

19 of 20

Putting it together: the 6 steps applied to My Customer project

Customer-facing agent on the B2B Commerce portal (Experience Cloud) — retailers & distributors self-serve on orders, shipments, invoices and support.

1

Executive alignment

& strategy

IN PRACTICE

Customer Service flagged as the priority: cut inbound volume on historical channels, deliver 24/7 self-service. Crawl-walk-run agreed with the sponsor.

2

Define top

use cases

IN PRACTICE

6 jobs scoped with the business: KB / troubleshooting, case creation, ticket tracking, RMA, order info, order cancellation — value-ranked before build.

3

Lay the

foundation

IN PRACTICE

Multi-LLM design: Einstein for transactional CRM acts, Azure OpenAI GPT-4o for RAG. One orchestrator prompt routes by intent. Data access scoped per topic.

4

Configure

& test

IN PRACTICE

Concise Always/Never instructions, modular actions in few topics. Testing Center for coverage: personas, negative cases, edge & guardrail/toxicity tests.

5

Go live

& beyond

PRACTICE

Phased rollout on the authenticated portal, close monitoring via event logs. Clear human-escalation paths. Hypercare & continuous tuning.

6

Plan

for scale

IN PRACTICE

Wins industrialised into reusable assets (Cloudity Agent Factory). Next use cases identified; one repeatable pattern drives real Agentforce consumption.

Takeaway: the same 6 steps that frame the method are exactly the path we walked with Vusion — from a single scoped use case to a repeatable, value-driven agent.

CzechDreamin

20 of 20

Thank You

Q&A

CzechDreamin

CzechDreamin