A banking client recently asked me a question that I suspect many Swiss business leaders are pondering: “We’ve deployed chatbots, we’re using automation—what exactly is this ‘agentic AI’ everyone keeps talking about, and why does Gartner predict that 40% of enterprise applications will have AI agents by the end of 2026?”
It’s a fair question, especially given that less than 5% of enterprise applications currently incorporate agentic capabilities. The term “agentic AI” has become one of those phrases that sounds simultaneously revolutionary and impenetrably technical—the kind of jargon that appears in analyst reports and vendor presentations but rarely with concrete explanations of what it actually means or when you should care about it.
The reality, as is often the case with emerging technologies, is both more straightforward and more nuanced than the marketing suggests. Agentic AI represents a genuine evolution in how businesses can leverage artificial intelligence, but it’s not appropriate for every situation, and understanding the distinction between appropriate and inappropriate use cases may be the most important strategic decision technology leaders make in 2026.
Key Takeaways
For business leaders: Agentic AI refers to AI systems that can autonomously pursue multi-step objectives with minimal human intervention—not just responding to prompts, but planning, executing, and adapting. Gartner predicts explosive growth (40% of enterprise apps by end of 2026), and the technology is already generating measurable ROI at firms like Goldman Sachs. However, approximately 40% of initial deployments will fail, primarily due to attempts to layer agents onto legacy systems rather than redesigning workflows. The highest-value first use cases for Swiss SMEs are accounts payable/receivable automation, compliance checking, and intelligent data extraction—but only when implemented with proper workflow redesign and human oversight checkpoints.
What Is Agentic AI? (And What It Isn’t)
The confusion surrounding agentic AI is understandable, given that the technology landscape already includes chatbots, robotic process automation (RPA), and various other forms of “intelligent” systems. To cut through the noise, it helps to establish clear distinctions.
The Fundamental Difference
Traditional AI systems (including most chatbots and current automation tools) operate on a fundamentally reactive model: they wait for a human to provide a prompt or trigger, they execute a predefined response or workflow, and then they wait for the next instruction. The human remains firmly in control of both the objective and the execution pathway.
Agentic AI systems, by contrast, operate with a degree of autonomy that represents a qualitative shift in capability. When you provide an agentic system with an objective, it can:
- Decompose that objective into constituent sub-tasks without requiring you to specify each step
- Plan a sequence of actions to accomplish those sub-tasks, adapting its approach based on intermediate results
- Execute those actions across multiple systems and data sources, potentially over an extended timeframe
- Monitor its own progress, recognising when a chosen approach isn’t working and autonomously adjusting its strategy
- Recover from errors or unexpected situations without immediately escalating to human intervention
This is not merely a quantitative improvement in automation capabilities—it represents a fundamentally different paradigm in how AI systems operate within business processes.
A Practical Example: Invoice Processing
To make this distinction concrete, consider the seemingly straightforward task of processing a supplier invoice:
Traditional RPA approach:
- Human: Extract data from invoice PDF
- System: Run OCR, populate form fields
- Human: Verify account coding is correct
- System: Create entry in ERP system
- Human: Initiate approval workflow
Each step requires explicit programming, and any deviation from the expected format (a handwritten note, an unusual item description, a non-standard layout) typically requires human intervention to handle the exception.
Agentic AI approach:
- Human: “Process this invoice according to our standard procurement policies”
- Agent autonomously:
- Extracts data from invoice regardless of format variations
- Identifies the relevant purchase order in the ERP system (even if PO number is missing or incorrect)
- Verifies that pricing matches contracted rates, escalating discrepancies above defined thresholds
- Determines appropriate account coding based on historical patterns and company policies
- Routes for approval based on amount, department, and budget status
- Monitors approval process and sends appropriate reminders if needed
- Creates final accounting entries once approved
- Updates vendor records with delivery performance data
The critical difference is that the agent handles the entire end-to-end process, making dozens of micro-decisions along the way, adapting to variations in invoice format and content, and only involving humans when genuine judgment calls exceed its authority thresholds.
How Agentic AI Differs from RPA and Chatbots
| Dimension | Chatbots | RPA | Agentic AI |
|---|---|---|---|
| Interaction Model | Conversational responses to user queries | Automated execution of defined workflows | Autonomous pursuit of objectives |
| Planning Capability | None (responds to each input independently) | Minimal (follows script) | High (creates and adapts multi-step plans) |
| Error Handling | Requests clarification or escalates | Typically fails and alerts human | Attempts alternative approaches before escalating |
| Scope | Single conversation turn | Single predefined process | Multi-process workflows spanning systems |
| Human Oversight | After each exchange | At defined checkpoints | At decision thresholds only |
| Learning | Pattern recognition within conversation | None (deterministic rules) | Adapts strategies based on outcomes |
The Market Reality: Explosive Growth (And Significant Failure Rate)
The analyst projections for agentic AI adoption are striking not merely for their optimism but for the speed of the anticipated transformation.
The Gartner Prediction
According to Gartner’s 2025 strategic technology trends report, the percentage of enterprise applications incorporating agentic AI capabilities is projected to reach 40% by the end of 2026, up from less than 5% in early 2025. This represents one of the fastest technology adoption curves Gartner has tracked in recent years.
Market Size and Growth
The financial projections tell a complementary story. The global agentic AI market is expected to grow from approximately $7.8 billion in 2025 to $52 billion by 2030, according to Markets and Markets research. This represents a compound annual growth rate exceeding 45%—the kind of trajectory typically associated with genuinely transformative technologies rather than incremental improvements.
The Reality Check: Why 40% Will Fail
However, before we get carried away with enthusiasm, it’s crucial to acknowledge a sobering statistic that Gartner includes alongside its growth projections: they estimate that more than 40% of initial agentic AI deployments will fail to deliver expected value, with the majority of these failures attributable to a single fundamental mistake.
The mistake? Attempting to layer agentic AI capabilities onto existing legacy systems and processes rather than redesigning workflows to take advantage of the technology’s unique capabilities.
This pattern should sound familiar to anyone who has been in technology leadership for more than a few years—it’s essentially the same mistake organisations made when attempting to “digitise” paper-based processes by simply creating digital versions of paper forms, or when attempting to “mobilise” desktop applications by shrinking the interface to fit on a phone screen.
The organisations that will succeed with agentic AI are those that ask not “how can we make our current process autonomous?” but rather “if we were designing this process from scratch with agentic capabilities available, what would it look like?”
Real-World Deployment: Goldman Sachs and Claude Agents
While much of the discussion about agentic AI remains speculative or focused on future possibilities, there are already compelling examples of large-scale production deployments delivering measurable business value.
Goldman Sachs’s Implementation
Perhaps the most notable early adopter is Goldman Sachs, which has deployed Claude-powered agents across multiple business functions, with particularly significant implementations in accounting reconciliation and regulatory compliance workflows.
The accounting reconciliation use case: Goldman Sachs processes millions of transactions daily across dozens of systems and legal entities. Traditionally, reconciliation required armies of analysts manually comparing records across systems, investigating discrepancies, and documenting their findings—work that was both time-intensive and error-prone due to the sheer volume and complexity.
The Claude-based agentic system they deployed:
- Autonomously accesses data across multiple accounting systems and trading platforms
- Identifies discrepancies that require investigation based on materiality thresholds and risk factors
- Researches potential causes by examining related transactions, system logs, and historical patterns
- Generates preliminary explanations for discrepancies, citing specific evidence
- Routes items requiring human judgment to appropriate specialists with full context
- Documents all findings in the required regulatory format
Results: Goldman Sachs reports that the system has reduced the time required for certain reconciliation processes by more than 60%, while simultaneously improving the completeness and consistency of documentation—a combination of benefits that would be difficult to achieve through either pure automation or additional human staffing.
The compliance checking use case: Financial institutions operate under extraordinarily complex and constantly evolving regulatory requirements. Goldman Sachs has deployed agentic systems that monitor transactions and communications for potential compliance issues, automatically referencing the relevant regulatory requirements, and escalating items that require legal review.
The critical distinction from traditional rules-based compliance systems is the agent’s ability to understand context and nuance rather than simply matching keywords or thresholds. This significantly reduces false positives while improving the detection of genuinely concerning patterns that might not match predefined rules.
What Makes These Deployments Successful
Several factors distinguish Goldman Sachs’s successful implementations from the failed projects that Gartner predicts:
- Workflow redesign: Rather than attempting to automate existing manual processes step-by-step, they reconceived the workflows to leverage the agent’s strengths
- Clear authority boundaries: The agents have well-defined decision thresholds beyond which they must escalate to humans
- Comprehensive context: The systems have access to all relevant data sources rather than operating in isolation
- Continuous learning: The implementations include feedback loops that improve performance over time
- Appropriate use cases: Both accounting and compliance involve substantial information processing and rule application—tasks well-suited to agentic capabilities
When to Deploy Agentic AI: The Best First Use Cases for Swiss SMEs
For Swiss small and medium enterprises considering agentic AI, the question is not whether the technology has merit—the Goldman Sachs example demonstrates that it does—but rather which specific applications offer the best combination of business value, implementation feasibility, and manageable risk for organisations without Goldman’s resources or technical depth.
High-Value First Use Cases
Based on implementations I’ve guided for Swiss SMEs and the pattern of successful deployments globally, the following use cases consistently deliver strong ROI while remaining achievable for organisations new to agentic systems:
1. Accounts Payable and Receivable Workflows
Why this works well: AP/AR processes involve substantial routine decision-making within well-defined policies, make use of data from multiple systems (invoices, purchase orders, contracts, ERP), and offer measurable time savings and error reduction.
What the agent does:
- Processes incoming invoices in any format (PDF, email, paper scan, EDI)
- Matches invoices to purchase orders and receiving documents (three-way matching)
- Identifies and researches discrepancies in pricing, quantities, or terms
- Applies company-specific approval routing based on amount, category, and budget status
- Generates payment batches according to payment terms and cash flow policies
- Handles vendor inquiries about payment status by accessing multiple systems
Expected results: 60-75% reduction in processing time per invoice, 80%+ reduction in payment errors, improved vendor relationships through faster query resolution.
Swiss-specific advantage: The agent can seamlessly handle invoices in German, French, and Italian, applying the appropriate VAT treatment based on vendor location and transaction type—a complexity that often requires significant manual attention in multilingual Swiss businesses.
2. Regulatory Compliance Checking
Why this works well: Swiss SMEs face substantial compliance obligations (FADP, GDPR for EU clients, industry-specific regulations), and compliance work is both high-stakes and time-intensive.
What the agent does:
- Monitors business processes and data handling against compliance requirements
- Identifies potential issues based on current regulations (which it accesses from updated sources)
- Generates compliance documentation and audit trails automatically
- Prepares responses to data subject access requests
- Flags contracts or business practices that may create compliance risk
- Maintains a compliance calendar and alerts responsible parties to upcoming deadlines
Expected results: 70-80% reduction in time spent on routine compliance activities, improved completeness and consistency of documentation, reduced audit findings.
Swiss-specific advantage: The agent can simultaneously evaluate compliance with both FADP (Swiss) and GDPR (EU) requirements, correctly applying the more stringent standard where they overlap—a nuance that frequently causes confusion when evaluated manually.
3. Intelligent Data Entry and Document Extraction
Why this works well: Data entry remains one of the largest time sinks for SME employees despite decades of automation efforts, because source documents vary widely in format and quality.
What the agent does:
- Extracts structured data from unstructured or semi-structured documents (contracts, applications, reports)
- Populates multiple systems with consistent data from a single source document
- Identifies missing or potentially incorrect information and requests clarification
- Learns from corrections to improve extraction accuracy over time
- Handles multilingual documents, translating as needed for consistent storage
Expected results: 80-90% reduction in manual data entry time, dramatic reduction in entry errors, faster processing of customer requests or supplier documents.
Swiss-specific advantage: The agent handles documents in German, French, Italian, and English without requiring language-specific configuration—particularly valuable for Swiss companies operating across linguistic regions.
4. Customer Service and Support Escalation
Why this works well: Customer inquiries often require gathering information from multiple systems, making judgment calls about appropriate responses, and escalating complex issues—precisely the kind of workflow where agentic capabilities shine.
What the agent does:
- Receives customer inquiries through any channel (email, chat, phone transcription, web forms)
- Gathers relevant context from CRM, order history, support tickets, and knowledge base
- Resolves straightforward inquiries autonomously within defined authority limits
- Escalates complex issues to appropriate specialists with full context and preliminary research
- Follows up to ensure customer satisfaction and updates knowledge base
- Identifies patterns in customer issues that may indicate larger problems
Expected results: 50-70% of routine inquiries resolved without human intervention, 60% reduction in time specialists spend gathering context, improved customer satisfaction through faster response times.
Swiss-specific advantage: The agent can conduct the entire interaction in the customer’s preferred language while maintaining consistent internal documentation in the company’s primary language.
Use Case Comparison and Selection
| Use Case | Implementation Complexity | Time to Value | Annual Savings (Est.)* | Risk Level |
|---|---|---|---|---|
| AP/AR Workflows | Medium | 2-3 months | CHF 40,000-80,000 | Low |
| Compliance Checking | Medium-High | 3-4 months | CHF 30,000-60,000 | Medium** |
| Data Entry/Extraction | Low-Medium | 1-2 months | CHF 50,000-100,000 | Low |
| Customer Service | Medium | 2-3 months | CHF 35,000-70,000 | Medium*** |
*Based on a company with 30-50 employees. Actual savings depend heavily on current process volume and efficiency.
**Risk is medium because compliance errors can have regulatory consequences; requires particularly careful validation.
***Risk is medium because customer-facing systems require robust error handling to avoid reputational damage.
When NOT to Deploy Agentic AI
Understanding where agentic AI creates value is only half of the strategic equation. Equally important—perhaps more important—is recognising situations where deploying agentic capabilities is likely to consume resources without delivering proportional value.
Inappropriate Use Cases
1. Processes requiring genuine human judgment or creativity
Agentic AI is remarkably good at applying policies, following procedures, and making decisions within defined parameters. It is not good at making judgment calls that require weighing competing values, considering long-term strategic implications, or generating genuinely novel approaches.
Bad example: Deploying an agent to make hiring decisions or negotiate complex contracts.
Why it fails: These processes require weighing intangible factors, reading between the lines of what people say, and making decisions that balance multiple stakeholder interests in ways that can’t be reduced to rules.
2. Low-volume, high-variability processes
The value proposition of agentic AI depends partly on volume—the time savings compound when a process occurs dozens or hundreds of times weekly. For processes that occur infrequently and differently each time, the implementation investment may exceed the benefit.
Bad example: Deploying an agent to handle annual strategic planning or one-off merger integration tasks.
Why it fails: The agent won’t have enough repetition to learn patterns or optimise its approach, and the human effort required to specify requirements and validate outputs may exceed the effort required to simply do the work manually.
3. Processes where explainability is critical
While modern agentic AI systems can provide reasoning for their decisions, there remain situations where you need to explain not just what decision was made but why that specific approach was chosen—often to external auditors, regulators, or in legal proceedings.
Bad example: Deploying an agent to make decisions about customer creditworthiness or insurance claims that may need to be defended in court.
Why it fails: While the agent can explain its reasoning, the explanation may not satisfy legal or regulatory requirements for transparency and accountability in high-stakes decisions affecting individuals.
4. Processes embedded in legacy systems with poor data quality
This is perhaps the most common reason for agentic AI failure: attempting to deploy sophisticated AI on top of systems where the underlying data is incomplete, inconsistent, or unreliable.
Bad example: Deploying an agent to optimise inventory across multiple legacy systems that don’t properly synchronise and contain conflicting information.
Why it fails: The agent’s decisions can only be as good as the data it operates on. If the underlying systems contain contradictions and errors, the agent will either make poor decisions based on bad data or spend all its time attempting to resolve data quality issues rather than accomplishing the intended objective.
The Readiness Assessment
Before investing in agentic AI for any use case, organisations should honestly evaluate the following readiness factors:
Process maturity: Is the process well-defined with clear policies and decision criteria? If you can’t articulate how a human should handle the process, an agent certainly won’t be able to figure it out.
Data availability: Does the agent have access to all information required to make good decisions? If critical context exists only in people’s heads or in unstructured formats, the agent will be operating partially blind.
Stakeholder acceptance: Are the people affected by automation willing to trust the agent’s decisions and work collaboratively with it? If key stakeholders are fundamentally opposed, they will find ways to work around the system rather than with it.
Error tolerance: Can the business tolerate occasional mistakes while the agent learns? If a single error could create severe consequences, you need more extensive validation and testing before deployment.
Success measurement: Can you clearly measure whether the agent is delivering value? If you can’t measure success, you can’t optimise the implementation or justify continued investment.
Critical Success Factors: Why Workflow Redesign Matters
The single most important insight from organisations that have successfully deployed agentic AI is this: the technology implementation is the easy part. The hard part is rethinking how work should be done.
The Layering Trap
The most common failure pattern I observe when organisations attempt to implement agentic AI is what I call “the layering trap”—attempting to insert an AI agent into an existing workflow that was designed for human execution, without fundamentally reconsidering how the process should work.
Example of the layering trap:
A company decides to “automate” its expense approval process. The existing process works like this:
- Employee fills out expense form with receipts
- Manager reviews and approves/rejects
- Finance team codes expenses to appropriate accounts
- Accounting team enters into ERP
- Payment team schedules reimbursement
The layering approach attempts to have an AI agent execute steps 3-5, but leaves steps 1-2 unchanged. This approach captures some efficiency but misses the larger opportunity.
The Redesign Approach
A workflow redesign asks: “If we were designing this process from scratch with agentic capabilities available, what would it look like?”
Redesigned expense process:
- Employee photographs receipts as they occur (even before trip is complete)
- Agent extracts data, requests clarification on anything unclear
- Agent codes expenses to appropriate accounts based on policy and historical patterns
- Agent routes exceptions above defined thresholds to manager for approval
- Agent schedules payment according to policy
- Agent alerts employee if any expense violates policy (proactive compliance)
The difference: The redesigned process eliminates the formal expense report entirely, reduces manager involvement to exceptions only, and makes compliance checking proactive rather than reactive. The time savings are an order of magnitude larger than the layering approach.
Principles for Successful Workflow Redesign
Based on successful implementations, several design principles consistently distinguish high-value redesigns from failed layering attempts:
1. Invert the human-machine relationship
Traditional automation: Machine assists human in human-designed process Agentic redesign: Human oversees machine in machine-optimised process
2. Move humans to exception handling
Traditional automation: Human handles every case, machine provides information Agentic redesign: Machine handles routine cases autonomously, human handles exceptions
3. Build in continuous learning
Traditional automation: Process is static until someone redesigns it Agentic redesign: Agent learns from outcomes and human feedback, continuously optimising
4. Design for recovery, not perfection
Traditional automation: Extensive upfront effort to prevent errors Agentic redesign: Accept that errors will occur, build in rapid detection and recovery
5. Make decisions at the lowest appropriate level
Traditional automation: Important decisions escalate to humans Agentic redesign: Decisions are made at the point of maximum information availability, which is often the agent level, escalating only when judgment exceeds the agent’s authority
Swiss Compliance Considerations for Agentic AI
For Swiss SMEs, regulatory compliance isn’t merely a constraint to be navigated—it’s a competitive differentiator that must be preserved and enhanced as new technologies are adopted. Agentic AI implementations must be designed from the outset to meet and exceed Swiss data protection standards.
FADP Requirements for Autonomous Systems
The revised Federal Act on Data Protection (FADP), which came into force on 1 September 2023, introduces several requirements that have particular implications for agentic AI systems:
Transparency and explainability: Data subjects have the right to know when decisions affecting them are made by automated systems. For agentic AI, this means:
- Clear disclosure when an AI agent rather than a human has made a decision
- Ability to explain the factors and reasoning behind the agent’s decision
- Documentation of the data sources the agent consulted
Purpose limitation: Personal data can only be processed for the purposes for which it was collected. For agentic AI, this means:
- Careful specification of what tasks the agent is authorised to perform
- Technical controls preventing the agent from accessing data beyond what’s necessary
- Audit logs documenting what data the agent actually accessed during execution
Accuracy and data quality: Controllers must ensure personal data is accurate and up to date. For agentic AI, this means:
- Mechanisms for the agent to identify potentially outdated or incorrect data
- Processes for humans to review and correct agent-identified data quality issues
- Regular validation that the agent’s decisions are based on current, accurate information
GDPR Implications for EU Client Data
Swiss SMEs that process personal data relating to EU citizens must also comply with GDPR requirements, which include provisions specifically relevant to automated decision-making:
Article 22 restrictions: GDPR provides data subjects the right not to be subject to decisions based solely on automated processing that produce legal or similarly significant effects. For agentic AI implementations serving EU clients, this means:
- High-stakes decisions must include human review before finalisation
- Clear documentation of which decisions qualify as “legal or similarly significant effects”
- Processes for data subjects to request human review of automated decisions
Data Protection Impact Assessments (DPIAs): When processing is likely to result in high risk to individuals, GDPR requires a formal assessment. Agentic AI deployments often trigger this requirement, necessitating:
- Systematic description of the processing operations and purposes
- Assessment of necessity and proportionality
- Identification of risks to data subjects’ rights and freedoms
- Measures to address identified risks
Swiss-Compliant Deployment Options
Several infrastructure options allow Swiss SMEs to deploy agentic AI while maintaining compliance with Swiss data protection requirements:
Azure Switzerland Regions: Microsoft Azure operates data centres in Zürich and Geneva, with the following compliance features:
- All data processing occurs within Swiss borders
- FINMA-compliant configurations available for regulated industries
- Data residency guarantees in customer agreements
- Swiss-resident support staff for compliance inquiries
AWS Zurich Region: Amazon Web Services provides a Swiss region with comparable capabilities:
- Full data sovereignty within Switzerland
- Compliance certifications including ISO 27001, ISO 27017, ISO 27018
- Detailed audit logging for regulatory reporting
- Direct connections to Swiss financial networks
On-Premise with Claude Integration: For organisations requiring maximum control, Anthropic offers enterprise deployment options that allow running Claude models on-premise while maintaining access to updates and improvements:
- Complete data control with no cloud transmission
- Integration with existing security infrastructure
- Customisation for industry-specific requirements
- Higher initial cost but potentially lower long-term compliance risk
Compliance Checklist for Agentic AI Deployment
Before deploying any agentic AI system that processes personal data, Swiss SMEs should ensure the following requirements are addressed:
Pre-Deployment:
- Data Protection Impact Assessment completed and documented
- Legal basis for processing identified and documented for each data type
- Data minimisation principle applied (agent accesses only necessary data)
- Purpose specification clearly defined and technically enforced
- Vendor contracts include data processing agreements meeting FADP/GDPR requirements
- Data subject rights procedures updated to account for automated decision-making
During Operation:
- Comprehensive audit logging of all agent actions and data access
- Regular review of agent decisions for accuracy and compliance
- Process for data subjects to request human review of automated decisions
- Incident response procedures specifically addressing AI-related risks
- Staff training on compliance requirements for agentic systems
Ongoing Governance:
- Quarterly review of agent decision accuracy and bias
- Annual DPIA review and update based on actual usage patterns
- Regular testing of data subject rights request procedures
- Documentation updates reflecting any changes to processing purposes or methods
Implementation Roadmap: From Strategy to Production
Successful agentic AI implementation follows a structured methodology that balances ambition with pragmatism, moving from careful validation to scaled deployment as confidence builds.
Phase 1: Strategic Selection (Weeks 1-2)
The objective of this initial phase is to identify the specific use case that offers the optimal combination of business value, implementation feasibility, and manageable risk for your organisation’s first agentic AI deployment.
Activities:
- Inventory business processes currently consuming significant employee time
- Evaluate each candidate against criteria including volume, decision complexity, data availability, and stakeholder readiness
- Conduct preliminary ROI analysis for top 2-3 candidates
- Select final use case based on expected value and probability of success
Deliverables:
- Process inventory with automation opportunity scores
- Detailed business case for selected use case including ROI projections
- Risk assessment identifying potential obstacles and mitigation strategies
- Stakeholder buy-in from affected department leaders
Success criteria: Clear consensus on which use case to pursue and realistic expectations for outcomes and timeline.
Phase 2: Workflow Redesign (Weeks 3-5)
Rather than automating the current process as-is, this phase reconceives how work should flow when an autonomous agent is available to handle routine decisions and execution.
Activities:
- Map current state workflow in detail, identifying all decision points and data sources
- Design future state workflow optimised for agent capabilities
- Define specific decision thresholds where human review is required
- Specify data access requirements and integration points
- Create preliminary test scenarios covering both routine and edge cases
Deliverables:
- Current state and future state workflow documentation
- Decision authority matrix (agent autonomous | agent recommends + human approves | human only)
- Data flow diagrams showing all system integrations
- Test scenario library for validation
- Updated business case reflecting redesigned process
Success criteria: Stakeholders agree that the redesigned workflow is both achievable and superior to current state.
Phase 3: Technical Implementation (Weeks 6-10)
Development follows an iterative approach, building small functional increments, validating them thoroughly, and expanding based on feedback rather than attempting to build the complete system before any testing.
Activities:
- Configure agent access to required systems via Model Context Protocol (MCP) integration
- Implement core decision logic and exception handling
- Build human oversight interfaces for approval workflows
- Develop audit logging and compliance reporting
- Conduct extensive testing against scenario library
Deliverables:
- Functioning agent integrated with required systems
- Human oversight dashboard for approvals and monitoring
- Comprehensive audit logging for compliance
- Test results documenting accuracy and edge case handling
- Security assessment confirming no unintended data access
Success criteria: Agent successfully handles 80%+ of test scenarios within acceptable quality thresholds.
Phase 4: Pilot Deployment (Weeks 11-14)
Rather than immediate full production use, a carefully monitored pilot allows identification and resolution of issues before they can cause significant problems.
Activities:
- Deploy agent to handle subset of actual workload (typically 20-30% of volume)
- Intensive monitoring of all agent decisions and actions
- Daily review meetings with pilot team to identify issues
- Rapid iteration to address discovered edge cases
- Continuous comparison of agent performance vs. historical human baseline
Deliverables:
- Pilot performance metrics (accuracy, speed, error rate, escalation rate)
- Issue log with resolution tracking
- User feedback documentation
- Refined decision thresholds based on observed performance
- Updated implementation plan for full deployment
Success criteria: Agent matches or exceeds human baseline performance on quality while delivering measurable time savings.
Phase 5: Production Scaling (Weeks 15-18)
With confidence established through successful pilot, the agent progressively handles increasing workload while monitoring ensures sustained performance.
Activities:
- Gradually increase agent workload from pilot levels to full production volume
- Train all affected staff on working with the agent effectively
- Establish ongoing performance monitoring and alerting
- Document standard operating procedures including agent oversight
- Capture lessons learned for future agent deployments
Deliverables:
- Agent handling 100% of appropriate workload
- Comprehensive staff training completion
- Performance dashboard with ongoing KPI tracking
- Standard operating procedures documentation
- Post-implementation review with ROI validation
Success criteria: Sustained performance meeting or exceeding pilot results at full production volume.
Phase 6: Continuous Optimisation (Ongoing)
Successful agentic AI is not “set and forget”—ongoing attention ensures the agent continues delivering maximum value as business conditions evolve.
Activities:
- Monthly performance reviews assessing accuracy, efficiency, and user satisfaction
- Quarterly evaluation of decision thresholds and authority boundaries
- Identification of adjacent processes suitable for agent expansion
- Regular updates to decision logic reflecting policy changes or new regulations
- Continuous stakeholder feedback collection and incorporation
Deliverables:
- Monthly performance reports with trend analysis
- Quarterly optimisation recommendations
- Annual ROI assessment with expansion opportunities
- Updated documentation reflecting process evolution
Success criteria: Sustained or improving performance with expanding scope over time.
The 2026 Competitive Reality
There is a truth that Swiss business leaders must confront directly: while you are evaluating whether to explore agentic AI, some of your competitors have already moved beyond evaluation to deployment, and the competitive gap this creates compounds over time.
The First-Mover Advantage Window
Unlike many technologies where being a fast follower is the prudent strategy, agentic AI offers genuine advantages to early adopters that become difficult for later entrants to overcome:
Cost structure advantage: Organisations deploying agentic AI in 2026 are fundamentally restructuring their cost base, reducing the marginal cost of processing additional transactions, serving additional customers, or entering additional markets. Competitors who delay will find themselves competing against companies operating with 30-40% lower operational costs in key processes.
Talent attraction and retention: Knowledge workers increasingly prefer employers who equip them with advanced tools and free them from tedious work. Organisations known for effectively leveraging AI will attract stronger talent while competitors struggle to fill roles that consist primarily of manual processing.
Learning curve effects: Agentic AI systems improve with use, learning from outcomes and feedback. Organisations that deploy early accumulate months or years of learning that can’t be easily replicated by competitors starting later.
Customer expectation setting: As leading companies deploy agentic capabilities to deliver faster response times, 24/7 availability, and more proactive service, they reset customer expectations. Competitors who don’t match these capabilities increasingly appear outdated.
What Goldman Sachs Teaches Swiss SMEs
The Goldman Sachs implementation described earlier offers several lessons applicable even to much smaller organisations:
Lesson 1: Start with processes you understand deeply. Goldman chose accounting reconciliation and compliance checking—areas where they had decades of domain expertise. They didn’t attempt to automate processes they were still figuring out manually.
Lesson 2: Invest in integration infrastructure. The agents’ effectiveness depends entirely on their ability to access data across systems. Goldman invested substantially in the integration layer via MCP integration before deploying agents.
Lesson 3: Maintain meaningful human oversight. Even with 60%+ time savings, humans remain involved at defined decision thresholds. The goal is human-agent collaboration, not human elimination.
Lesson 4: Measure rigorously. Goldman established clear baselines before deployment and continues to measure agent performance against both historical human performance and ongoing human validation.
The encouraging news for Swiss SMEs is that while Goldman’s implementations are large-scale and sophisticated, the fundamental principles apply at any scale. A 30-person company can deploy an agentic system for AP/AR processing following the same methodology Goldman used for reconciliation, achieving comparable percentage improvements with dramatically lower absolute investment.
Next Steps: Your Agentic AI Readiness Assessment
The strategic question facing Swiss business leaders in 2026 is not whether agentic AI will transform how work gets done—that transformation is already underway. The question is whether your organisation will lead that transformation, follow it, or find itself disrupted by it.
The 10-Minute Readiness Self-Assessment
Before engaging external consultants or vendors, evaluate your organisation’s readiness by answering the following questions honestly:
Process Maturity:
- Do you have at least one high-volume, repetitive process that consumes significant employee time?
- Can you clearly articulate the policies and decision criteria for that process?
- Would different employees handling the same case reach consistent decisions?
Data Infrastructure:
- Is the data required for good decisions available in accessible systems rather than locked in paper or people’s heads?
- Is that data generally accurate and up to date?
- Do you have the ability to integrate disparate systems or are they isolated silos?
Organisational Readiness:
- Are business leaders willing to redesign workflows rather than just automating current processes?
- Are the employees who would work with the agent receptive to the idea rather than fearful?
- Can you dedicate a project team for 3-4 months to see an implementation through?
Success Measurement:
- Can you measure the current time and cost of your target process?
- Do you have the ability to track agent performance against defined KPIs?
- Is there executive commitment to a 12-18 month learning period rather than expecting immediate perfection?
If you answered “yes” to most questions in each category, your organisation is likely ready for agentic AI and should proceed to detailed use case evaluation and vendor selection.
If you answered “no” to multiple questions in a category, focus first on building that foundational capability rather than rushing into implementation. Attempting to deploy agentic AI without adequate process maturity, data infrastructure, or organisational readiness is the path to becoming part of Gartner’s 40% failure statistic.
Starting Your Agentic AI Journey
For organisations ready to move forward, I recommend the following immediate next steps:
1. Identify your champion process (1 week): Select the single use case that offers the best combination of business value, implementation feasibility, and stakeholder support. Resist the temptation to pursue multiple use cases simultaneously in your first implementation.
2. Map the current state exhaustively (1-2 weeks): Document precisely how the process works today including all decision points, data sources, exception handling, and quality checks. This baseline is essential both for redesign and for measuring improvement.
3. Engage experienced implementation support (ongoing): While agentic AI technology is increasingly accessible, successful implementation still requires expertise in workflow redesign, integration architecture, and change management. The difference between success and failure often comes down to having experienced guidance through the inevitable challenges.
4. Build your business case rigorously (1 week): Quantify both the hard savings (time, cost) and soft benefits (quality, scalability, employee satisfaction) you expect. Establish clear success metrics and commit to measuring them throughout implementation.
5. Secure executive sponsorship (ongoing): Agentic AI implementation requires sustained commitment through inevitable setbacks and learning periods. Executive sponsorship ensures the project receives necessary resources and support when challenges arise.
Get Expert Guidance for Your Implementation
Ready to explore what agentic AI could mean for your organisation?
I invite you to schedule a complimentary 60-minute strategy session where we will:
- Evaluate your specific processes against proven agentic AI use cases
- Identify the highest-ROI opportunities for your organisation
- Discuss implementation approaches that maintain Swiss data protection compliance
- Review successful case studies from similar organisations
- Provide realistic timelines and investment expectations
- Answer your questions about the technology, risks, and change management
This session is not a sales presentation—it’s a practical, technical discussion designed to help you make an informed decision about whether and how to pursue agentic AI for your specific situation.
During our conversation, we’ll use the eflury Method for agentic implementation, which has successfully guided Swiss SMEs from initial assessment through production deployment while maintaining the rigorous quality and compliance standards Swiss businesses demand.
Emanuel Flury is Switzerland’s first dedicated Claude automation consultant, specialising in agentic AI implementations that deliver measurable ROI while maintaining Swiss data protection standards. Based in Grenchen, he works with businesses across Switzerland and the broader DACH region to design and deploy autonomous AI systems that enhance rather than replace human capability.
References
-
Gartner. (2025). Top 10 Strategic Technology Trends for 2026: Agentic AI. Retrieved from gartner.com
-
Markets and Markets. (2025). Agentic AI Market - Global Forecast to 2030. Retrieved from marketsandmarkets.com
-
Anthropic. (2025). Goldman Sachs Case Study: AI Agents for Accounting and Compliance. Retrieved from anthropic.com
-
McKinsey Global Institute. (2025). The State of AI in 2025: Agents and Autonomous Systems. Retrieved from mckinsey.com
-
State Secretariat for Economic Affairs SECO. (2023). New Federal Act on Data Protection (nFADP). Retrieved from kmu.admin.ch
-
European Commission. (2024). General Data Protection Regulation (GDPR): Automated Individual Decision-Making. Retrieved from gdpr.eu
-
Microsoft Azure. (2025). Azure Switzerland Regions: Zürich and Geneva. Retrieved from azure.microsoft.com
-
Amazon Web Services. (2025). AWS Zurich Region: Swiss Data Residency and Compliance. Retrieved from aws.amazon.com
-
Forrester Research. (2025). The Total Economic Impact of Agentic AI Platforms. Retrieved from forrester.com
-
Federal Data Protection and Information Commissioner FDPIC. (2024). Guidance on Automated Decision-Making Under the Revised FADP. Retrieved from edoeb.admin.ch