How to Avoid the Biggest Pitfalls in Agentic AI Deployment

In June 2026, SSON hosted a webinar with Hypatos and EY on "The Power of Thinking Differently: Why Process Mapping Falls Short in the Age of Agentic AI." 

Global Business Services leaders keep making one costly mistake with agentic AI: treating it like the next RPA. The pattern is familiar: find the broken step, bolt on automation, measure the local win. It works, but it leaves most of the value on the table. For those of you who missed the webinar, here is what the speakers, Moritz Treutwein, Head of Partnerships, Hypatos, Nitee Gupta, EY and Jimmy Marquis, EY, are changing in this playbook.  

RPA Thinking vs. Agentic Thinking

• RPA vs. Agentic AI: RPA needed stable, linear, fully-mapped processes. Change one column header and the whole bot chain breaks.  Agentic AI, meanwhile, can interpret  human language, so your work instructions can be easily created and adapted by anyone in the GBS team. Furthermore, it can reimagine the process, skip irrelevant steps, and work toward an outcome non-linearly.
• The single biggest implementation mistake: applying RPA-style rigor (rigid rules, exhaustive edge-case scripting) to agents. It caps the value the technology can deliver.  Therefore, instead of coding every exception, give agents policies and principles that let them handle a whole population of edge cases at once.

Redefine the Outcome, Not Just the Task

• The old playbook was: find where invoices get stuck, then deploy AI on that node, thereby lowering cost-per-invoice. This was valid, but incremental.
• The better playbook: start from a bigger outcome (e.g. the vision of paying invoices fast enough to capture 2% savings through 10-day early-payment discounts) and work backward. This may mean touching cash positioning, approvals, and human validation steps that were never in scope before.
• This means when you're starting to implement Agentic AI, you should ask, "Is there a greater outcome we can drive for the organization?" and not just "where is work getting stuck?"

Metrics: Track Evolution, Not Just Automation Rate

• Traditional KPIs (cost per invoice, invoice per FTE) still matter. They're not being replaced, but they can't be the only things you look at to prove Agentic AI value.
• The real question is what business impact does that automation rate actually produce? Without standardized processes and accessible knowledge, a high automation rate doesn't translate into business value.
• Before releasing headcount, decide what that capacity gets redirected toward; otherwise the ROI case erodes over time. Are you able to scale without recruiting more people, take on more judgement-based activities? What is it that your freed-up people are now responsible for? 

Data Readiness: Don't Wait for "Perfect"

• It's a myth that you need clean and complete data before you can start with Agentic AI. In reality, agents can pull data just-in-time (the same way skilled staff do today), and they can actively flag and help fix inconsistencies as they process.
• Nearly every implementation surfaces vendor master-data inconsistencies; expect it, don't let it block the start.
• Incomplete data is generally safer than inaccurate data: agents can be directed to fill gaps and self-check; inaccurate data risks confident, undetected errors (hallucination-style behavior, since models tend to try to satisfy the request either way).
• A second data dimension is often missed: institutional knowledge. SOPs, handbooks, and the tacit "how we actually do this" living in people's heads need to be codified into a knowledge layer; agents will treat whatever they can access as ground truth and won't question it.

Readiness: Six Areas to Check Before You Start

1. Process readiness: Your processes need to have a clear start, clear end, clear outcomes/decisions, and governance that lets the AI operate within real boundaries (without this, "AI will just scale the chaos").
2. People readiness: AI literacy needs to be enterprise-wide, starting with leadership vision, not confined to a technical team. The goal is business-fluent staff who can direct and judge agent output.
3. Technology readiness: Architecture that integrates with existing systems and is built for what's coming next, with enterprise-grade data security from day one.#
4. Policy readiness: Most organizations believe their processes and policies are documented; in practice, there's usually a real gap between what's written and what's actually followed. Surface this early.
5. Data readiness: This covers three things: (a) baseline data quality, (b) willingness to acquire new data sources agents can access in real time, and (c) codifying institutional knowledge into a usable knowledge layer.
6. Leadership readiness: You need visible, committed leadership that gives people agency in the transformation rather than having it done to them. Leadership also needs to clearly communicate the purpose of Agentic AI to prevent fear of job loss and displacement. 

Picking the First Use Case (It's Not What You Think)

• Be bold: Whilst many organizations tend to trial Agentic AI on smaller, edge use-cases, our panel presented an alternative approach. The best proof-of-concept and ROI comes from high-volume, fairly standardized processes with historical data available (e.g., invoices, sales orders, other semi-structured documents). With these high-volume use-cases you'll clearly be able to show ROI and get the buy-in for future initiatives.  
• Use the first deployment to start building durable agentic capability (an internal "Agentic CoE") that can monitor, identify new use cases, and drive continuous improvement, not just to ship one automation.

Timeline Reality Check

• With prerequisites in place (documented processes, accessible knowledge, reasonable data), a pilot can go from proof to production in roughly 8–12 weeks, followed by a wave-based rollout.
• Proving a solution works against real, messy data can happen within days, but production-readiness still depends on the same readiness factors above.
• Deployment isn't a one-time event: ongoing monitoring and continuous improvement are part of the model, not an afterthought.

For more insight into how to accelerate your journey to an Agentic GBS, watch the full webinar here.