EY GDS on Agentic AI, AI Value and operational transformation shows how autonomous voice agents and agentic systems are reshaping Global Capability Centers and customer-facing operations. This shift moves beyond chatbot deflection into genuine work automation: agents that pick up calls, understand intent on the first attempt, capture data to a CRM, and hand work to the right human or complete it entirely. The practical question is no longer whether agentic AI works, but where it delivers measurable return and where it still fails.

The technology is now moving into production at scale. As IT Voice Media reported, EY GDS is documenting how agentic AI creates value across cost reduction, speed, and customer satisfaction. Enterprises in India and across Southeast Asia are moving beyond pilots, with voice AI and conversational AI becoming standard infrastructure rather than experiment. The barrier is no longer capability. It is integration, data quality, and honest assessment of what tasks remain human-only.

What Agentic AI Actually Does in a Business Context

Agentic AI is not a chatbot that asks clarifying questions and passes customers to a queue. It is software that receives a task, decides whether it can handle it, executes steps autonomously, and reports the outcome. In a voice context, this means an AI agent answers a phone on the second ring, listens to the customer's full request, writes the key details to a built-in CRM, checks inventory or account status, and either solves the problem or pre-books a human callback with all context included. The distinction matters because it determines where time and money actually get saved.

Traditional IVR systems ask customers to press buttons or speak keywords. They route calls based on a menu tree. Agentic AI hears the full problem first. If a customer calls a support line and says their invoice is wrong, the agent retrieves the account, runs a validation check, identifies the error, and either corrects it or escalates with full documentation. A human does not spend 90 seconds re-gathering information already captured. This is where operational efficiency compounds.

The core mechanism is context retention. An agentic system holds the caller's situation in working memory throughout the conversation, adapts its responses, and flags escalation criteria before handing off. Competitors report that 75 percent of users repeat themselves across conversation turns because the system loses context. When context is maintained properly, first-contact resolution rates improve measurably. Industry benchmarks put first-call resolution at 65 to 75 percent with traditional support; agentic deployments typically report 80 to 88 percent on defined task categories.

EY GDS on Agentic AI: What the Guidance Actually Covers

EY's Global Data Services division has published analysis on how agentic AI maps to real business outcomes, moving beyond vendor claims into measured implementation. The guidance addresses three core areas: cost structure, capability limits, and organizational readiness. This is valuable not because EY has invented anything new, but because they have documented what succeeds and what remains stuck in pilot.

Cost reduction is the first metric. A typical customer service operation spends 40 to 60 percent of its budget on handling routine inquiries: password resets, invoice queries, order status checks, appointment rescheduling. If agentic AI can reliably handle 50 to 70 percent of these calls end-to-end, labor cost per contact drops by roughly 35 to 50 percent on that subset. For a GCC handling 10,000 calls per month at an average labor cost of £4 per contact, moving 3,000 calls to automation saves £42,000 monthly. This is real money, not theoretical efficiency.

The second metric is speed. A human taking a call goes through a mental checklist: identify the customer, pull their account, understand the issue, check systems, explain the resolution, and document the outcome. This typically takes 4 to 8 minutes for routine work. An agentic system can do the same in 90 to 120 seconds because it reads intent in parallel and takes no time to context-switch. If those 3,000 routine calls happen daily, reducing each by 3 minutes frees 150 labor hours per month without hiring anyone new.

The third metric, which EY emphasizes, is accuracy and compliance. Humans make data entry errors. They miss security checks. They deviate from scripts in regulated industries. Agentic systems execute the same process every time, log every action, and flag compliance issues before they become problems. For financial services or healthcare GCCs, this is often worth more than the pure labor saving.

Where Agentic AI Creates Value: Four Concrete Scenarios

Agentic AI works best on tasks where the outcome is deterministic and the data is clean. A customer calls to check a delivery status. The agent looks up the order number, sees it shipped on a certain date, provides tracking, and ends the call. Outcome: solved. A second customer calls to update their address on file. The agent captures the new address, validates it, updates the CRM, and sends a confirmation email. No ambiguity. These tasks, repeated thousands of times, are where agentic systems deliver measurable return.

Tech support for password resets is another classic. A user cannot log in. The agent verifies their identity through security questions or email link, resets the password, and walks them through a test login. This is procedural. The agent never needs to improvise. Time-to-resolution: 2 to 3 minutes with an agent, 60 to 90 seconds with agentic AI, because the system runs verification and reset in parallel rather than serially. On a 500-person SaaS company, this represents roughly 15 hours of human time per month reclaimed.

Appointment scheduling and rescheduling is a third high-value target. A customer calls to book a service appointment. The agent checks availability, presents three slots, captures confirmation, and sends calendar invitations. An agentic system does this faster and without the human lag between checking a calendar and offering times. Studies on call center automation show booking tasks see the highest first-call completion rates, typically 90 percent plus, because the outcome is unambiguous: the appointment either books or it does not.

Billing and invoice inquiry handling is the fourth scenario. A customer disputes a charge or requests clarification. The agent pulls the invoice, cross-references terms of service, explains the charge, and either corrects an error or confirms the amount is correct. Agentic systems do this with higher consistency than humans because they reference the same data source every time and flag unusual amounts for human review rather than making judgment calls.

The AI Value Question: Where ROI Actually Appears

Calculating AI value requires separating three types of return: labor cost reduction, revenue protection, and risk mitigation. Labor cost reduction is the most obvious and most often quoted. If an agentic system handles 1,000 calls per month that would otherwise need a human, and each contact costs £5 in labor, the system saves £5,000 monthly. Multiply by 12, subtract infrastructure and licensing costs, and you have a business case. Most vendors focus here because the math is clean.

Revenue protection is less visible but sometimes larger. When appointment no-show rates drop because customers receive automated reminders and confirmations, revenue from those appointments stays in the business. When order inquiry resolution improves, fewer customers abandon carts due to frustration. When billing disputes are resolved faster, cash flow improves and customer lifetime value increases. These are not spectacular numbers individually, but they compound.

Risk mitigation is critical in regulated industries. A compliance violation in a financial services call center can cost tens of thousands in fines plus legal time. If an agentic system enforces every required check and logs every step, that risk cost is lower. Similarly, if documentation is always complete and consistent, disputes are easier to defend. For a GCC in a regulated sector, this can justify investment before labor savings are even counted.

The honest assessment: most organizations see ROI between 6 and 18 months on pure labor cost basis, assuming they deploy on high-volume, low-complexity work. If they also capture revenue protection and risk reduction, it accelerates. If they deploy on low-volume or high-complexity work, payback extends past 24 months or may not materialize at all.

GCC Transformation: Why Agentic AI Reshapes Capability Center Economics

Global Capability Centers exist because labor is cheaper in certain geographies. This advantage has been narrowing as competition for GCC talent increases and wage pressure builds. A customer service agent in Bangalore now costs £6,000 to £8,000 annually versus £12,000 to £15,000 in Western markets, but that gap is closing. Agentic AI changes the math by shifting work from people to infrastructure, which has more favorable unit economics the higher the volume.

A GCC with 200 agents handling routine work can reduce that team to 100 to 120 agents using agentic systems for first-contact resolution, keeping the same throughput. The 80 freed agents move to higher-value work: complex troubleshooting, relationship management, complaint handling that requires empathy and judgment. This restructuring is not cost elimination. It is capability redistribution, which is more sustainable than pure headcount reduction because it retains institutional knowledge and improves agent engagement.

For GCC operators, agentic AI also solves a staffing volatility problem. Hiring and training an agent costs £2,000 to £4,000 and takes 4 to 8 weeks. Attrition in GCCs runs 20 to 35 percent annually in high-cost regions. If a system can absorb routine work without requiring new hires, staffing becomes more stable and hiring costs drop. For a 1,000-person GCC with 25 percent annual attrition, avoiding 250 hires saves £500,000 to £1,000,000 in recruitment and training.

The transformation also changes GCC positioning. Instead of competing on labor cost alone, GCC operators now compete on technology integration, process quality, and data security. This is more defensible long-term because labor cost arbitrage always has limits. Sysevo and similar platforms embed voice agents with CRM functionality so GCC teams spend less time on data hygiene and more on judgment-based work. This reshaping is already underway in India and Southeast Asia, according to CIO and industry reporting on AI agent sprawl and production readiness.

Conversational AI Trends: What's Shifting in 2026

Conversational AI is moving from narrow task automation toward broader context handling. The key shift is context retention across multiple turns. As bestmediainfo.com reported, 75 percent of Indian users repeat themselves across conversation turns because existing systems drop context. This is a data engineering problem disguised as a UX problem. Modern agentic systems hold caller context throughout the session, reducing frustration and improving perceived intelligence.

A second trend is multi-channel integration. Voice is no longer separate from chat, email, or messaging. A customer initiates a request via voice, continues it via chat, and resolves it via email without re-explaining the problem. This requires unified context across channels. Systems that manage this seamlessly see higher satisfaction scores and faster resolution. Those that treat channels as separate silos create friction.

The third trend is security and authenticity. As voice AI becomes more realistic, scam risk rises. Regulatory pressure is building for voice systems to identify themselves, disclose when they are AI rather than human, and maintain robust call recording and audit trails. Compliance with these requirements is becoming table stakes. Systems that cannot provide proof of consent, identity verification, and full audit logs will face regulatory friction in regulated sectors.

The fourth trend is integration depth. Early agentic systems integrated with specific CRMs or ticketing platforms. The direction now is toward CRM-first design where the CRM is not a destination for data but the native brain of the agent. Built-in CRM functionality means the agent doesn't copy data into a separate system. It reads and writes the CRM as its native interface. This eliminates sync delays and data loss in hand-offs.

Voice AI News and Real-World Deployment Lessons

Deployment speed is accelerating. The barrier is no longer whether voice AI works in a lab. It is whether it works in production on your messy data with your specific customer base. As reported in industry coverage of AI agent production struggles, many organizations deploy pilots successfully but stall moving to production because they underestimate data quality and integration work. A pilot might run on clean test data. Production requires handling typos, incomplete information, regional accents, and edge cases the test set never covered.

The most common failure pattern is poor intent detection. An agent misunderstands what a customer is asking and takes the wrong path. This happens when the training data is too narrow or the training process conflates similar requests. A system trained on 5,000 example calls might work for common scenarios but fail on the 100th distinct variation of the same basic request. Production systems need 20,000 to 50,000 quality examples and ongoing refinement.

The second failure pattern is escalation delay. A system should recognize quickly when a request requires human judgment rather than attempting multiple failed attempts. If an agent tries to solve a problem five times and fails each time, the customer is frustrated. A well-designed system recognizes failure after one or two attempts and escalates with full context. Training the system to know what it cannot do is as important as training it to do what it can.

Organizations reporting success emphasize process design over technology selection. Before implementing an agentic system, they map current workflows, identify which steps are truly deterministic versus judgment-based, and build the agent only for deterministic work. They also staff for hand-offs: if 20 percent of calls require escalation, they plan for that load and ensure human agents have full context when they pick up. Technology choices matter, but process design matters more.

When Agentic AI Fails: Trade-Offs and Honest Limits

Agentic AI does not work well on ambiguous or relationship-based work. A customer calls angry about poor service. They do not want a system to acknowledge their complaint and file a ticket. They want empathy, reassurance, and a commitment that a human will take ownership. An agent can do this. An AI system typically cannot, at least not in a way that feels genuine. Attempting to automate this work creates frustration rather than value.

Complex troubleshooting also remains difficult. A software developer calls technical support with a weird error that appears in three systems simultaneously. Diagnosing this requires creativity, hypothesis testing, and often lateral thinking. Agentic systems follow decision trees. They can ask scripted questions and check documentation, but they rarely synthesize across domain knowledge the way experienced humans do. This is where escalation is correct, not failure.

Data quality constraints are real. If your CRM has 30 percent missing phone numbers or customer records are scattered across three systems with no unique identifier, agentic AI will struggle. The system cannot look up an account if your database has duplicate entries or the lookup code is unreliable. Many organizations discover this constraint during pilot, realizing they need to fix data quality before the system can be effective. This work is not trivial. It is also not the vendor's job.

Small organizations should be cautious about early adoption. An agentic system makes economic sense at volumes above 5,000 interactions per month. Below that, the infrastructure cost relative to throughput makes the business case weak. A small legal practice with 200 calls monthly may be better served by a human receptionist than by maintaining a sophisticated voice AI system. The technology is not universally better. It is better in specific contexts at specific scales.

Implementation Path: From Pilot to Production

The standard implementation timeline is 12 to 20 weeks from evaluation to production. The first 4 to 6 weeks cover requirements gathering, data audit, and integration architecture. You are answering: Which workflows will the agent handle? What systems must it integrate with? What data quality work is needed? This phase is often underestimated and over-important. Getting the scope right here prevents rework later.

Weeks 7 to 12 cover build and training. Your team and the vendor's team define conversation flows, train models on your call data and documentation, and test in staging. Testing should include not just happy paths but error cases: what happens when the customer's account is locked? What if the system cannot find matching records? What if the customer says something completely unexpected? Testing rigor here determines production stability.

Weeks 13 to 16 cover pilot deployment. You route 10 to 20 percent of calls to the agent while monitoring performance. Key metrics: call completion rate, escalation rate, customer satisfaction, and data accuracy. If the system escalates 40 percent of calls, it is not ready. If it completes 80 percent correctly, you are on track. Use this window to gather real performance data and refine handling of edge cases.

Weeks 17 to 20 cover full rollout and stabilization. You gradually increase traffic to the agent while monitoring for degradation. You also staff for escalations, train agents on how to handle transferred calls, and track key metrics weekly. Stabilization is not immediate. Most systems improve significantly between week 2 and week 8 of production as they encounter real-world variation and the team tunes responses.

Technology Selection: What to Evaluate

Start with integration scope. How deeply does the system integrate with your CRM, ticketing platform, and backend systems? Integration that requires API calls and custom code is slower but more flexible. Integration via pre-built connectors is faster but more limited. If your CRM is custom or proprietary, integration will be harder and more expensive. Factor this into your evaluation, and do not assume easy integration unless the vendor proves it.

Second, evaluate context handling. Ask the vendor to demonstrate how the system maintains context across 10 conversation turns. Does it lose detail? Does it confuse who said what? Does it refer back to earlier details correctly? This is not a feature that shows well in marketing. It shows in practice. Request a trial with your own call recordings, not synthetic demo data, and listen to how the system performs on realistic variation.

Third, examine escalation design. How does the system decide when to hand off to a human? Is it based on keywords, confidence scores, or explicit rules? Can you customize escalation criteria? Can you change which human teams receive which escalations? An inflexible escalation system will frustrate your teams because calls will route to the wrong people. Flexibility here is worth paying for.

Fourth, consider governance and audit. What call recordings are available? How is consent managed? Can you export interaction logs for compliance? Can you audit which decisions the system made and why? For regulated sectors, these are non-negotiable. For consumer-facing businesses, they still matter because disputes will arise and you need to prove what happened.

Cost Models and ROI Calculation

Agentic AI is typically priced as a per-minute or per-interaction fee, plus a base platform cost. A realistic model: £0.10 to £0.30 per minute of agent conversation, plus £2,000 to £5,000 monthly platform fee. On 10,000 calls per month averaging 3 minutes each, that is £3,000 to £9,000 for the minutes plus the platform fee, totaling £5,000 to £14,000 monthly. If those calls would otherwise require a human at £5 labor cost each, the human-only alternative costs £50,000 monthly. Even at the high end, the automation saves 60 to 65 percent of contact cost.

Implementation and training costs are separate. Expect £15,000 to £40,000 one-time for integration, configuration, and team training, depending on your system complexity. This is not optional. It is not negotiable. If a vendor quotes £5,000 for implementation, either they are not building for your specific needs or they are under-scoping. Ask detailed questions about what is included.

Break-even is typically 4 to 9 months on labor cost basis alone. If your GCC spends £300,000 annually on routine customer support labor and an agentic system reduces that by £180,000, the system pays for itself in 2 to 3 months on labor, and the remaining 9 months of the year is pure savings. Add in risk mitigation and compliance value, and ROI becomes even clearer.

However, calculate what happens if the system solves 40 percent of calls instead of 60 percent. Your savings are proportional. If you projected 60 percent but achieved 40 percent, your break-even extends to 7 months instead of 4. This is why pilot performance is so important. Do not assume vendor projections will match your reality without testing.

Building the Business Case: Metrics That Matter

Start with baseline metrics. How many calls does your center handle monthly? What percentage are routine versus complex? What is the average handling time and labor cost per call? What is your first-call resolution rate today? What is your customer satisfaction score? These are your control group. Everything you measure after implementation is compared to these.

Then define agentic system targets. If the system will handle routine calls, project that routine calls represent 50 to 70 percent of volume. Project that first-call resolution on those calls will improve from 70 percent to 85 percent. Project that handling time per call will drop from 5 minutes to 2 minutes for calls the system handles. These are realistic targets based on typical deployments, not vendor claims.

Calculate the financial impact. If you handle 10,000 routine calls per month at £5 per contact with 70 percent current resolution, and the system handles 80 percent of volume at £1.50 per contact with 85 percent resolution, the impact is significant. You move from £50,000 monthly cost to roughly £21,000, saving £29,000 monthly or £348,000 annually. Subtract system costs of £120,000 annually, and net savings are £228,000 in year one.

But also calculate risk. What if resolution is 75 percent instead of 85 percent? What if the system handles only 60 percent of volume because it is overly conservative? What if integration takes twice as long as planned, delaying ROI by three months? A business case should include base case, upside case, and downside case. This helps executives understand confidence in the projection and makes you credible when outcomes vary.

Next Steps: Moving From Decision to Implementation

If you are evaluating agentic AI for your organization, start by auditing your top call drivers. Which calls are handled most frequently? How many of them are truly routine versus judgment-based? Which ones do your best agents handle the same way every time? These are your candidates for automation. Focus here first and you will see clearer ROI than trying to automate complex work early.

Second, assess your data readiness. Are customer records clean and complete? Can you reliably look up an account by phone number or customer ID? Can you update records in your CRM programmatically? If the answer to all three is yes, you are ready. If any is no, plan data remediation first. This work is not exciting, but it is mandatory.

Third, identify your integration points. Which systems must the agent read from or write to? Can the vendor integrate with those systems easily? Request a detailed integration technical specification from the vendor and have your team review it. Do not assume it will work. Verify it in writing before committing.

Fourth, start with a conversation. Book a call with the Sysevo team to discuss your specific scenario, call volumes, types of calls, and integration requirements. Do not start with a demo. Start with a requirements conversation. The right system for your business depends on your exact use case, and a good vendor will listen to that before showing slides.

Frequently Asked Questions

How is agentic AI different from a regular chatbot or IVR?

A chatbot answers questions based on pattern matching. An IVR routes calls based on button presses or keywords. Agentic AI listens to the full request, understands intent, integrates with backend systems, takes autonomous action, and reports the outcome. It solves problems rather than just routing them. This distinction determines whether a system saves labor or just deflects work.

What percentage of calls can an agentic system handle end-to-end?

On well-defined routine tasks like appointments or billing inquiries, systems typically handle 60 to 75 percent end-to-end with first-contact resolution. On broader mixed workloads, it is lower, around 40 to 55 percent, because mixed workloads contain more edge cases and complexity. Pure labor cost savings come from the high-percentage case. Mixed workloads need careful scoping.

How long does it take to see ROI?

Break-even on labor cost basis is typically 4 to 9 months, depending on call volume and implementation cost. If your GCC is large and handles high volumes of routine work, you reach break-even at the faster end. If your work is mixed or volumes are lower, ROI takes longer. Most organizations see positive ROI within 12 months on a reasonable deployment.

Do I need to retrain my human team if I deploy an agentic system?

Yes, but not the way you might think. Your team needs training on how to handle escalations from the system with full context pre-loaded. They also need to understand when to trust the system and when to override it. Most importantly, your best agents should help train the system by providing examples of how they handle calls. This is often a shift in their role rather than a threat.

What happens if the system makes an error or upsets a customer?

The system should escalate to a human as soon as the customer shows frustration. If the system detects it made an error, it should acknowledge it and hand off. Full call recordings and logs mean you can review what happened, identify the failure pattern, and refine the system. Errors happen in production. The system that learns from them improves over time.

Is agentic AI suitable for small businesses?

Agentic AI makes economic sense above 5,000 interactions per month. Below that, infrastructure cost relative to throughput makes the business case weak. A small business with 500 calls monthly might be better served by hiring a part-time receptionist. The technology is not universally better. It is better at specific volumes and for specific types of work.

How do I measure whether the system is working?

Track four metrics: call completion rate (percentage of calls handled without escalation), first-contact resolution (percentage of calls where the customer's issue was solved), customer satisfaction on calls handled by the system, and cost per contact. If all four are improving or meeting targets, the system is working. If any one is underperforming, investigate why before scaling.