What problem are we actually solving with GenAI in service?
Executives want faster, more accurate resolutions at lower cost without creating risk. Agents want trusted answers in-flow. Customers want clear steps that finish the job and a clean handoff when the bot cannot. Generative AI can summarise, draft, and guide, but it predicts text rather than truth. Trustworthy service needs models that are grounded in approved sources, measured against customer outcomes, and governed for privacy and safety. The NIST AI Risk Management Framework states that trustworthy AI must be valid, reliable, safe, secure, and accountable across the lifecycle.¹ ISO/IEC 23894 adds a formal process to identify, assess, and treat AI risks so controls are explicit and auditable.² These standards shift the conversation from novelty to operating discipline.
What can GenAI do today that reliably improves CX outcomes?
Generative AI helps through four proven mechanics. Agent assist retrieves relevant passages and drafts responses with citations so resolvers accelerate to the first useful step while staying accurate. Retrieval augmented generation reduces hallucination by grounding outputs in evidence from your corpus.³ Summarisation compresses long threads into decision-ready notes that speed wrap without losing nuance. Structured extraction converts free text into fields that drive forms or entitlements. Guided next steps turn policy into small, sequenced actions that agents can follow under pressure. A recent survey of hallucination research warns that ungrounded generation can produce fluent errors, which makes retrieval and citation non negotiable in regulated environments.⁴
Where does GenAI belong in the contact centre stack?
Platforms slot GenAI beside knowledge, routing, and WFM. The pattern is simple. A retrieval layer indexes approved sources with metadata. An orchestrator checks policy and identity. The model composes a draft answer from retrieved chunks and shows citations. A logging unit records the sources used, prompts, and outcomes for audit. This loop depends on good content and good retrieval. Studies show that models can miss the most relevant spans when context is long, which is why ranking and chunking matter as much as generation.⁵ The outcome is practical. Agents see the right paragraph, not a search result page. Customers see clear steps and a link to the source.
Which use cases are practical right now?
Start with frequent jobs that have clear rules and verifiable end states. Billing explanations, password resets, outage lookups, appointment changes, ID and entitlement checks, and policy explanations are stable. Retrieval augmented generation drafts the answer and cites the exact policy or knowledge article so resolvers trust it.³ Agent wrap summaries and post interaction emails benefit from summarisation because the system already holds the transcript. Complex complaints and vulnerable customers benefit from guided next steps because policies are strict and consequences are high. The goal is to raise First Contact Resolution and reduce repeat within seven days by making the first capable resolver decisive.⁶
How do we measure value without vanity?
Use HEART’s goal, signal, metric discipline to bind GenAI to outcomes. Set a goal such as faster resolution. Pick signals that move in days such as grounded answer rate and time to first useful step. Track lagging outcomes such as task completion and First Contact Resolution after handoff.⁷ FCR remains the crisp proof that a case resolved first time when human help was required.⁸ Measure repeat within seven days on the same issue to confirm the bot did not defer work. This pairing shows value to both operations and finance.
What is the minimal technical pattern to ship safely?
Teams can deliver value with a small, auditable flow. Index approved sources. Retrieve the top passages for a question. Compose a draft that quotes or cites those passages. Refuse to answer if retrieval fails. Redact personal information in prompts and outputs. Restrict retrieval to content a user can already access. OWASP’s guidance for LLM applications lists concrete defenses against prompt injection and data exfiltration that should live in this flow.⁹ The pattern is not expensive. It is careful.
How do privacy and law shape deployment in Australia?
The Australian Privacy Principles require informed, specific, current, and voluntary consent, with purpose limitation and rights to access and correction. Deployments must check purpose at entry and at send, not just at onboarding.¹⁰ Logs should prove that consent existed and that prompts did not expose disallowed data. This is practical engineering. Add a consent record to the context. Filter sensitive fields before prompts. Mask card data and use out of band capture for payments so PCI duties remain controlled. Programs that instrument privacy on day one scale faster because risk and audit stay aligned.
What operating model makes GenAI sustainable?
Run GenAI as a product, not a project. A cross functional team ships small changes weekly. Product owns the journey and the KPI set. Data and ML own retrieval quality, evaluation, and drift. Platform owns integration, reliability, and cost. A knowledge lead owns content standards because retrieval quality depends on clear titles, plain language, and active maintenance. KCS provides a practical cadence for continuous article improvement in the flow of work.¹¹ MLOps practices add versioning, tests, and monitoring so changes do not surprise customers.¹² The rhythm is stable. Observe, fix, ship, measure, repeat.
What does a 60 day rollout look like?
Days 1–20: Baseline and guardrails.
Pick two intents. Inventory sources and owners. Chunk long articles. Add synonyms customers actually use. Enable retrieval, citations, and fail closed behaviour when sources are missing.³ Add PII redaction and role based retrieval.⁹ Log consent and purpose.¹⁰
Days 21–40: Agent assist first.
Launch in the agent desktop. Measure grounded answer rate, citation coverage, time to first useful step, and wrap time. Fix retrieval and content issues before customers see them.⁷
Days 41–60: Customer facing thin slice.
Expose one intent with explicit escalation. Measure task completion, FCR after handoff, and repeat within seven days against matched controls. Promote only when outcomes move in the right direction.⁸
What are the most common pitfalls and how to avoid them?
Teams deploy ungrounded chat and then wonder why answers drift. Fix by mandating retrieval and citations.³ Teams chase containment rather than completion. Fix by measuring completion and FCR after handoff as the primary outcomes.⁷ ⁸ Teams roll out everywhere without privacy controls. Fix by instrumenting consent and purpose and by redacting PII in prompts and outputs.¹⁰ Teams treat content as a one off. Fix by adopting KCS so frontline teams improve articles in the flow of work.¹¹ Teams assume long context will fix retrieval. Fix by improving chunking and ranking because models lose relevant facts in the middle.⁵
What impact should executives expect when these patterns are followed?
Expect earlier movement in grounded answer rate and time to first useful step within weeks. Expect measurable gains in task completion and First Contact Resolution after handoff for targeted intents within one to two cycles. Expect lower repeat within seven days where answers cite sources and where escalation passes identity and last step to the agent. Expect cleaner complaint trends for status opacity and policy misunderstanding as explanations become consistent. These gains reduce cost to serve because the system helps the first capable resolver finish the job.
FAQ
What is the fastest safe way to use GenAI in service?
Start with agent assist on one intent. Require retrieval and citations, redact PII, and restrict retrieval by role. Measure grounded answer rate and time to first useful step, then expand after outcomes improve.³ ⁷ ⁹
Why insist on retrieval augmented generation instead of pure chat?
RAG grounds the model in approved sources and cites them. This reduces hallucination and creates auditability, which is essential in regulated environments.³ ⁴
Which metrics belong on the board pack?
Task completion, First Contact Resolution after handoff, repeat within seven days, grounded answer rate, time to first useful step, and privacy and safety controls such as redaction success and prompt injection blocks.⁷ ⁸ ⁹ ¹⁰
How do we keep answers current?
Adopt KCS. Keep articles short, scannable, and written in customer words. Assign owners and review high reuse content regularly. This improves retrieval and agent trust.¹¹
Does longer context always help?
No. Research shows models can get “lost in the middle.” Good chunking, ranking, and focused context windows outperform naive long context.⁵
What governance frameworks should we align to?
Use NIST AI RMF for trustworthy AI functions and ISO/IEC 23894 for risk management. Map privacy to the Australian Privacy Principles.¹ ² ¹⁰
Sources
-
Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST, 2023, National Institute of Standards and Technology. https://www.nist.gov/itl/ai-risk-management-framework
-
ISO/IEC 23894:2023 — Information technology — Artificial intelligence — Risk management — ISO/IEC, 2023, International Organization for Standardization. https://www.iso.org/standard/77304.html
-
Retrieval-Augmented Generation for Knowledge-Intensive NLP — Patrick Lewis, Ethan Perez, Aleksandra Piktus, et al., 2020, NeurIPS. https://proceedings.neurips.cc/paper_files/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html
-
A Survey on Hallucination in Natural Language Generation — Zhijing Jin, Nelson F. Liu, et al., 2023, ACM Computing Surveys. https://dl.acm.org/doi/10.1145/3571730
-
Lost in the Middle: How Language Models Use Long Context — Nelson F. Liu, Kevin Lin, et al., 2023, arXiv. https://arxiv.org/abs/2307.03172
-
Stop Trying to Delight Your Customers — Matthew Dixon, Karen Freeman, Nicholas Toman, 2010, Harvard Business Review. https://hbr.org/2010/07/stop-trying-to-delight-your-customers
-
Measuring the User Experience at Scale (HEART Framework) — Kerry Rodden, Hilary Hutchinson, Xin Fu, 2010, Google Research Note. https://research.google/pubs/pub36299/
-
First Contact Resolution: Definition and Approach — ICMI, 2008, ICMI Resource. https://www.icmi.com/files/ICMI/members/ccmr/ccmr2008/ccmr03/SI00026.pdf
-
OWASP Top 10 for LLM Applications — OWASP Foundation, 2023, OWASP. https://owasp.org/www-project-top-10-for-large-language-model-applications/
-
Australian Privacy Principles — Office of the Australian Information Commissioner, 2023, OAIC. https://www.oaic.gov.au/privacy/australian-privacy-principles