The Problem Nobody Is Talking About Loudly Enough
Enterprise AI systems are getting smarter, faster, and more connected to business data than ever before. Yet a quietly persistent problem is undermining their reliability in live customer operations: context decay. According to a recent conversation on CX Today between Francesca Roche and Matt Graney, Chief Product Officer at Celigo, AI can produce surprisingly unreliable outcomes even when it has access to enormous volumes of business data. The culprit is not bad data or underpowered models. It is the degradation of contextual relevance as AI moves from simple question-answering into complex, multi-step agentic tasks.
Context decay refers to what happens when an AI system loses the thread of a conversation, a workflow, or a customer situation across multiple interactions or process steps. The model may have all the raw data it needs — order history, account status, prior tickets, sentiment signals — but as the task grows longer or more complex, earlier context gets diluted, misweighted, or simply dropped. The result is an AI that sounds confident but is operating on an incomplete or distorted picture of reality.
Why This Matters More in Customer Service Than Anywhere Else
In a back-office automation context, context decay is a nuisance. In a live customer interaction, it is a trust-destroying event. Consider an agent-assist tool that pulls the right account data at the start of a call but fails to carry forward the customer's stated urgency, their prior complaint history, or the resolution that was promised two contacts ago. The AI's output becomes subtly — or not so subtly — wrong. Human agents acting on that output make errors. Customers feel unheard. Escalations rise.
The problem compounds as organisations move toward fully autonomous AI agents handling end-to-end service journeys. Each hand-off point, each new data pull, each shift in conversation channel represents an opportunity for context to decay further. What started as a coherent customer story becomes a patchwork of disconnected data fragments the AI is trying to stitch back together in real time. The longer the journey, the more stitching required — and the more likely the seams will show.
For CX operations leaders, this is not a theoretical concern. If your AI deflection rates look good but your CSAT scores are flat or falling, context decay is worth investigating as a root cause. The AI is resolving — but not always correctly, and not always in a way that feels coherent to the customer on the other end.
The Hybrid Response: Humans as Context Anchors
Here is the operational insight that gets overlooked in the rush to automate: skilled human agents are extraordinarily good at maintaining context. They remember what the customer said three minutes ago. They notice the emotional shift mid-call. They carry the thread across a complex multi-issue interaction without losing the plot. These are not soft skills — they are precision cognitive capabilities that current AI architectures genuinely struggle to replicate at scale across long, variable service journeys.
This is precisely where a hybrid human-plus-AI model earns its keep. Rather than positioning AI as the sole handler of complex interactions, smart operations teams are deploying AI to handle the structured, short-context tasks it does reliably well — FAQs, status checks, routing, summarisation — while keeping trained human agents in the loop for interactions where context continuity is critical. The human becomes the context anchor; the AI becomes the force multiplier.
Multilingual complexity adds another layer. When customers are communicating across languages, nuance, idiom, and implied meaning are easily lost in AI processing — accelerating context decay and increasing the risk of a misread situation. Human agents with genuine language fluency and cultural literacy provide a quality backstop that no current model can fully replace.
What Operations Leaders Should Do Now
Start by auditing where your AI-handled journeys are longest and most multi-step. Those are your highest context-decay risk zones. Map the hand-off points, identify where AI confidence scores drop, and consider whether a human review layer — even a lightweight one — should sit at those junctures. Build your automation stack not around the question of "how much can AI handle?" but around the smarter question of "where does context integrity matter most, and who is best placed to protect it?" The answer, for now, is almost always a human.