The rise of generative artificial intelligence is profoundly transforming customer relations roles. Promising efficiency, availability, and personalisation, it is now establishing itself as a major strategic lever for CX and service departments. Yet behind the enthusiasm, the operational reality remains mixed: many projects struggle to move beyond the pilot stage, while the promised gains often remain difficult to measure. Why do generative AI projects fail in customer relations?
The question is less technological than organisational. Recent research — from RAND Corporation to MIT Project NANDA — converges on a finding: failure rarely lies with the model itself, but with the way it is integrated, governed, and aligned with the real needs of users. Poorly defined problems, insufficient data, technology fascination, infrastructure debt, or the absence of guardrails: the causes are known, but still underestimated.
In a context where companies want to rapidly industrialise their AI agents, understanding these root causes of failure and identifying the success factors becomes essential. This document offers a cross-reading of recent studies (RAND, MIT, Deloitte, NIST, Salesforce, Zendesk) to identify the concrete levers for moving from a showcase chatbot to a genuinely productive, compliant, and value-creating AI Agent.
Why Do So Many Generative AI Projects Fail in Customer Relations (Chatbots & AI Agents) and How Can They Succeed?
The 2024 RAND report identifies 5 root causes of AI project failure: poorly defined problem, lack of useful data, technology fascination, infrastructure debt, and technical limitations. These pitfalls are even more visible in customer relations chatbots.
Only 13-14% of organisations are “AI-ready” despite a near-universal urgency to deploy AI; most remain stuck at the pilot stage. Furthermore, studies on customer experience remind us that bots can degrade satisfaction if they appear focused on cost reduction. Empathy and conversational design, however, change the picture. “All-AI” predictions (e.g. >50% of cases resolved by AI by 2027) coexist with research (MIT 2025) asserting that 95% of GenAI pilots have no P&L impact, which underlines the importance of a measurable, focused, and integrated roadmap.
Furthermore, a 90-day plan (targeted use cases, RAG, guardrails, HITL, CX metrics such as containment & CSAT) makes it possible to move from the “showcase” chatbot to a genuinely productive AI Agent that is compliant with the AI Act, according to NIST.
Context and Stakes
What RAND Research Says and Why It Is Critical in Customer Relations
The RAND report interviewed 65 practitioners (data scientists and ML engineers) and synthesises 5 root causes of failure.
First, failure would be caused by a poorly defined problem or poor metrics. Second, it would be linked to insufficient or unusable data. Third, one of the causes would be the shiny object syndrome (chasing technology rather than the problem). Fourth, one would be infrastructure debt related to data governance, deployment, or MLOps. Finally, the intrinsic limitations of AI on certain problems could also be a cause of failure.
In this case, the authors recommend aligning goals and business context and choosing “durable” problems. As a result, they also emphasise the importance of investing in infrastructure and understanding its technical limitations.
Ultimately, once transposed to chatbots, these problems have significant consequences.
- Poorly defined problem: a bot optimised for deflection can degrade CSAT if the customer’s actual intent is complex.
- Poor data: knowledge bases become outdated and produce incorrect or “hallucinated” responses.
- Technology fascination: launching a “turbo-charged” LLM without an escalation procedure or empathy design can significantly degrade customer relations.
- Infrastructure debt: consequently produces unreliable RAG, missing observability tooling, and few guardrails.
- AI limitations: certain patterns (emotional complaints, multi-entity cases) require a human intervention known as human-in-the-loop by design, according to RAND Corporation. Without this, the user remains stuck with their problem.
In conclusion, RAND reminds us that “according to some estimates, more than 80% of AI projects fail” — this is not their own measure, but a risk signal to be taken seriously.
State of Play: Much Urgency, Little Preparation
Given these specific findings, two structural observations stand out:
- Cisco AI Readiness Index: 97-98% of organisations say AI urgency has increased, but only 13-14% declare themselves fully ready to make the leap.
- McKinsey (2024): GenAI adoption is progressing (65% of respondents say they use it), but value creation remains concentrated among the best-prepared players.
Also, on customer relations, projections are ambitious: Salesforce estimates that 30% of cases were already resolved by AI in 2025, and 50% by 2027. These figures describe a trend (not an established fact), useful as a compass but not as a guarantee.
Customer Experience: What Studies Tell Us (and What We Too Often Forget)
First, it is important to address perception bias. Academic research shows that customers often perceive bots as a cost-cutting lever “against” quality, which reduces service evaluation at equal performance. Ultimately, it is therefore important to frame the intent and explain the role of the bot upfront.
Next, empathy and conversational design must not be overlooked. Integrating empathy signals improves satisfaction; it is a product skill as much as a model choice.
Finally, the contrarian signal also plays a role. The Zendesk 2025 and Salesforce reports describe a wave of investments and positive perceived returns from decision-makers. Indeed, 73% of CX leaders believe that companies that scale AI will survive competitive pressure, according to the studies conducted. However, these figures should be treated with caution as they are perceptions from surveys.
In sum, while conversational AI represents a clear strategic lever, its success depends less on technical performance alone than on the perception and trust it inspires. The challenge is therefore not simply to automate, but to design experiences aligned with users’ emotional and cognitive expectations. It is on this condition that bots will become genuine extensions of customer relations, rather than mere efficiency tools.
The MIT 2025 “Bombshell” vs Optimistic Studies: Who to Believe?
The MIT Project NANDA (2025) report sounds the alarm: according to its conclusions, 95% of GenAI pilot projects have no measurable impact on P&L. The reason: a lack of deep integration into business processes and the absence of genuine learning loops from the field. Only 5% of initiatives manage to create value, by tackling clearly defined problems and relying on appropriate technology partners. A salutary warning against the risk of the “eternal pilot”.
Conversely, the Deloitte 2025 report paints a more optimistic picture: among GenAI initiatives already scaled, almost all show tangible ROI, with nearly 20% exceeding 30% return on investment. In other words, it is not the AI itself that fails, but organisations’ ability to industrialise its use cases.
In conclusion, this divergence is explained primarily by the maturity level of the projects (pilot vs. large-scale deployment). It also lies in the relevance of the use cases selected, the degree of integration into the information system, and the measurement of performance. In other words, value does not come from the model, but from execution.
Compliance & Risks: AI Act, NIST, and Response Reliability
The European AI Act marks a decisive step. From 2 August 2025, obligations relating to general-purpose AI models (GPAI) and governance will progressively come into force. Systems already on the market will need to be fully compliant by 2 August 2027. Certain provisions, notably targeted prohibitions and measures relating to AI literacy, will apply from February 2025. In practical terms, this requires sector players to strengthen traceability, documentation, risk assessment, and transparency throughout the model lifecycle.
In parallel, the NIST AI Risk Management Framework (AI RMF) offers an operational and pragmatic approach to risk management, structured around four key functions: Govern, Map, Measure, and Manage. This structure provides a robust framework for establishing guardrails, planning tests, ensuring monitoring, and effectively handling model-related incidents.
Finally, on the scientific front, research in 2024-2025 (notably from Oxford University) has made considerable progress in detecting and reducing hallucinations via abstention strategies or retrieval-augmented generation (RAG). However, zero hallucination remains a theoretical ideal: the real challenge is now to manage uncertainty, by equipping systems to detect, flag, and escalate ambiguous cases with discernement.
Actions and Best Practices to Put Into Practice
The Metrics That Matter for a Customer Relations Chatbot/AI Agent
A few useful definitions:
- Containment (CCR): % of interactions fully handled by the bot without human intervention. Standard formula: conversations resolved by the bot / total conversations initiated.
- Resolution rate, CSAT, CES, AHT, escalation rate, cost per interaction (compared to agent cost).
Note that containment alone is not enough. If poorly optimised, it degrades satisfaction and increases repeat contacts; combine it with post-interaction CSAT and First-Contact Resolution.
From “Chatbot” to AI Agent (Agentic AI): What Changes and What Stays the Same
AI Agents plan, use tools (CRM, payment, logistics), maintain a state of the conversation and operate with rules — ultimately a step beyond pure Q&A. Consulting firms (McKinsey, Capgemini) see agentic AI as a end-to-end automation lever through complete resolution rather than mere information retrieval. However, the hygiene building blocks remain essential: RAG, policy engine, HITL, monitoring, observability.
Some customer relations use cases well-suited to agentic AI:
- address change & identity verification & CRM update,
- refund & credit note (rules, ceilings, dual control),
- parcel rerouting & notifications,
- appointment scheduling with multiple constraints (calendar & stock APIs).
Note: certain cases should not be “agentified” first: sensitive complaints, disputes with high emotional weight. In these situations, the human still plays a non-negligible reassurance role in customer relations.
90-Day Execution Plan tolk.ai (Customer Relations Special)
Scoping & measurement baseline (weeks 0-2)
- Choose 1-2 use cases with high volume and low risk (e.g. order tracking, delivery status).
- Baselines: CSAT, AHT, escalation rate, cost per contact, self-service rate. (RAND Corporation)
Data & architecture (weeks 2-6)
- Build a clean RAG (index FAQ/KB, policies, CRM extracts), version the sources, define freshness SLAs.
- Implement guardrails (lists of authorised actions, PII-masking, controlled refusals) + explicit HITL.
Conversational design & empathy (weeks 4-8)
- Flows with intents + empathy (acknowledgement, rephrasing, options).
- User testing (sensitive scripts) and offline evaluation on real corpus. (ScienceDirect)
Controlled pilot (weeks 6-10)
- Canary launch on 10-20% of traffic, A/B vs. reference channel.
- Daily monitoring: containment, post-chat CSAT, escalation causes, hallucination flags.
