SiliconANGLE Contact Center AI ROI Analysis Exposes the Knowledge Gap

SiliconANGLE contact center AI ROI analysis has shifted the argument from model capability to a less glamorous constraint: whether an organization can govern its knowledge.

The September 11 report distilled an interview with analysts Bob Laliberte and Zeus Kerravala at the AI ROI in Contact Center Summit. Their central point was direct. Contact centers cannot automate their way to better outcomes when agents receive incomplete, stale, or disconnected information.

That conclusion challenges the market’s dominant sales pitch. Vendors such as Five9, Cisco, Microsoft, Salesforce, and their rivals are adding agents, copilots, orchestration, and automated quality controls. Yet those products still depend on the policies, customer records, workflows, and operating rules beneath them.

The real contest is therefore not one contact center platform against another. It is automation volume against resolution quality. A system can shorten conversations, contain more customers, and still leave more problems unresolved.

That distinction matters because executives are demanding financial returns from AI deployments that have moved beyond controlled pilots. Customers, meanwhile, judge the same systems through effort, accuracy, and whether they need to call again.

What the SiliconANGLE Contact Center AI ROI Report Changed

The report reframed knowledge management as an operating dependency, not a supporting feature.

In the original interview, Laliberte described contact centers as unusually visible tests for enterprise AI. They combine large interaction volumes, substantial labor costs, and direct encounters with customers.

A weak internal experiment can remain hidden inside a department. A weak customer-service agent cannot. It can give the wrong refund rule, miss an account restriction, or force a caller to repeat information.

Those failures expose a basic problem with contact center automation. The language model generates an answer, but the enterprise must supply the facts, permissions, and approved next actions.

Knowledge management covers the processes used to create, update, organize, retrieve, and retire that information. For AI, it also determines which sources the system can use and how those sources are ranked.

The resulting knowledge layer includes much more than a collection of help-center articles. It can include product policies, regional restrictions, account data, troubleshooting procedures, compliance rules, and escalation criteria.

A customer asking about a delayed shipment illustrates the difference. A generic model can explain common delivery problems. A useful agent must retrieve the actual order, identify the carrier rule, and know which remedy the customer qualifies for.

The same requirement applies to human assistance. An AI copilot can suggest language during a call, but its speed has little value when it surfaces an outdated policy.

SiliconANGLE’s report also challenged the metrics traditionally used to judge those interactions. Kerravala questioned whether a shorter call represents success when the underlying issue remains unresolved.

Average handle time measures how long an agent spends on an interaction and related work. It is useful for staffing, but it does not prove that the customer received a correct answer.

First-contact resolution gets closer to the outcome because it asks whether one interaction solved the problem. Even that measure needs careful definitions, especially when customers return through another channel.

Containment creates a similar trap. It records whether an automated system avoided a human transfer, but a customer who abandons a failing bot can still count favorably.

The report therefore did not announce a new model or platform. It identified a change in the standard used to judge the entire category.

Deployment is no longer enough. A contact center must connect automation to accurate resolutions, customer effort, repeat demand, and business results.

This shift also carries an important disclosure. SiliconANGLE stated that theCUBE was a paid media partner for the summit, while sponsors lacked editorial control over its coverage.

Readers should treat the interview as informed industry analysis, not an independent audit of vendor deployments. Its argument remains testable against broader market data and individual production results.

Why Knowledge Management Now Controls AI Returns

Contact center AI produces value only when trusted knowledge can reach the correct workflow at the correct moment.

A language model does not know which version of a return policy is active unless the enterprise provides that context. It cannot infer a customer’s permissions safely from a general training corpus.

Retrieval-augmented generation, often called RAG, lets an AI system retrieve approved information before composing an answer. This design can ground responses in company material, but retrieval alone does not guarantee accuracy.

The source may be obsolete. Two departments may publish conflicting instructions. Access controls may block relevant records or reveal information to the wrong user.

Metadata can also be missing. Without effective labels for region, product, effective date, audience, and owner, a retrieval system can select a plausible but inappropriate document.

That makes content governance part of the AI system itself. Each important source needs ownership, review rules, version history, access controls, and a process for retirement.

The problem becomes harder when an AI agent can take action. An answer about changing an address carries limited risk. An automated address change affects customer data and may require identity verification.

Agentic AI refers to systems that can select steps and perform actions toward a goal. In a contact center, those steps can cross telephony, customer relationship management, billing, and fulfillment systems.

A reliable agent must know more than what to say. It must know which operation is permitted, what information is required, when approval applies, and when to transfer the customer.

Workflow redesign is therefore inseparable from knowledge management. Existing procedures were often written for trained employees who could interpret ambiguity and seek informal guidance.

Automated agents require explicit conditions. Teams must convert undocumented judgment into rules, approved tools, exception paths, and accountable handoffs.

The knowledge base also needs feedback from live interactions. Failed searches, repeated transfers, corrected answers, and disputed summaries can expose missing or misleading material.

That feedback loop turns a static repository into operating infrastructure. Without it, contact center AI repeats the same knowledge failure at greater scale.

The relationship appears in wider 2026 data. A global CX survey covered 815 enterprise decision-makers across 12 markets and 19 industries.

Ryan Strategic Advisory conducted the research for TELUS Digital during the first quarter of 2026. The survey found a gap between AI adoption and the operational systems needed to optimize it.

Fifty-one percent of respondents planned investment in intelligent knowledge management and search. Only 34% reported current deployment, creating a 17-point gap.

Real-time AI copilots showed a similar gap. Fifty-six percent planned investment, while 38% currently used real-time knowledge retrieval and guidance for agents.

Only 32% reported using AI-powered quality assurance and coaching. That leaves many organizations without automated monitoring at the scale created by AI-assisted interactions.

The figures are vendor-sponsored survey findings, not universal market measurements. Still, they support the same mechanism described by Laliberte and Kerravala.

Enterprises have invested heavily in customer-facing AI. The underlying knowledge, evaluation, and improvement layers have developed more slowly.

A governed AI knowledge base offers a useful model for this dependency. Its value comes from provenance and retrieval discipline, not document volume alone.

The practical question for buyers is no longer whether a platform includes generative AI. Most major platforms do.

Buyers need to ask how the system chooses sources, resolves conflicts, enforces permissions, records actions, and measures incorrect answers. Those capabilities decide whether an attractive demonstration survives production.

Automation Volume Is Losing to Resolution Quality

The primary tension is between automating more interactions and producing outcomes customers can trust.

For years, contact centers optimized queues through average handle time, call deflection, occupancy, and cost per interaction. Those measures helped managers plan staffing and control operating expenses.

AI makes several of those numbers easier to improve. It can summarize calls, recommend replies, route requests, or answer repetitive questions without a human agent.

The danger is that the dashboard improves while the customer experience deteriorates. A short automated conversation is not valuable when it produces a second call, a complaint, or an avoidable refund.

An AI agent can also increase containment by making escalation harder. That result reduces apparent labor demand while increasing customer effort and damaging trust.

Laliberte emphasized the visibility of these failures. A poor AI interaction can add work for the customer, harm confidence, and weaken the brand.

Kerravala’s criticism of average handle time points toward a broader measurement reset. Efficiency still matters, but it must sit beneath an outcome hierarchy.

The first layer should measure whether the requested task was completed correctly. The second should capture repeat contact, transfers, corrections, and customer effort.

Financial measures should then include the full cost of resolution. That includes model inference, integrations, monitoring, knowledge maintenance, human review, and escalations.

This approach changes how teams interpret a common scenario. Suppose an AI agent handles an address update without a person.

The interaction appears successful when the conversation ends quickly. It becomes a failure if the system updates the wrong account, misses verification, or forces a later correction.

A credible ROI model must therefore follow the outcome beyond the initial conversation. It should identify downstream rework and customer behavior that conventional containment reports miss.

The financial pressure behind automation remains substantial. A widely repeated Gartner forecast projected $80 billion in contact center labor savings from conversational AI during 2026.

That projection was originally made in 2022, when Gartner estimated that one in ten agent interactions would become automated by 2026. It was a forecast, not a verified accounting of savings already achieved.

Gartner’s more recent cost warning adds important tension. The firm predicts generative AI cost per resolution will exceed three dollars by 2030.

Its reasoning includes rising data center expenses, more complex use cases, vendor pressure for profitability, and the specialized labor required to maintain deployments.

Gartner also expects regulation to increase assisted-service volume. Its forecast says AI-related rules will raise such volume by 30% by 2028 as customers exercise access to human agents.

Neither forecast invalidates contact center automation. Together, they show why labor substitution cannot serve as the only business case.

AI can create value through faster agent preparation, more consistent answers, better quality coverage, and improved routing. Those benefits still require defensible measurements.

Outcome scoring should also separate simple and complex work. An order-status request does not carry the same risk as a disputed medical bill or suspected financial fraud.

Combining both interactions into one containment rate can conceal where automation works. It can also hide where human judgment remains necessary.

Contact centers should segment results by intent, channel, customer group, language, and risk level. They should also compare automated outcomes with appropriate human baselines.

This is where the knowledge foundation becomes financially visible. Accurate sources raise the chance of resolution, while clear escalation rules prevent expensive failures.

The winning operation will not necessarily automate the largest share of calls. It will automate the right calls while preserving trustworthy access to people.

Voice AI Raises the Cost of Bad Knowledge

Voice agents turn knowledge errors into live customer experiences, often before a supervisor can intervene.

Voice remains one of the hardest channels because the system must recognize speech, interpret intent, retrieve information, respond quickly, and manage interruptions.

Text interfaces allow users to reread an answer. A caller experiences delay, confusion, and repetition in real time.

Background noise, accents, unusual names, and poor connections can distort transcription. A correct knowledge source cannot help when the system misunderstands the customer’s request.

Five9 highlighted this dependency in its March 2026 deployment guidance. The company described speech-to-text accuracy as the foundation for routing, qualification, and knowledge retrieval.

That claim comes from a vendor selling the relevant technology. It still identifies a concrete failure point that buyers can test with their own audio and call mix.

In June, Five9 introduced an updated Voice AI release. The system uses coordinated agents for multi-step customer interactions and includes testing, versioning, monitoring, and rollback controls.

Five9 said 65% of organizations in its research were implementing and releasing at least one AI use case. It also reported self-service automation as a leading use case at 42%.

Those findings should be treated as company-sponsored research. They show how the vendor market is positioning production controls as part of the product, however.

Cisco, Salesforce, Microsoft, NICE, Genesys, and other platform providers are following the same broad direction. They are connecting conversational systems to enterprise records, routing, workforce tools, and quality management.

The competition will expose an important divide. A platform can supply connectors and governance features, but the customer remains responsible for its records and operating decisions.

No vendor can automatically decide which conflicting policy reflects an organization’s true intent. It cannot assign ownership where the enterprise has left responsibility unclear.

The handoff to a human agent creates another knowledge test. A customer should not need to repeat the entire conversation after automation reaches its limit.

The human needs the transcript, verified customer context, actions already attempted, and the reason for escalation. Raw conversation history alone can create more work rather than less.

An AI-generated summary can help, but it must distinguish customer statements from verified facts. Otherwise, the next agent may treat an incorrect assumption as confirmed account information.

Regulated environments add further requirements. Financial, healthcare, and public-sector contact centers must control access, preserve audit trails, and document consequential actions.

These controls can reduce apparent speed. They also protect the organization from turning one inaccurate answer into a repeated policy violation.

The same logic applies to multilingual service. Translating a flawed source does not fix it. It can multiply the error across markets and create conflicting customer commitments.

Testing must therefore cover more than conversational fluency. Teams need scenario suites built around common requests, edge cases, policy exceptions, adversarial prompts, and handoffs.

They should test the complete chain from speech recognition through retrieval and action. A model-only benchmark cannot reveal failures inside account systems or workflow permissions.

Production monitoring should also distinguish refusal from failure. An agent that safely escalates a restricted request may perform better than one that completes it incorrectly.

This complicates simple automation leaderboards. Higher containment does not always mean better operational performance.

Knowledge freshness provides another measurable signal. Teams can track how many active sources have owners, review dates, effective dates, and documented replacement rules.

They can also measure retrieval success. Useful indicators include unanswered searches, conflicting sources, supervisor corrections, and repeat inquiries after automated resolution.

These operational metrics connect content maintenance to financial outcomes. They reveal whether a model problem is actually a knowledge problem, an integration problem, or a process problem.

The ROI Case Still Has Serious Verification Gaps

The knowledge-management thesis is persuasive, but the current evidence does not establish one universal formula for AI returns.

SiliconANGLE’s article centers on analyst commentary from a sponsored event. It does not provide audited results from a defined group of contact centers.

The TELUS Digital research offers broader quantitative evidence, but the sponsor also sells customer-experience services. Its figures describe reported adoption and investment priorities, not causal proof.

Vendor case studies have another limitation. They often highlight successful deployments, selected use cases, and customers prepared to speak publicly.

Those cases can reveal implementation mechanics. They do not establish that an average buyer will receive the same results.

Organizations also define resolution, containment, and customer satisfaction differently. A metric that looks identical across two reports may use different observation windows or exclusions.

The underlying work mix changes comparisons further. A retailer handling order status has different automation potential from a hospital managing sensitive care questions.

Labor savings are especially easy to overstate. Reduced handle time creates capacity, but capacity does not become cash automatically.

A contact center realizes direct savings only when it avoids hiring, reduces overtime, shifts work, or changes staffing. Otherwise, it may gain service capacity without lowering expenditure.

That capacity can still be valuable. Faster responses, more coaching, or additional customer outreach may improve the business without appearing as a payroll reduction.

AI costs also extend beyond software licenses. Enterprises need integration, security reviews, evaluation data, knowledge owners, compliance work, and continuing model supervision.

Some of those costs move rather than disappear. Fewer repetitive calls may create more complex work for human agents and greater demands on supervisors.

Customer behavior can change as well. Easy automated access may increase the number of interactions, offsetting savings from a lower cost per contact.

The industry also lacks a stable benchmark for long-running autonomous systems. Models, vendor architectures, policies, and customer expectations are changing at the same time.

That makes short pilots poor predictors of mature operations. A pilot may use a narrow intent set, curated knowledge, and unusually close supervision.

Production introduces policy changes, seasonal demand, new products, uncommon accents, system outages, and customers who behave unpredictably.

The most credible evaluation design starts with a limited set of high-volume intents. Teams should establish human baselines and record the complete cost of current resolution.

They can then compare AI-assisted, autonomous, and human-only groups using the same definitions. The comparison should extend long enough to capture repeat contacts and correction work.

Quality review needs representative sampling across customer groups. An average score can hide weaker performance for certain languages, accents, products, or accessibility needs.

The organization should also maintain a clear stop condition. A rising correction rate, compliance incident, or repeated retrieval failure should trigger rollback or narrower autonomy.

This is not an argument for delaying every deployment. It is an argument for matching authority to evidence.

Summarization and agent assistance generally leave a person in control. Autonomous refunds, account changes, or eligibility decisions create more consequential exposure.

A staged model lets trust grow with measured performance. It also gives the knowledge team time to fix the content gaps that real interactions reveal.

The key uncertainty is therefore not whether knowledge matters. It is how much disciplined knowledge work improves each outcome within a particular operation.

Executives should ask for that local evidence before accepting broad claims about savings. They should also resist treating missing measurement as proof that benefits do not exist.

A balanced scorecard can capture both financial and service value. It should connect resolution quality, customer effort, employee outcomes, risk, and fully loaded operating cost.

Three Signals Will Show Whether Contact Center AI Pays

The next phase will be decided by production evidence, knowledge discipline, and customer outcomes rather than feature announcements.

The first signal is a change in reported metrics. Vendors and enterprise buyers should publish verified resolution rates alongside containment and average handle time.

Those reports should include repeat contact, escalation quality, correction work, and customer effort. They should also explain definitions and measurement windows.

If outcome reporting becomes standard, it will strengthen the SiliconANGLE contact center AI ROI thesis. It will show that buyers no longer accept automation volume as a substitute for value.

If vendors continue emphasizing conversations handled without publishing downstream results, the verification gap will remain. Buyers will struggle to compare platforms or validate returns.

The second signal is investment in knowledge operations. The 17-point deployment gap found in the 2026 survey provides a clear baseline for future research.

Watch whether organizations move beyond buying search features. The stronger indicator is whether they assign content owners, track freshness, resolve conflicts, and connect corrections to workflows.

Growth in those practices would support the argument that enterprise knowledge has become production infrastructure. Flat adoption would suggest that companies still expect models to compensate for weak foundations.

The third signal is how platforms handle human transfer and governance. Product releases should be judged by observable controls, not by claims of greater autonomy.

Useful evidence includes versioning, rollback, source citations, permission enforcement, action logs, evaluation tools, and complete handoff context.

Regulatory developments will also test these controls. Requirements for human access can expose deployments built around keeping customers inside automation at any cost.

A well-designed operation should treat transfer as a valid resolution path. It should reserve autonomy for tasks where the system has sufficient knowledge, authority, and evidence.

These signals matter to more than contact center executives. Developers must design retrieval and tool access around changing business rules rather than static demonstrations.

Enterprise buyers need procurement criteria that cover knowledge ownership and evaluation. A feature list cannot reveal whether a product will work with fragmented internal information.

Knowledge workers also have a direct stake. AI can remove repetitive searching and post-call documentation, but it can push harder cases toward employees without adequate context.

Customers face the clearest consequences. They need systems that resolve ordinary requests quickly and provide accountable human help when circumstances become complex.

The larger lesson reaches beyond customer service. Enterprise AI depends on the quality of the information and decisions that organizations can make available to it.

Contact centers simply make that dependency easier to observe. Every wrong answer, failed transfer, and repeat call becomes measurable evidence.

SiliconANGLE contact center AI ROI analysis captures the market at that transition point. The model is no longer the whole product, and deployment is no longer the whole result.

The practical next step is to inspect one high-volume customer journey from beginning to end. Identify every source, decision, system action, exception, and handoff involved.

Then ask whether each source has an owner and whether every automated action can be reviewed. Measure correct resolution before expanding the system’s authority.

If those answers remain unclear, adding another AI agent will not repair the business case. Better-governed knowledge, clearer workflows, and honest outcome metrics must come first.