Next-gen AI: Driving real ROI with zero-copy data

We’re in an era of instant gratification, and its impact can be felt almost everywhere, especially in business. Chief information officers (CIOs) and enterprise architects are all too familiar with it. They now face unprecedented pressure to deploy artificial intelligence that moves past the hype and noise and delivers instant, measurable ROI.

And it’s not just business leaders that recognise this; findings from our latest State of Sales report indicate a clear increase in the number of customers demanding more measurable ROI and service outcomes than in previous years. As a result, many businesses can’t afford the time or luxury to experiment with AI concepts and tools anymore. The risk of being left behind is too high.

The highest immediate returns are often found by automating case resolutions and deploying high-accuracy forecasting. However, a critical reality remains. Artificial intelligence, in its current guise, is only as powerful as the data it can access and the guardrails that govern it. That often makes true integration stilted and patchy.

But there is a solution, and we’ll cover it in this article. We’ll explore the proven AI strategies and tools that modern enterprises are successfully incorporating into their operations, and we’ll look at the key regulation and compliance concerns that can hinder many AI ambitions. We’ll also provide some core metrics and KPIs that IT leaders can use to evaluate the success of their implementation.

• An AI engine is only as powerful as the context it can access; transitioning from legacy ETL to a unified, zero-copy architecture establishes a real-time ‘source of truth’ across ERP and CRM systems without the risk of data duplication
• Successful deployment requires a visible trust architecture for both internal and external parties
• Integrating your customer service data with sales pipeline analytics creates a crucial commercial bridge and subsequent increased revenue
• A single, unified platform removes the shortcomings and risks of non-standardised stacks and workflows

For next-generation AI to execute complex tasks to its maximum capability, it needs a comprehensive context to work with. This means that enterprise leaders must shift from siloed data architectures to a unified data engine that can harmonise both structured data (like ERP records and CRM fields) and unstructured data (like customer emails and call transcripts).

Historically, feeding this level of operational context to enterprise applications required the traditional extract, transform, load (ETL) process, where raw data is transformed and cleansed to fit a standardised format. While this provided a foundation for data automation, the duplication of massive data sets and the subsequent synchronisation lag created problems.

Market leaders have started to champion what is known as ‘zero-copy’ architecture in response. Rather than physically moving or copying enterprise data located within data warehouses, zero-copy technology allows a business’s AI engine to access and query the data at the source. This innovative approach ensures that data never leaves its secure home, so it eliminates the security risks and infrastructure costs associated with traditional data pipelines.

This advanced level of data hygiene is vital for a successful transition to more unified platforms. Our research suggests that high-performing teams are 1.5 times more likely to prioritise data hygiene to improve AI outcomes and establish an immutable single source of truth (SoT).

For an AI agent to resolve customer queries instantly, it can’t wait for yesterday’s ledger to sync. It needs live, fluid access to systems that run your business, like your legacy enterprise resource planning (ERP) platform or your contact centre as a service (CCaaS) telephony stacks.

This fundamental, modern AI function is often more attainable than many businesses realise. Rather than initiating a blunt rip-and-replace policy, leading enterprises are choosing to wrap their legacy infrastructures in modern API-led connectivity and event-driven architectures. By combining tools such as MuleSoft for orchestration with zero-copy data virtualisation, enterprises are allowing AI to interact securely with legacy systems in real time.

At the same time, unifying CCaaS telephony systems with a CRM platform and digital messaging nodes like WhatsApp creates an omnichannel presence that’s often asynchronous. So whether the customer initiates an inbound call, sends an SMS, or interacts with a chatbot, the AI agent can readily ingest a customer’s history with the company and respond accordingly.

That eliminates conversational fragmentation, the need for customer repetition, and latency. But access to personalised customer profiles is essential for success.

According to our State of the AI Connected Customer report , 73% of customers believe that brands treat them as individuals, a figure that substantially increases year on year. However, only 49% feel that brands use their information in a positive way, suggesting that there’s still a long way to go to deliver the personalised experience customers really want.

When evaluating the infrastructure required to connect legacy systems to modern AI engines, enterprise architects should look to prioritise these four main features.

Market leaders are successfully able to layer autonomous agents over a unified data foundation to reduce case resolution times. Because most next-gen AI agents and platforms are directly integrated with real-time CRM data and legacy ERP systems, they can quickly and independently resolve Tier 1 queries. This can include processing complex order exchanges, cross-referencing inventory, or updating tier accounts, all done in seconds.

However, even with the most sophisticated AI technology out there, agents are still likely to encounter edge cases, such as a highly nuanced problem or a customer who becomes upset during the exchange. Such is the nature of conflict resolution.

So while a quick resolution is clearly a major benefit, architects also need to implement a deterministic handoff architecture to handle these cases. It might look something like this:

1. Trigger detection: The AI recognises an escalation trigger, such as a negative sentiment threshold or a query that falls out of its grounded knowledge base remit
2. Context aggregation: The system compiles the agent’s full conversation transcript, the underlying retrieval paths, and the live data fetched from the ERP/CRM into a unified case briefing
3. Disclosure of transfer: The agent explicitly notifies the customer that it’s transferring them to a specialist
4. Human handoff: The case is routed to the human agent’s console via an omnichannel gateway, and the human representative will already have the full summarised context, eliminating the need for the customer to repeat their problem

Keeping human service reps involved in case escalation and resolution is vital. For our State of the AI Connected Customer report, 72% of customers said that it’s important for them to know when they’re communicating with an AI agent, and 46% said they’re more likely to use an AI agent when they know there’s an escalation path to a real person. IT leaders must commit to working this into any CCaaS platforms that they opt to implement.

Utilising AI agents for case resolution and escalation also can’t come at the expense of transparency or brand protection. Even though AI is so prevalent across all areas of business, customers are still somewhat sceptical about its use, with 31% of customers saying they would trust AI more if they could get a clear explanation of outputs from the AI platform.

To meet these customer expectations while still protecting the enterprise, IT leaders must implement a multi-layered trust architecture. This should include features such as:

• Grounding and retrieval-augmented generation (RAG): Instead of allowing the underlying large language model (LLM) to generate responses freely from its training data, RAG restricts the AI’s knowledge base and access to call upon verified enterprise documents, active knowledge articles, or live ERP/CRM records only. This helps to eliminate hallucinations and problematic responses.
• PII masking and tokenisation: Before a customer’s prompt is even transmitted to the LLM, a strict privacy layer automatically strips out any PII data. This guarantees that sensitive data never enters the model’s environment.
• Toxicity mitigation: Real-time semantic filters must sit between the user and the agent, checking inbound (and outbound) messages. This is to identify, block, or reroute potentially inappropriate language or hostile inputs.
• Role-based access control (RBAC): An agent should only be able to know of and access what it needs to achieve its objective, and nothing more.

By anchoring autonomous agents with strict frameworks, explainable logic, and frictionless human escalation, organisations can confidently scale their operations while retaining a high level of customer trust and sentiment.

Modern AI agents and platforms are also proving their worth when it comes to strategic planning. Findings from our State of Sales Report indicate that 91% of sales professionals believe AI directly benefits their sales planning and predictive modelling, highlighting the enormous potential this area of AI usage has.

Previously, enterprise sales teams relied on historical trends and the gut feelings of individual sales representatives. Advanced machine learning algorithms are starting to strip away this well-intended ambition, replacing human bias with automated, real-time predictive insights.

To sell effectively, agents and mobile workers need to deliver personalised and trusted customer service. That’s possible when data is integrated across all support channels, as well as sales, ecommerce, and marketing. This provides a comprehensive, real-time view of the customer that AI can leverage in an ethically responsible manner.

While AI is often embraced for time-saving and cost reduction, it’s also helping to shift the perception of service organisations from cost centres to revenue generators. AI can suggest contextually aware ‘next best offers’ to agents based on detailed customer histories and recent website engagement. And with AI automating much of the administrative busywork, agents have more time to upsell and cross-sell.

Fisher and Paykel are a great example of a company proving that service and sales are stronger together. By unifying its sales, service, and marketing teams around a single view of the customer, they’re now able to provide self-service buying options across every channel. This creates a better user experience for customers, who can easily locate what they want and need.

This contextual intelligence highlights a key missing link in modern commercial design: the integration of customer service data with sales pipeline data. When service endpoints are deeply connected via unified data architecture, every customer support interaction flows directly into downstream predictive forecasting models.

As a result, the business gains an accurate, data-driven look at pipeline health, bridging the gap between customer retention and growth.

When reporting project success back to the CIO or CFO, IT leaders need concrete, data-backed evidence that illustrates how next-gen architecture directly impacts both financial efficiency and infrastructure durability. General sentiment metrics like customer satisfaction offer a broad insight, but they’re not enough to justify investment in holistic, powerful AI platforms and tools.

To present a comprehensive picture of performance, teams should categorise their metrics and KPIs into two distinct groups:

• Operational AI metrics: These are data points that track how well the machine learning model performs in relation to daily business operations and functions, and they’re used to evaluate the model’s direct business impact
• Architectural health metrics: These are data points that monitor the mechanics underlying the machine learning model, and they’re used to evaluate a system’s stability, data security, and integration performance

Below you’ll find some of the core metrics and KPIs across both groups that you should pay particular attention to when monitoring for AI success.

It’s also smart to keep track of other metrics, such as your Retrieval-Augmented Generation (RAG) precision score (the percentage of agent answers that can be mapped back to verified knowledge sources), your end-to-end data ingestion latency (in relation to your data pipeline), and the average cost per API call or resolution.

A business may spend countless hours and a lot of money to build the best possible modern AI infrastructure available. But the other essential component needed for AI adoption is the workforce itself.

We all prefer familiarity. Once employees are fully accustomed to a system or tool, it can be hard to let it go and embrace something new (particularly when it involves AI, which still creates job security concerns). But real ROI is often achieved when technical capability and human adoption find a sweet spot.

Driving adoption often starts with building an intuitive, low-friction workspace for your team. If a sales representative or customer service agent has to cycle through multiple tabs or is confronted by unwieldy data visualisation, they’ll simply go back to their legacy routines.

Successful enterprises will often position AI as an assistant rather than anything else. You don’t want your workforce to see it as a threat. Instead, they need to understand how the AI platform will make their jobs easier.

A co-pilot user interface is often a great way of steadily introducing the technology. Architects can design the initial interface as a side-by-side companion panel where the AI can state recommendations, draft emails, or cross-reference ERP data directly inside the employee’s main CRM screen. At this early stage, all AI tasks will still require a human click to approve or send.

There should also be a clear commitment to maintaining transparency and explainability, particularly during early implementation. The interface should present clear grounding sources to all employees, allowing them to see exactly where certain AI decisions are being drawn from. This will help to eliminate ‘black box’ anxiety that can cause employees to reject AI suggestions.

Transitioning from an abstract architectural blueprint to a live, high-ROI enterprise deployment requires a shift away from fragmented point solutions. Stitching together the various platforms, tools, and systems can create a disorganised integration that can cause instability throughout a company’s operations.

Instead, market leaders are choosing to leverage a single, unified metadata architecture, such as the Salesforce Platform, to deploy these capabilities natively, securely, and at scale.

1. The Data and Integration Foundation

• MuleSoft serves as the modern API and event-driven gateway that wraps legacy ERP architecture and CCaaS telephony infrastructures.
• Data 360 functions as the real-time, zero-copy data engine. Data 360 harmonises structured CRM and ERP records with unstructured transcripts and email chains into a single source of truth.

2. The agent and guardrail layer

• Agentforce is the core, autonomous agent platform used to build and deploy role-specific digital agents. Architects can easily tailor pre-built agents to execute advanced workflows using trusted enterprise data.
• AI Trust Layer is the native security architecture providing effective compliance boundaries. It automatically executes real-time PII masking, tokenisation, and grounding checks before any data is processed by the LLM.

3. The predictive forecasting engine

• Tableau Pulseis a dedicated machine learning engine that eliminates human bias from the sales pipeline by analysing historical trends, consumption metrics, and customer behaviours to deliver precise predictive models.
• Tableau is the downstream data visualisation engine that translates complex predictive data into clear, explainable insights, ensuring business leaders understand the exact logic driving pipeline adjustments.

Because these layers are natively built onto a shared metadata framework, your security permissions, data classifications, and role-based access controls automatically shift from your data layer up to your autonomous agents. This unified foundation eliminates the risk of data leaks and reduces overall deployment timelines.