80% Accuracy Works for Consumer Use Cases, But Is a Complete Disaster for Enterprise Financial Closing
After years of enterprise AI pilot programs, most projects have never moved past the trial phase. Two senior executives from SAP and Infosys joined a MIT Technology Review podcast to break down the core bottlenecks: fragile data foundations, missing governance frameworks, and the paradigm shift of treating AI as a digital coworker rather than a simple automation tool.
There is a stark statistic for enterprise AI: 86% of companies remain stuck in pilot, proof-of-concept, or fragmented trial stages, with only 14% having successfully scaled AI to production across their operations. This is the finding from a joint Infosys and HFS survey of 500 Global 2000 enterprises.
Data from BCG paints a similar picture: 74% of organizations are trapped in pilot mode with no formal scaling roadmap. A separate Anthropic survey shows 42% of businesses are blocked by poor data access and low data quality.
Taken together, these figures point to the same core truth: Enterprises are not short of AI tools, but lack the infrastructure and operating models required to make AI run reliably inside real business workflows.
Thorsten Leiduck, Global Head of Ecosystem for SAP Business AI, and Satinder Pal Singh, Vice President at Infosys North America, laid out this problem clearly in the MIT Technology Review Business Lab podcast. Co-produced with Infosys, the conversation centers on exactly what enterprises need to reconfigure as they move from isolated AI pilots to agent-driven, automated workflows.
Fix Your Data Layer First, Before Deploying Agents
Leiduck’s assessment is straightforward: AI performance is entirely dependent on robust data foundations. Enterprises need a connected, semantically rich data layer that accurately reflects their unique business reality — otherwise agents can only operate based on generic large model patterns and will fail to deliver tangible business value.
He uses model accuracy as an illustrative example: In consumer-facing scenarios or daily office productivity tools, 80% accuracy may be perfectly acceptable, since human users can step in to fix errors. But in mission-critical processes such as supply chain management or financial closing, 80% accuracy is a catastrophic outcome.
This distinction is critical. Many enterprises directly port the consumer AI experience into their core business systems, ignoring the strict requirements enterprise workflows have for accuracy, auditability, and embedded business logic. SAP’s approach is to bake business semantics and industry context into knowledge graphs, so agents operate within the real-world constraints of the company, rather than relying solely on generic LLM reasoning.
The AI Department Trap
Singh attributes the 86% pilot failure rate to a common organizational pattern: the AI Department Trap. Companies create a dedicated AI team or center of excellence, and isolate it from core business units. The team builds technically impressive pilots, but the final output is unrecognizable to frontline business teams. The end result is that AI solutions go looking for problems, instead of actual business problems driving AI adoption.
His recommendation is to reframe AI not as a technical initiative, but as a full operating model transformation. AI is no longer something the IT team builds for business teams, but something every function and every workflow is designed around from the start.
When evaluating use cases, Infosys filters for three criteria simultaneously: impact, feasibility, and scalability. Impact measures whether an agent can move the needle on key business metrics such as revenue, cost, and working capital risk. Feasibility assesses whether current data infrastructure, platform capabilities and talent can support the use case. Scalability asks whether a successfully validated use case can be replicated across the organization via its architecture. As Singh puts it: A pilot that cannot be scaled is just an experiment.
He also shares a very practical analogy: manage your AI project portfolio the same way you manage an investment portfolio. Fast, quick-win projects fund the overall program, mid-term bets maintain consistent delivery momentum, and one or two bold long-term positions keep stakeholders invested for outcomes that will materialize after three years.
Two Real-World Production Deployments
The first use case comes from a large global diversified manufacturer with an extensive SAP footprint. Its core pain point was month-end financial close: more than 80 separate SAP transaction codes, the closing team regularly working overtime including weekends, and over 300 L1 support tickets generated every single month as part of the close cycle.
Infosys deployed Topaz Fabric on the SAP Business AI platform, using a master agent to orchestrate five specialized sub-agents that cover WBS accruals, revenue recognition, intercompany reconciliation, master data matching, and SOX audit trail generation. The agents run autonomously, only escalating true edge cases to human staff. The end result: the entire L1.5 support tier was eliminated, zero human intervention required to resolve the 300+ monthly tickets. The financial closing team shifted from reactive fire-fighting to proactive financial analysis, generating more than $1 million in annual savings.
The second use case involves a North American utility company going through a S/4HANA transformation. Its bottleneck was development speed: each complex SAP functional specification document required more than 80 person-hours of work. Using the AI-native development environment on the business AI platform paired with a Claude-powered retrieval system, Infosys cut the time required for specification documents from 86 person-hours to just 6, a 92% reduction, with 95% accuracy validated across 51 distinct SAP task types. This architecture is now being replicated for other clients.
Singh also referenced a global S/4HANA transformation project that used AI-assisted delivery tools to complete more than 480 RICEF objects in 5 months, a timeline impossible with standard delivery methodologies. Another customer uses the AI platform to automatically classify more than 120,000 SAP support events per month, with zero analyst billable hours required. The combined effect cut delivery timelines from 12-18 months down to 4-6 months.
Governance Is Not a Brake — It Prevents Forced Full Outages Down the Line
Governance is often framed as the opposite of rapid scaling. But Singh’s data tells a different story: the Infosys Knowledge Institute Generative AI Radar found that 48% of companies that have deployed generative AI have no formal governance policies in place. Nearly half of all enterprises are running AI systems without documented frameworks, oversight, accountability mechanisms, or compliance checks.
His assessment is that missing governance does not accelerate AI adoption — it creates a backlog of accumulated risk that will eventually force a far more disruptive full system shutdown and mandatory audit.
SAP’s Agent Hub is designed as the control center for enterprise-grade AI and agent governance. It can manage both SAP-native agents and non-SAP agents integrated via A2A protocols. Its built-in capabilities include an agent registry, LLM registry, MCP registry, verifiable agent identities and scoped access permissions, full observability, and model drift detection. Leiduck emphasizes all these capabilities are natively embedded in the business AI platform, not bolted on as an afterthought.
Singh breaks governance down into three layers: The policy layer: organization-wide AI usage policies that clearly define what agents can and cannot do, how data can be used for training, disclosure requirements, and accountability for outputs, with buy-in from legal, compliance, technology teams and executive board sponsorship. The platform layer: using Agent Hub as the enterprise control tower to give leadership full real-time visibility even in complex multi-agent environments. The process layer: quarterly AI risk reviews, regular bias audit cycles, and recurring reviews of model performance against core business KPIs.
Once these three layers are fully implemented, governance actually speeds up deployment. Every new use case has a clear approval path and access to pre-built governance guardrails.
Treat AI as Digital Colleagues, Not Automation Tools
Both executives highlight a critical mindset shift: frame AI agents as digital colleagues, not advanced automation tools.
This is far more than just rhetoric. Traditional automation tools only execute pre-defined rules. A digital colleague needs to exercise judgment in the context of your unique business, escalate appropriately when facing anomalies, and leave a fully auditable trail for all its actions. Singh notes that this mindset shift accounts for 70% of all enterprise AI implementation challenges — no platform technology can solve that for you.
Leiduck shares a clear outlook for the future: industry focus will shift away from individual monolithic applications to the agent value layer, where work is orchestrated across systems. Delivery will become industrialized, with the rise of agent factories and new outcome-based payment models. Customer-specific knowledge graphs and ontologies will become first-class enterprise assets, because they encode workflows, business semantics, policies and data relationships, forming the blueprint for secure, high-performance enterprise agent architectures.
Singh’s vision is more decision-focused: Legacy enterprise systems told leaders what already happened. The next generation of systems will help leaders understand what is happening right now, what is likely to happen next, and exactly what actions to take. He describes a scenario where a CFO opens their system in the morning, and an agent that continuously analyzes financial, supply chain, procurement and workforce data proactively surfaces the three most high-impact issues that will affect the current quarter, paired with recommended actions and quantified financial impact. That is not a dashboard — that is an entirely new decision-making model.
Both executives agree on one core point: when designing your AI architecture, build it for the capabilities AI will have 18 months from now, not the capabilities available today. Strong data foundations, a well-built governance framework, and flexible architecture are the prerequisites to absorb future innovations as they emerge.
This article is based on the MIT Technology Review Business Lab podcast, co-produced with Infosys.
发布时间: 2026-11-16 21:00