TL;DR: The promise of AI-driven business intelligence has often been derailed by LLM hallucinations and data ambiguity. Anthropic’s recent revelations on deploying Claude for internal self-service analytics offer a robust, battle-tested blueprint. By shifting the focus from code generation to stringent data foundations and a human-governed semantic layer, Singaporean SMEs can circumvent the local talent crunch, automating up to 95% of rote data queries and transforming disparate software silos into unified, conversational intelligence.
A morning walk through the bustling Central Business District of Singapore, perhaps darting into a meticulously curated independent coffee roastery on Telok Ayer or an agile logistics startup near Tanjong Pagar, reveals a paradox of modern enterprise. On the surface, these small and medium-sized enterprises (SMEs) are triumphs of digitalisation. Propelled by the government's Smart Nation initiatives and Productivity Solutions Grants (PSG), they operate on a sleek veneer of cloud-based Point of Sale systems, automated HR platforms, and accounting software like Xero.
Yet, ask the founder or the operations director a seemingly simple question—"What is our true cost of customer acquisition factoring in last month’s supply chain disruptions?"—and the digital facade crumbles. The answer is inevitably buried across three different platforms, requiring days of manual Excel wrangling.
Enter the promise of Generative AI. The prevailing narrative suggests that one simply points a Large Language Model (LLM) at a data warehouse, and voila: instant, conversational insights. But as Anthropic, the pioneering AI research company behind the Claude models, recently detailed in their engineering blog, the reality is far more complex. Allowing an AI agent unchecked access to a data warehouse often creates a false sense of precision. Initial elation gives way to dread as stakeholders realise the AI is confidently pulling from deprecated tables or misinterpreting business definitions.
Anthropic’s data science team, however, has managed to automate 95% of their internal business analytics queries using Claude, achieving an astonishing 95% accuracy rate. They achieved this not by building a more creative AI, but by rigorously structuring their data environment. For Singaporean SMEs—operating in a hyper-competitive, talent-constrained environment where hiring a five-person data engineering team is an economic impossibility—Anthropic’s agentic analytics stack is not just a technical curiosity. It is a highly practical blueprint for survival and scale.
In this briefing, we will dissect Anthropic’s methodologies and apply them directly to the Singaporean SME landscape, exploring how local businesses can engineer a self-service data architecture that actually works.
The Core Problem: Data is Not Software (And Why LLMs Hallucinate in BI)
To understand how to deploy Claude for analytics, one must first unlearn the lessons of AI in software engineering. Anthropic sharply observes that coding is an open-ended solution space. It rewards an LLM's creativity, and natural guardrails like unit tests and compilers catch hallucinations.
Analytics, conversely, is heavily deterministic. There is usually only a single correct answer, derived from a single correct source, but no automated compiler exists to prove the AI got it right. The complexity lies not in generating the SQL code, but in the ambiguity of the data itself.
Anthropic identifies three primary failure modes when LLMs attempt data analytics:
- Concept vs. Entity Ambiguity: An SME's data model might contain hundreds of viable tables and thousands of fields. When a marketing manager asks Claude for the number of "active users," the AI faces a semantic void. Does "active" mean a user who logged in within 30 days? Does it include fraudulent or test accounts? Does it count users who opened an app but made no transaction? Without explicit mapping, the AI guesses—often incorrectly.
- Data Staleness: Singapore’s business environment is notoriously fast-paced. A retailer might pivot from sourcing in Guangzhou to Johor Bahru, fundamentally altering how logistics data is categorised. Business definitions change, schemas are updated, and if the AI’s context goes stale, it will return subtly wrong answers based on historical truths that no longer apply.
- Retrieval Failure: Sometimes, the correct data and the correct annotations exist, but the search space is so vast that the agent simply fails to find them.
For a mid-sized local enterprise—say, a boutique wellness chain with outlets from Orchard Road to Katong—these failures mean the difference between optimising staff rosters for peak hours and over-staffing a quiet Tuesday. If Claude is to become the ultimate analyst, the underlying infrastructure must be purposefully designed to eliminate these three errors.
Layer 1: Data Foundations—Taming the SME Data Sprawl
The foundation of Anthropic’s success is not a proprietary AI breakthrough, but a return to rigorous, old-school data engineering, recalibrated for an AI end-user.
Standard practices like dimensional modelling and shift-left testing remain vital. However, the paradigm shift is that the consumer of the data is no longer a human data scientist who can intuitively spot an anomaly. The consumer is an AI agent acting on behalf of a non-technical stakeholder.
To solve Entity Ambiguity, SMEs must create Canonical Datasets. A common trap for growing businesses in Singapore is the proliferation of "near-duplicate" data. The marketing team has a spreadsheet for revenue, the finance team relies on Xero, and the operations team looks at Shopify dashboards. When an AI searches for "revenue," it finds three different answers.
Anthropic advocates for curating a small set of canonical, single-source-of-truth datasets. These must be clearly owned, consumption-ready, and highly discoverable. Any downstream rollups or caches must derive mechanically from these canonical models. The objective is brutal simplicity: when Claude searches for a concept, it must hit a single, governed answer.
Furthermore, these standards must be strictly enforced. In a fast-moving SME, it is tempting to bypass governance for a quick ad-hoc analysis. Anthropic warns that governance without enforcement quickly decays back into ambiguity. Tooling must route the agent structurally to the canonical models first.
Finally, Metadata must be treated as a first-class product. LLMs excel at reading codebases because developers write READMEs, docstrings, and type signatures. An SME's data warehouse must be equally legible. Column descriptions, valid value ranges, and model tiering must be maintained with rigorous discipline. A table named
txn_vol_final_v2 means nothing to Claude without robust metadata explaining that it represents daily settled transaction volumes in Singapore Dollars, excluding refunded orders.Layer 2: Sources of Truth (The Semantic Layer)
If the data foundations represent the physical warehouse of information, the Sources of Truth represent the catalogue the AI consults to navigate it. This layer is the critical bridge between human business language and database architecture.
The crown jewel of this layer is the Semantic Layer. This is where metrics and dimensions are explicitly defined and compiled. If an SME director asks Claude about "Monthly Recurring Revenue," the semantic layer ensures the agent calls a specific function that retrieves the exact same number every single time, matching the figures the CFO reports to the Inland Revenue Authority of Singapore (IRAS).
Crucially, Anthropic shares a vital lesson from their own experimentation: Do not let the AI define the metrics.
Initially, Anthropic attempted to bootstrap their semantic layer by having an LLM auto-generate metric definitions from raw tables and historical query logs. The result was a set of plausible-sounding definitions that secretly encoded the exact ambiguities they were trying to escape. The lesson for Singaporean businesses is clear: use Claude to help write the documentation, but a human must ultimately own and hard-code the business definition. "Profit" must be defined by human commercial logic, not probabilistic text generation.
When the semantic layer does not cover a specific, esoteric question, Anthropic relies on Lineage and Transformation Graphs. This allows Claude to reason about which upstream models feed a concept, essentially letting the AI conclude, "I don't know the exact metric, but I know which governed model to aggregate the data from."
Interestingly, Anthropic found that giving Claude raw retrieval access to thousands of past SQL queries (a Query Corpus) barely improved accuracy. Unstructured retrieval struggled to map new questions to old precedents. Instead, distilling past queries into structured, per-domain reference documents proved far more effective.
Layer 3: Agent Skills and Addressing Staleness
Equipping the data environment is only half the battle; the AI must be taught how to navigate it. Anthropic achieves this through specific "Skills" (instructions or tool-calling capabilities granted to the LLM).
A foundational skill is instructing Claude to always consult the semantic layer first before attempting to write custom SQL. If a metric exists in the semantic layer, the agent must use it. Only if the query falls outside defined metrics should the agent use lineage tools to find the canonical tables and write its own query. This structured routing acts as a digital guardrail, keeping the AI on the path of governed truth.
But how does an agile SME, constantly shifting its strategies in the dynamic ASEAN market, prevent this meticulously crafted environment from going stale?
Anthropic’s primary defence is Colocation of Artifacts. All data code—the models, the semantic layer, the reference docs, and the dashboard definitions—lives in a single repository. When a data engineer (or an outsourced IT vendor managing the SME’s infrastructure) changes a core business logic, automated Continuous Integration (CI) checks ensure that any downstream dashboards or documented metrics that would be broken by the change are flagged immediately. The fix must ship in the same update.
For a Singaporean SME, this means shifting away from disjointed software management. Rather than having a freelance web developer handle Shopify, a separate accountant handle Xero, and an internal manager handle inventory on Excel, the business must move towards modern data stack tools (like dbt combined with a warehouse like BigQuery or Snowflake) that centralise business logic.
The Economic Imperative for Singaporean SMEs
Why should a medium-sized enterprise in Tampines or Jurong care about the internal data architecture of a San Francisco AI lab? The answer lies in the structural realities of Singapore’s economy.
Singapore is currently navigating a severe talent crunch. The Ministry of Manpower (MOM) has steadily tightened Dependency Ratio Ceilings (DRC) and raised qualifying salaries for Employment Passes. Hiring a dedicated team of data scientists and analytics engineers is a luxury reserved for multinational corporations and heavily funded tech unicorns.
Traditional BI tools promised self-service but ultimately required end-users to learn SQL or navigate maddeningly complex dashboard interfaces. As a result, data requests bottlenecked at the desk of the solitary IT manager or the mathematically inclined co-founder.
Claude, when deployed atop Anthropic’s agentic analytics stack, shatters this bottleneck. By handling 95% of rote, repetitive business questions—"How did the recent Hari Raya promotion affect margin across our top three product lines?"—the AI functions as an always-on, infinitely patient data analyst. It allows the human talent within the SME to focus on strategic foresight: causal modelling, supply chain forecasting, and high-level negotiations.
Furthermore, this setup aligns perfectly with the Singapore government’s push for high-value productivity. Investing in the foundational data engineering required to support Claude is highly grant-eligible work, positioning forward-thinking SMEs to leverage state support for structural transformation.
Key Practical Takeaways
- Audit Your Data Sprawl: Identify where your "near-duplicate" data lives across different SaaS platforms. Your first step to AI readiness is consolidating these into a single cloud data warehouse.
- Establish a Human-Governed Semantic Layer: Define your core business metrics (e.g., Gross Margin, Active Customers, Churn) centrally. Do not allow the AI to guess these definitions; hard-code them so the AI acts purely as a retrieval engine.
- Invest in Metadata: Treat descriptions, tags, and column definitions as critical infrastructure. If a new employee wouldn't understand what a table column means, Claude won't either.
- Mandate the Semantic Layer: Instruct your AI agents to always query the semantic layer first before attempting to write raw SQL. Use prompt engineering to enforce this hierarchy of trust.
- Colocate Data Logic: Keep your data models, metrics, and documentation in a single repository. Ensure that a change to how data is collected immediately triggers an update to how it is defined for the AI.
Frequently Asked Questions
Why can’t I just connect Claude directly to my existing SQL database and let it figure things out?
Without a governed semantic layer and explicit metadata, an LLM will suffer from "concept-entity ambiguity." It will guess which tables and columns hold the right answers, often pulling from deprecated or duplicate datasets, resulting in confident but mathematically incorrect insights that can damage business decision-making.
Is building this "agentic analytics stack" too expensive for a typical Singaporean SME?
Not anymore. Cloud data warehouses (like BigQuery) and transformation tools (like dbt) offer pay-as-you-go pricing that scales with your business. Furthermore, local SMEs can tap into IMDA's SMEs Go Digital programme or the Enterprise Development Grant (EDG) to subsidise the initial setup costs of centralising their data infrastructure.
Should we use AI to write our business definitions and metric formulas to save time?
No. Anthropic’s own testing revealed that having an LLM auto-generate metric definitions from raw tables encodes ambiguity and leads to errors. Humans must define the strict commercial logic (what constitutes "revenue" or "profit"), while the AI should be used to retrieve that data and perhaps generate the supplementary documentation describing it.
For Further Reading:
No comments:
Post a Comment