なぜエージェンティック分析は、適切に管理されたデータレイヤーから始まるのか

As AI changes how executives interact with data, analytics is moving out of the dashboard era and into a far more dynamic operational model. Natural-language interfaces, AI-driven insights, and agentic workflows promise broader access to intelligence, but they also expose a problem many organizations have lived with for years: fragmented definitions, inconsistent metrics, and governance models that were never designed for AI scale.

To unpack what that means in practice, I spoke with Nick Eayrs, Vice President of Field Engineering for Asia-Pacific and Japan at Databricks. With nearly 25 years in leadership across multiple regions, Eayrs has a broad view of how data insights can be an accelerator within organizations, and what it takes to succeed in the new era of agentic analytics.That background gives him a broad view of how data and AI strategies play out across markets, operating models, and levels of enterprise maturity.

The throughline of our conversation was his conviction that AI is not eliminating the need for semantics and governance. It is making them far more important. In his view, organizations will not get trusted AI outcomes until they fix the data layer beneath them: the business definitions, lineage, access controls, and open standards that allow intelligence to scale without collapsing under cost and complexity.

AI Is Rewriting the Rules of Analytics

Catherine Brown: Why does AI put semantics and governance pressure on analytics in a way that legacy BI never had to deal with?

Nick Eayrs: Legacy BI was really a world of static dashboards and predefined reports. Business users had to navigate fairly complex interfaces, and if they had a follow-up question or wanted to explore something more deeply, they usually needed specialist support. There was very little true self-service.

The semantic layer underneath traditional BI was also relatively static and slow to change. If the business needed a new definition for revenue, churn, or customer lifetime value, that usually meant going back to IT or specialist teams to update the semantic layer and rebuild reports. It was a very predetermined model.

AI changes that completely. It no longer has to be static, and it no longer has to be purely descriptive. Traditional BI is often rear-view-mirror analytics. It tells you what happened. With AI, you can start to predict what might happen, ask why it happened, and understand what to do next. You can reason across much more data yourself and generate insights in real time.

But semantics do not go away in that world. If anything, they matter more. AI and agents are still informed by the data underneath them. That gets back to the old principle of garbage in, garbage out. The more trusted, high-quality data you have, with the right business context around your products, services, taxonomy, and terminology, the better the AI experience will be.

If someone asks, “Why did we miss our Q3 targets?” the system needs to understand what “targets” means in that organization, what period the user is referring to, and how those metrics are defined. Without that semantic context, the system is just guessing. It may produce generic answers, but not trusted ones.

There is another important point here as well. In the Databricks view, the semantic layer should be open and interoperable. Traditional BI vendors often lock the semantic model into their own tool, which means everything has to flow through that interface. That becomes a major constraint. If you want AI and agentic experiences to scale, a strong custom example in APJ would be Takeda. With the right data foundations and guardrails in place, they were able to build out multiple AI use cases across commercial, R&D, manufacturing, and back office functions.

Catherine: Can you talk more specifically about the governance pressure AI puts on analytics?

Nick: On both the BI side and the AI side, governance comes down to trust, lineage, and traceability.

If you are producing dashboards or business intelligence insights, you need to understand how they were built. Which underlying data was used? How were the metrics defined? If you do not know that, then you cannot trust what you are looking at.

The same is true on the AI side. You are not going to trust the output from a model, an agent, or an agentic application if you cannot understand how that output was derived. Which table did it come from? Which features were used? Which model was serving the inference? That end-to-end lineage is essential.

There is also a compliance dimension. In highly regulated industries, organizations are increasingly going to be required to prove that traceability. If an AI-driven decision is being exposed externally to consumers, citizens, or patients, you have to be able to stand behind it and audit how it was created. AI is putting more pressure on analytics because the expectations around trust and traceability are rising.

Fragmented Metrics Are Slowing Decisions

Catherine: What are the most common conflicting metrics patterns you see, and what do they cost organizations?

Nick: The biggest issue is fragmentation. Most organizations have multiple BI tools in the estate, and each of those tools may have its own semantic model and its own interpretation of business metrics. That means you end up with no single source of truth and a lot of duplicated logic that may not align.

One dashboard might define revenue one way. Another tool may define it differently. Someone in finance may be working from another version in Excel. At that point, trust starts to erode very quickly. Decision-making slows down because people are no longer debating the decision itself. They are debating which number is right.

Why Legacy BI Models Break at AI Scale

Catherine: Why does dashboard logic, when it is trapped in tools, collapse under AI scale?

Nick: Traditional BI tools often extract data out of source systems, aggregate it for a specific reporting outcome, move it into proprietary storage, and then layer proprietary semantics and dashboards on top of that. Everything gets locked into the tool.

That becomes a real problem in an AI world because users always have follow-up questions. They want to go deeper. They want to expose that logic to other systems. They want data scientists or machine learning teams to build on it. If everything is trapped in one proprietary layer, that does not work well. You have to keep going back to the source, pulling more data, transforming it again, and rebuilding the logic. It becomes repetitive and expensive.

If, instead, everything is built on open data formats and open interfaces, then BI, AI, notebooks, agents, and data science teams can all work from the same governed foundation. You store and process the data once. Everyone can interact with it in natural language. Everyone can build on it. That is a much better model for scale.

There is also a significant engineering burden in the old way of doing it. You end up maintaining lots of synchronization pipelines and a lot of custom code just to keep fragmented systems aligned. That complexity becomes very hard to justify.

What a Machine-Readable Semantic Layer Looks Like

Catherine: What does a machine-readable semantic layer look like in practice?

Nick: First, business metrics have to be treated as a foundational pillar. That means the definitions of things like revenue, churn, or customer lifetime value need to be explicitly defined, certified, and reusable across the organization.

Second, those metrics need to be accessible through standard languages, primarily SQL, and they need to be consumable not just by BI tools but by AI interfaces, notebooks, and agents as well. If they are not accessible and reusable, you have not really solved the problem.

Third, you need openness and interoperability. You do not want to push all of your business logic into a system that you cannot get it back out of. Open standards matter because they give you optionality and a safe exit strategy if you ever need to change systems or providers.

You also need AI-enabled governance. In an agentic world, you may have thousands of models or agents interacting with the semantic layer all the time. Keeping metadata, comments, and business metrics current is a huge challenge if that is all done manually. AI can help generate and maintain that metadata so the semantic layer stays usable at scale.

And then, of course, you need conversational and contextual intelligence on top so that agents and applications can interact with that layer through APIs and natural-language interfaces.

Catherine: Where does evaluation fit into this? Does the certification of the data happen first, and then the AI layer and evaluations come after?

Nick: Yes. The data foundations come first. You need the metadata, the business logic, the comments, and the business metrics in place before AI can use that data well.

Then you build the AI or agentic layer on top of it. After that, the evaluation frameworks come into play to validate whether the outputs are aligned with expectations and to refine what the system is doing. But the evaluation layer is not a substitute for getting the foundations right. It depends on those foundations.

Why Per-Seat BI Models Limit Adoption and Value

Catherine: Per-seat BIモデルは、導入と価値創造を積極的に妨げているのはどのような場合ですか？

Nick: データとAIの民主化の目標は、組織内のすべてのナレッジワーカーにインテリジェンスを提供することです。Per-seatモデルは、その目標に直接反します。

組織がどのユーザー、チーム、またはビジネスユニットにアクセス権を与えるかを強制するため、民主化を制限します。また、ライセンスの可用性ではなく、ビジネス価値に基づいてどのプロジェクトを進めるかを決定するため、イノベーションも制限します。

それは価値創造にも影響します。多くの場合、多様なチームがビジネス上の問題を中心に集まったときに、最良の結果が得られます。それらのチームの一部しかシステムにアクセスできない場合、コラボレーションが制限され、組織の価値創造能力が制限されます。

もう1つの問題は効率です。消費ベースのモデルでは、使用した分だけ支払います。使用量がスケールアップすれば、使用量の増加分を支払います。ゼロにドロップすれば、ゼロを支払います。これは、過小評価されているか過剰にプロビジョニングされている可能性があるシートライセンスに対して支払うよりも、はるかに合理的なモデルです。

Catherine: ライセンス制限が実質的にガバナンスレイヤーとして機能していると主張する組織もあるかもしれません。それについてどう思いますか？

Nick: ライセンスを制限することによってデータへのアクセスをガバナンスしようとしているのであれば、失敗するでしょう。それは間違った制御ポイントです。

優れたガバナンスは、プラットフォームとデータレイヤーから始まります。ロールベースおよび属性ベースの制御、IDシステムに連携された認証と承認、データ資産の明確な分離と分類から始まります。権限とポリシーの施行は、事前に解決します。

それを適切に行えば、人々が本来見るべきものだけを見ていることを保証しながら、アクセスを広範囲に展開できます。シートライセンスをガバナンスメカニズムとして使用することはスケーラブルではなく、根本的なガバナンス作業の代わりにはなりません。

信頼性を向上させ、コストを削減する最も速い方法

Catherine: 組織が信頼性を向上させ、分析コストを同時に削減するために行うことができる最も速いアーキテクチャの変更は何ですか？

Nick: 最も重要な変更は、強力なガバナンス基盤に基づいた統一されたセマンティックレイヤーを確立することです。

それはカタログの決定から始まります。データとAI資産をどのようにガバナンスしますか？カタログが配置されたら、そこでセマンティクスを定義し、そこでビジネス指標を認定し、単一の真実のソースを作成できます。Databricksモデルでは、その真実のソースはオープンで相互運用可能であり、それは非常に重要です。

それを実行すると、いくつかのことが起こります。ラインエージ、ガバナンス、監査可能性、および認定された定義があるため、信頼が得られます。不要な重複や繰り返しのETLを回避するため、簡素化が得られます。誰かが新しい質問をするたびにロジックを再構築する必要がなくなるため、ITの負担が軽減されます。

実装パターンはかなり明確です。まず、データ基盤を正しく設定します。次に、セマンティックレイヤーを構築し、ビジネス指標を認定します。3番目に、AIをレイヤー化し、評価フレームワークを使用してそれらの出力を監視および洗練します。そのシーケンスが重要です。NTT Docomoはその良い例です。Databricks Lakehouse、Unity Catalog、およびワークフローを使用してログ分析を自動化することにより、月間66時間の手動処理時間を6時間に短縮し、分析効率を90パーセント向上させました。これは、ガバナンスと基盤がはるかに迅速な意思決定を可能にする強力な例です。

APJがデータとAIの収益化をより速く進めている理由

Catherine: APJの企業は、データレイヤーをAIに収益化する上で、どのような違いや速さで取り組んでいますか？

Nick: APJは、非常に多様な市場であるため、興味深い市場です。非常に異なる国、言語、成熟度レベル、規制環境を扱っています。しかし、共通のパターンは、組織がデジタルトランスフォーメーションに非常に迅速に進む傾向があり、この地域の多くの政府が明確な国家AI戦略を策定していることです。

顧客から見ると、多くの場合、ガバナンスとデータ基盤レイヤーから始め、その基盤が整ったらAIネイティブアプリケーションに迅速に進みます。そのシーケンスが重要です。

金融サービスのような業界でもそのパターンが見られます。そこでは、顧客はガバナンスされたデータレイヤーの上に分析を統合し、アクセスを民主化しています。

もう1つの例は、日本のNet One Systemsです。基盤を整えた後、他のシステムと統合されたAI搭載のナレッジツールを構築し、サポートクエリへの応答時間を75パーセント短縮し、年間10,000時間の労力を節約しました。

APJに特にユニークな点の1つは、多言語の側面です。顧客は、日本語、中国語、広東語、タイ語、その他の現地語で機能を構築しています。これは強力ですが、基盤となるデータレイヤーがそれをサポートするのに十分なガバナンスと構造化がされている場合にのみ機能します。

APJの顧客は、基盤を迅速に正しく設定し、その上にAIファーストのアプリケーション開発に迅速に移行する傾向があります。多くの場合、他の地域よりも速く進んでいます。

締めくくり

Nickのポイントは、技術的かつ戦略的です。AIから価値を創造している組織は、分析、セマンティクス、ガバナンスを別々の会話として扱っていません。それらを1つの基盤として扱っています。経営者にとって、それは価値があるのは、より良いアーキテクチャだけではないからです。それは、より迅速な意思決定、洞察へのより広範なアクセス、および大規模な分析コストの削減です。AIは断片化されたデータレイヤーを修正しません。それを露呈させます。実験から信頼できるインテリジェンスへと最も速く移行する企業は、メトリックを明確に定義し、それらを中央でガバナンスし、分析とAIが同じ真実の上に構築できるように十分にオープンにする企業です。

効果的な運用モデルの構築について詳しくは、Databricks AI Maturity Modelをダウンロードしてください。

(このブログ記事はAI翻訳ツールを使用して翻訳されています) 原文記事