SeriesThe Data Mesh DiariesPart 1

Data Architecture Data Mesh

The Data Lake That Became a Swamp

Enterprises spent billions on data platforms. Most of that data sits unused, untrusted, and completely unfit for AI agents. This is not a technology problem.

When a human encounters a data swamp, the consequence is a delayed report. When an autonomous AI agent encounters one, the consequence is an automated business failure — executed at scale.

55% of enterprise data is never used (Gartner, 2024). Not looked at. Not analysed. Not trusted enough to build a single report on. And this number has not improved in five years — despite billions spent on cloud platforms, data lakes, and tools that promised to fix exactly this.

Historically, this was an expensive nuisance. A human analyst would open a ticket, wait three weeks, notice a discrepancy, and manually fix it in a spreadsheet. Humans adapt. We compensate for messy centralised data architectures using tribal knowledge, emails, and guesswork.

But we are no longer just building dashboards for humans. We are entering the era of Agentic AI — autonomous AI agents designed to perceive, plan, and execute business actions independently. And this is where the old model hits a brick wall.

I have spent close to a decade helping enterprises build data infrastructure. I have seen organisations invest in world-class platforms. I have sat in rooms with impressive architecture diagrams on the screen — modern, cloud-based, well-designed. And I have watched, again and again, as those same platforms quietly stopped delivering. The data team was buried in requests. Business teams were still working off old spreadsheets. And the newly deployed AI agents sat completely paralyzed — unable to safely automate a single workflow because the data underneath them could not be trusted.

The platform was not the problem. The way it was run was.


55%

of enterprise data is never analysed or used (Gartner, 2024)

68%

of enterprise data considered completely unused or underused (Splunk, 2023)

65%

of government data stored but never applied to decisions (Forrester, 2023)


How We Got Here

To understand why modern AI is breaking on legacy infrastructure, it helps to trace how enterprise data management has evolved.

It started with data warehouses in the 1980s and 1990s. Bill Inmon — widely known as the father of the data warehouse — had a clear idea: collect data from all your systems, clean it up, bring it together in one place, and use it to make decisions. For a long time, this worked.

But companies kept growing. And so did their data. The warehouse handled clean, organised information well. It struggled with everything else — website logs, emails, social media posts, sensor readings, documents. This newer type of data did not fit neatly inside it.

So the data lake was born. The idea was simple: store everything in one place — organised or not — and figure it out later. No strict rules upfront. Just pour everything in. It sounded like the perfect solution for both big data and machine learning.

In practice, it created a structural bottleneck nobody expected.

It was Zhamak Dehghani — a software engineer and architect at ThoughtWorks — who first put a name to why this kept failing. In a 2019 article that became a foundational text in the data industry, she argued that the problem was never the technology. It was the model. Centralising data, she said, creates the same bottlenecks and quality problems that centralising software development once did. Her answer was a different way of organising data entirely — and she called it Data Mesh.

"We called it a data lake. What we actually built was a data swamp — a place where data goes in, and nothing useful comes out."

How a Lake Becomes a Swamp

In most enterprises, data is organised this way: every department — Sales, Finance, Operations, HR, Customer Service — generates data as part of their daily work. That data flows into a central data lake, managed by a single central data team. Once it arrives, that team owns it completely. They build the pipelines, define the structures, answer the requests, and maintain the quality. The business domains that generated the data move on. They consider their job done.

The central team is typically made up of skilled data engineers and analysts. But they are generalists. They understand technology. They do not always understand the business deeply enough to know what the data actually means — and that distinction matters more than most people realise.

A swamp is not a lake that was built badly. It is a lake that was never looked after. And the journey from one to the other follows the same path — in every industry, every country, every size of organisation.

The first few months feel exciting. Data comes in from the ERP, the CRM, the finance system. The central team works hard. Early reports look good. Leadership is happy.

Then more teams join. More sources are connected. The lake gets bigger. But nobody set clear rules about who is responsible for what. Who checks the quality? Who keeps the definitions consistent?

Slowly, the problems appear. One department calls something a "customer." Another calls the same person a "client." One system shows gross revenue. Another shows net. Which number is correct? Nobody can say for certain.

Human teams go back to their spreadsheets. Meanwhile, the data science team tries to connect an AI agent to the lake — only for the agent to hallucinate or break because it cannot reconcile the conflicting data definitions.

The lake is full. The lake is useless. A new platform is proposed.

Three Reasons Centralised Data Fails — and Why It Kills Agentic AI

After years of watching this pattern, I believe centralised data management fails for three core reasons. None of them are technical. But all of them are fatal to autonomous AI.

The Latent Context Gap — no one truly owns the data. When a central IT team is responsible for all data pipelines, no business team feels responsible for the data itself. The Sales team sends their raw data to the lake and considers their job done. When you centralise data, you strip away its latent context — the business knowledge behind the numbers. A human analyst can call a colleague to ask why a field was overridden. An AI agent cannot. It reads the data literally. And it acts on it — at speed, at scale, without hesitation.

API Blindness and Stale Data — one team cannot serve everyone. Every new report, every schema change, every new data source goes to the central team's backlog. As the organisation grows, this becomes a serious bottleneck. But there is a deeper problem for AI. Centralised platforms are engineered for historical, batch-processed data — pipelines running overnight. Agentic AI operates in continuous, real-time loops. An agent trying to optimise supply chain routes or flag fraud using a 24-hour-old data snapshot is essentially acting blind.

Structural Misalignment — the builders do not understand the business. Central data engineers are technically skilled. But they do not deeply understand each individual business domain. A metric can be technically flawless according to database rules and completely wrong in business context. Autonomous agents require precisely defined, programmatically enforced interfaces — known as Data Contracts — to function safely. A central team simply cannot maintain hundreds of distinct data contracts for domains they do not intimately understand.

A Pattern That Plays Out Everywhere

I have seen this in large government entities. I have seen it in global banks. I have seen it in fast-growing technology companies. The geography and the industry change. The pattern does not.

In many of the large organisations I have worked with — especially the incumbents — where approval chains are long and decisions travel upward, teams naturally avoid taking on data ownership. Accountability without real authority is a burden, not an opportunity. If something goes wrong with the data, the team that raised their hand will be answerable. If nothing goes wrong, nobody notices. That is not an incentive to volunteer.

The same dynamic plays out in organisations everywhere. Hierarchical structures, centralised IT functions, and procurement models that reward platform purchases over organisational change — these are global patterns. The platforms keep getting better. The ownership model stays the same. And now, with AI agents depending on that data to take autonomous business actions, the cost of that unchanged model is no longer just slow reports. It is automated decisions made confidently on wrong information.

So What Is the Answer?

There is a better way to think about enterprise data. Instead of asking "where does the data live?", ask "who is responsible for it — and who benefits from it?" Instead of treating data as an IT storage problem, treat it as an organisational product that must be owned, maintained, and delivered by the people who understand it best.

That way of thinking has a name: Data Mesh. It was conceived by Zhamak Dehghani, and it is the subject of Part 2. I have learnt a great deal from her writing and her work — and this series owes much of its intellectual foundation to her thinking.

Data mesh is not a magic fix, and it is not right for every organisation. I will cover its real limitations later in the series. But its central idea — that data should be owned and delivered like a product by the people who understand it best — is the most important shift in data thinking of the past decade. And in a world of autonomous AI agents, it is no longer optional.

Understanding why the old model breaks is the first step. We carry that understanding into Part 2.


Further reading: Zhamak Dehghani's original 2019 article — How to Move Beyond a Monolithic Data Lake to a Distributed Data Mesh — is the piece that started this entire conversation.


Coming next — Part 2: Everybody Is Buying Data Mesh. But Are They Implementing It? where we explore what data mesh is.

Knowledge Check
0 / 4 ☆☆☆☆
Q1 Multiple Choice

An AI agent is deployed to automate fraud detection using a centralised data lake that runs overnight batch updates. The agent flags transactions as fraudulent based on 18-hour-old data. What is the root cause of this failure?

A
The AI model was not trained correctly on fraud patterns
B
The centralised data architecture is engineered for batch processing, not real-time agent decision-making
C
The fraud detection rules in the pipeline were outdated
D
The central data team did not test the agent before deployment
💡 Insight

Centralised data platforms are designed for historical, batch-processed ingestion — pipelines that run overnight or on a schedule. Agentic AI operates in continuous real-time loops. When you connect a real-time agent to a batch-driven data source, the agent is always acting on stale context. This is not a model problem or a rules problem — it is a structural mismatch between the architecture and the use case.

Q2 True / False

A technically correct data metric — one that passes all database validation rules — is always safe to use as the basis for an autonomous AI agent's decisions.

True
False
💡 Insight

Technical correctness and business correctness are different things. A field can pass every database constraint and still be wrong in business context — for example, a 'closed' deal that was manually overridden by a sales manager for legitimate reasons. A human analyst would know to question it. An AI agent reads it literally and acts on it. This is the Latent Context Gap — and it is one of the primary reasons centralised data is dangerous for autonomous AI.

Q3 Multiple Choice

A large organisation's central data team is struggling to maintain data contracts for 200 data sources across 15 business domains. Which of the following best explains why this is structurally difficult — regardless of team size?

A
Data contracts are too technically complex for most data engineers to write
B
200 data sources is simply too large a volume for any team to manage
C
The central team lacks the domain-specific business knowledge needed to define and enforce contracts meaningfully
D
The organisation's data catalog technology is not advanced enough
💡 Insight

The problem is not volume or technical complexity — it is domain knowledge. A data contract for the Sales domain must encode what 'active customer' means in that specific business context, what the acceptable staleness window is for a pricing feed, what constitutes a valid churn event. Only someone embedded in Sales every day knows these things. A central team managing 15 domains cannot have deep enough context across all of them to define contracts that are meaningful rather than superficial.

Q4 True / False

Investing in a more advanced cloud data platform will solve the core problem described in this article.

True
False
💡 Insight

The core problem is organisational — unclear data ownership, a central bottleneck, and domain context stripped away from the data. These are not infrastructure problems. A more advanced platform stores the same unowned, contextually stripped data more efficiently. The pattern of data lakes becoming data swamps has repeated across every generation of platform technology. The platform is not the variable that changes the outcome. The ownership model is.

Q5 Reflection

Your organisation has just deployed an AI agent to automate customer churn prevention. Two weeks in, the agent is sending retention offers to customers who already churned months ago — and missing genuinely at-risk customers entirely. The CTO wants to retrain the model. You believe the problem is deeper. How would you diagnose the real issue and what would you recommend investigating first?

Model Answer

A strong answer does not start with the model. It starts upstream — with the data the model is acting on. The investigation should cover: How is 'churned' defined in the data the agent consumes, and who defined it? Is the data the agent uses current, or is it batch-processed on a delay? Does the central data team who built the pipeline understand what 'at-risk' means in this business's specific customer lifecycle? Are there conflicting definitions of churn across different source systems being merged without reconciliation? The recommendation is to audit the data contract (or lack of one) for the churn data product before touching the model. In most cases, the model is performing exactly as trained — on the wrong data.

💡 Insight

Retraining a model on bad data produces a better-trained model that makes the same wrong decisions more confidently. The diagnostic instinct here is to treat the AI agent as a symptom-revealer, not the root cause. Every hallucination or bad decision an agent makes is pointing at a data quality, data freshness, or data definition problem upstream. That is the problem worth solving first.

0
out of 4 correct

Question 1 of 5