Six Signs Your Company Has a Data Problem
- 7 days ago
- 7 min read
A plain-language guide for Australian business leaders who want honest answers about why AI initiatives stall — and what to do first
It usually starts with a board meeting.
Someone puts AI on the agenda. The case is compelling: competitors are moving, vendors are showing impressive demos, and the question is no longer whether to adopt AI but how quickly. The business commits. A pilot is approved. A tool is subscribed to or a vendor is engaged. Six months later, the results are disappointing. The AI assistant gives inconsistent answers. The dashboard contradicts itself depending on how you ask the question. The automation that looked seamless in the demo breaks on real data in ways nobody anticipated.
Leadership starts to wonder whether AI actually works — or whether their organisation got something wrong.
In most cases, the organisation did not get the AI wrong. They skipped a step.
We work with mid-sized organisations across professional services, resources, logistics, and property — and the pattern is almost universal. The organisations that struggle with AI initiatives are not struggling because they chose the wrong tool or the wrong vendor. They are struggling because the data the AI needs to work with is not in a state that any model can make reliable use of.
Here are the six signs that the problem is your data, not your AI.
Data Problem Not AI Problem
Before getting to the signs, it is worth being direct about something that vendors rarely say: AI systems do not fix data problems. They expose them — loudly, expensively, and in front of the people you least want to be embarrassed in front of.
Every AI system, whether it is a predictive model, a generative assistant, or an automated decision engine, operates on data. The quality of what comes out is directly determined by the quality, consistency, and accessibility of what goes in. There is no model capable of overcoming a broken data foundation. There is no prompt clever enough to compensate for systems that do not agree with each other.
What AI readiness actually requires is data that is accurate and consistently defined across systems, data that can be accessed without a two-day IT request, data that is connected — so a customer in your CRM corresponds to the same customer in billing and your service history — and data that has someone responsible for its ongoing quality. When those foundations exist, deploying AI is relatively straightforward. When they do not, every AI project becomes a data cleanup project in disguise.
The Six Signs
These are the patterns we encounter most consistently when working with organisations that have struggled to gain traction with AI. Most organisations will recognise at least three. Many will recognise all six.
1. Your teams are arguing about the numbers
The monthly reporting meeting takes longer than it should because sales, finance, and operations are looking at different figures. Someone built a report in Excel. Someone else pulls from the CRM. Finance uses the ERP. All three numbers are defensible in isolation — they simply do not agree with each other.
This is not a reporting problem. It is a foundational data problem: there is no single source of truth, and the definitions that govern how numbers are calculated are not shared across the business. When AI is deployed on top of this inconsistency it does not resolve the disagreement — it amplifies it. An AI assistant asked to summarise monthly revenue will produce a different answer depending on which system it reads from, and nobody can tell it which one is authoritative. The technology performs exactly as designed. The problem is the environment it is operating in.
2. Answering a board question takes two days
A request comes in: how many active clients do we have in Queensland, broken down by service line, across the last three quarters? The answer exists somewhere in your systems. Getting it out, however, requires someone who knows which system to query, which table to join it to, and how to reconcile the inconsistencies in between. Two days later, an answer arrives — with caveats.
When basic reporting requires specialist intervention every time, data is not operationally accessible. AI tools that promise self-service analytics — ask any question, get an instant answer — cannot deliver on that promise if the data underneath is not structured, governed, and current. The vendor demo looked effortless because the vendor used clean, curated data. Production data in most mid-market organisations does not start that way.
3. Every system has its own definition of "customer"
Your CRM, your ERP, your billing platform, and your customer service tool all have a customer record. But they were built at different times, by different teams, with different assumptions baked in. A customer in the CRM is a contact. In the ERP it is an account. In billing it is a contract entity. These records overlap, but they are not the same thing — and there is no automated mechanism to reconcile them.
This is invisible day-to-day because your people know the difference. AI systems do not. An AI reasoning across these systems will either fail to join them correctly, or join them incorrectly and give a confident-sounding wrong answer. Both outcomes erode trust in the technology far faster than the technology deserves. The AI is not at fault. The architecture is.
4. Your data team is a bottleneck, not an enabler
Every data request goes through the same two or three people. They know where the data lives, how to extract it, and what quirks and exceptions to watch for. They are the institutional memory of the data estate. They also have a backlog measured in weeks, not days, and no clear path to clearing it.
This is not a headcount problem — it is a data architecture problem. When data is not properly modelled, documented, and made accessible, the knowledge required to navigate it concentrates in a small group. That dependency cannot be bypassed by an AI tool, because the AI requires the same navigational knowledge the team holds, just codified into the platform. Until it is codified, the bottleneck remains — and every AI initiative that depends on that data inherits it.
5. AI pilots perform well in demos and fail in production
This is the most common story we hear from organisations that have already started their AI journey. The vendor demo was compelling. The proof-of-concept looked credible. The pilot gets approved, real data gets connected, and performance falls well short of expectations.
The cause is almost always data quality. Vendor demonstrations use curated, cleaned datasets prepared specifically to showcase the technology at its best. Production environments contain gaps, duplicates, inconsistent date formats, and historical exceptions accumulated over years of real operations. AI models that perform well on clean data behave unpredictably when inputs are messy. This is not a flaw in the model.
It is a predictable consequence of skipping the data preparation step — and it happens consistently enough across organisations that we consider it a reliable diagnostic signal.
6. You are buying AI tools on top of broken foundations
Microsoft Copilot. Power BI AI features. A GPT-connected knowledge base. These are genuine tools with genuine value — in the right environment. When the underlying data is in order, they deliver on the promise. When it is not, each one becomes a new surface for the existing problem to show up in.
Copilot surfaces documents that are outdated or internally contradictory. Power BI AI features return figures that do not reconcile across reports. A GPT integration connected to your systems gives confident-sounding answers drawn from subtly inconsistent source data. The tool is not the problem. The sequence is. The data foundation must come before the AI layer — not run alongside it, and not follow it.
What the Path Forward Looks Like
The organisations we work with that have had genuine success with AI share one characteristic: they treated data readiness as a prerequisite, not a parallel workstream.
In practice, this means three things.
First, getting an honest picture of where data actually lives today — not where the documentation says it is, but where it genuinely lives across cloud platforms, legacy systems, spreadsheets, and third-party tools. Understanding the real state of the data estate is the starting point for every decision that follows.
Second, establishing a layer of connectivity and consistency. This does not require replacing every existing system. It means building a platform layer where data from disparate sources is brought together, aligned on common definitions, and made accessible in a governed and auditable way. On Azure Databricks — the platform we build on — this takes the form of a medallion architecture: raw data is ingested, cleansed and conformed, and surfaces as a trusted, queryable asset. The outcome matters more than the specific technology: one place where data is trusted, current, and ready to be used.
Third, deploying AI incrementally on the prepared foundation. We do not recommend waiting two years before AI gets any airtime. We recommend getting one business domain — customers, transactions, assets, workforce — into a reliable state, deploying an AI capability on that foundation, proving the value, and expanding systematically. This is the Connect → Optimise → Activate methodology we apply with clients: connect the data reliably, optimise the platform layer, then activate AI on a foundation that is actually ready for it.
Organisations that follow this sequence do not spend six months wondering why the pilot underperformed. They spend six months building something that works.
What To Do Next
The most useful first step for most organisations is an honest assessment of where their data actually stands — not where the documentation says it is, but where it lives in practice. This means examining data completeness, system connectivity, definition consistency, and governance maturity across the platforms the business actually depends on.
This is exactly what our Data Readiness Assessment is designed to do — in a fixed timeframe, at a fixed price, without the open-ended engagement that traditional consulting typically involves. In two to three weeks, we identify the specific gaps between your current data state and what your AI initiatives actually require, and we give you a prioritised roadmap you can act on immediately.
If any of the six signs in this article felt familiar, it is worth having a conversation.
Cypher Agency is a boutique data and integration engineering firm helping mid-sized businesses build reliable, governed data and integration environments — without the cost of building an internal team.
Comments