A Big 4 firm offered its audit clients a discount if they kept their financials in a discoverable data warehouse. This way, auditors could skip the integration work and finish faster. 

Few clients took the deal. Most paid the full bill because they were not mature enough to manage their own financial data.

That story captures 30 years of enterprise IT in a nutshell. Companies pay full price for unresolved internal data problems they could fix with the right guidance. Gradually, a mountain of technical debt accumulates. AI vendors now count on most companies never finding a way to a real solution.

Take back control

Now is your chance to tackle the information siloing, communications and integration problems your organization has had for 30 years. The window of opportunity is open in a way it hasn’t been since the 1990s.

There’s a reason to make the transformation now. Semantic graphRAG and related next-generation data and knowledge management based on open standards is commercially available and in use by large enterprises such as AstraZeneca, EY, S&P Global, and Roche. I heard these and other major case studies at the Graphwise AI Summit just last week. 

Boards want to take advantage of agents and know those agents need the right contextualized data. The companies above described how others could follow their example.

Most large enterprises waste money on application software that should be spent on data architecture transformation. Enterprise software spending keeps climbing even as core technology costs fall. Between 1999 and 2025, revenue for the top seven software companies surged nearly ninefold, driven by magic bullet syndrome, complex legacy systems, application sprawl, and deeply ingrained buying habits.

The 12 enterprise software sins

Enterprise software buying habits have created a web of systemic issues. These issues stem from two primary failures: meaning gaps and solution gaps. Meaning gaps occur when software vendors define data differently from your business, forcing expensive customization and custom integration. Solution gaps arise when tools fail to address the core problem, driving teams to buy peripheral software that creates additional silos.

Together, these gaps lead directly to the 12 enterprise software sins: application sprawl, redundant functionality, fragmented data models, vendor lock-in, shadow IT, spiraling maintenance costs, data siloing, lost context, brittle integrations, over-customization, compliance risks, and wasted software licenses. Organizations repeatedly fall into these traps by treating software as a quick fix rather than addressing underlying architectural weaknesses.

Auditing lessons for data control

Reclaiming control requires a disciplined auditing approach. First, audit your software portfolio for true data utility, not just feature sets. Identify where vendor-defined models override your own operational meaning, and eliminate tools that create artificial boundaries between datasets.

Second, establish explicit, shared semantics across your data layer before adding new AI or automation tools. Auditing principles dictate that if you cannot explicitly define control boundaries and rules within the data itself, you cannot govern autonomous agents or prevent costly software waste.

Organizations routinely pay premium prices for redundant tools and bloated code bases instead of streamlining their core data models. Taking back control starts with simplifying data architecture and eliminating unnecessary software waste.

Agents make the stakes higher. Agents need to be told what not to do in no uncertain terms. Vectors and probabilities cannot do that, because they are never certain about anything. Only data that carries explicit, shared meaning can ensure agent governance.

Why so many providers claim AI will fix everything

Over 8,800 venture-funded startups were chasing the AI market by my last count. Hundreds of incumbents now claim AI capabilities too. Perhaps 300 providers might actually help you. These 300 offer hybrid, deterministic and probabilistic capabilities. They aren’t just creating probabilistic monoliths.

Why does everyone pitch the same promise? Because they think some will believe them. After all, enterprises were gullible enough before.

How we got here

In the 2000s, a handful of incumbents and a few upstart ecommerce, web search and social media companies had the capital and savvy to launch massive data farming, mobile application and public cloud operations. Today those companies still hold a stranglehold on enterprise IT. Now they push their version of AI: naive, monolithic, and dependent on wasteful data centers to supply it at scale.

Decades of passivity about information technology and in particular ignoring the data layer made this possible. Enterprise became passive about IT in the 2000s and ceded effective control of their networks, code and data.

Some may remember Nick Carr’s HBR article and the follow-on book of the same name, IT Doesn’t Matter. Carr’s advice: buy what everyone else buys. Enterprise IT is a commodity. Focus on your core business and leave IT to others, because it is not a competitive differentiator.

Most enterprises did what Carr recommended.

• Suites first. ERPs, CRMs, HRMs, SCMs, CMSes and dozens of other three-letter suites arrived. A major ERP implementation could cost tens of millions of dollars, and consulting firms gladly helped. Some organizations outsourced IT operations to consultancies entirely.
• Cloud and SaaS next. In the late 2000s, Amazon Web Services and software as a service took hold.
• Data lakes after that. In the 2010s, Big Data and massively parallel commodity server farms took over. Enterprises that had modeled their data for warehouses started dumping raw copies into data lakes. The argument: skip up-front modeling, because at least the data is accessible.

Each step handed more control to someone else. Each step also left the data less described than before.

How to avoid getting taken this time

Test every pitch against these questions.

1. Does the vendor cover the whole AI picture? Probabilistic methods get you partway. Logic-based, symbolic methods give you the certainty an auditor needs. Almost all of those 8,800 startups cover only the probabilistic half. By contrast, only a few hundred also provide the deterministic reasoning and precision description necessary to trusted agentic computing.
2. Who owns the semantic layer? Proprietary vendors want to own every layer. A BI vendor will offer you its semantic layer and ask you to keep using its software. That layer will not let agents operate across company boundaries. Open standards such as RDF will, in a precise and trustworthy way.
3. Does the platform help you add feedback loops? Enterprise search failed in the 2000s because enterprises lacked the data and the feedback loop mechanisms with humans and machines in the loop  to improve results. Today’s workplace search tools repeat that pattern unless they enrich your data over time.
4. Can the vendor show an enterprise case, not just a demo? Ask for the customer story. AstraZeneca, EY, Roche, and S&P Global with Graphwise’s help built working systems. Ask what problems appeared and where the pushback came from.

Describe the walls in your data and lock in the rules agents must follow

The recent Hugging Face breach showed what happens when agents do not know where the walls are. Agents focus so hard on the mission that they find any route to it, breaking rules along the way. Guardrails bolted on from outside resemble old perimeter security. If you do not work at the data layer, you cannot control the agent.

Do the work at the data layer. Your content, data and knowledge management teams already hold the pieces. Content teams build taxonomies. Data teams model. Knowledge teams build ontologies. Bring them together in a knowledge graph built on open standards, and describe your data well enough that agents know what they can and cannot touch.

That gets you control. It also lowers cost. A database lookup via semantic graphRAG and MCP costs pennies. You can still ask a model to interpret your question, but asking a model to guess costs dollars and may return a wrong answer.

You do not need to pick the vendor the vendor everyone else does. You need to own your data, describe it openly, choose an architecture that supports articulated data and knowledge sharing at scale, and ask harder questions than your predecessors did.

For More Information:

Carr, Nicholas G. May 2003. IT Doesn’t Matter. Harvard Business Review.

Graphwise. Graphwise AI Summit 2026, accessed October 11, 2026.

Morrison, Alan. 2026. The Hugging Face breach: a case study in why agents need grounding. The GraphRAG Curator, August 1, 2026.

Morrison, Alan. 2026. Why Executives Need Some Data Therapy to Solve their AI Problems. The GraphRAG Curator, September 17, 2026.