You Can't Build AI on a Broken Foundation

The company that did everything right - and still failed.

ClearLedger had a good run. Founded in 2013 to help mid-market businesses automate accounts payable and cash flow reconciliation, the UK-based fintech spent twelve years doing what good software businesses do: shipping features, winning logos, and layering on integrations. By 2024 it had north of 400 customers, solid net revenue retention, and a product that genuinely worked.

It had also accumulated a digital infrastructure that looked less like a modern data business and more like a geological cross-section. Layer upon layer of systems, each added at a different point in time, each solving the problem that existed at that moment, none designed to talk to each other.

This is not a story about a badly run company. It is a story about a completely normally run company. And it matters because of what happened next.

The board mandates Ai

By 2024, the pressure on ClearLedger's board was building. Enterprise boardrooms everywhere had absorbed the implications of large language models. AI-native competitors were beginning to emerge, marketing themselves on intelligence rather than automation. The board's response was swift: lean into AI. Use the twelve years of transaction data to build smarter features, faster insights, defensible differentiation. The CTO estimated six to nine months to first production.

What no one had mapped out - not properly, not honestly - was the terrain they were about to build on.

What happened when they started digging

Three AI initiatives were launched across 2024 and into 2025. Each made commercial sense. Each hit a wall within months — blocked not by the model, but by the data sitting underneath it.

Churn Prediction Model

The team needed usage, support and billing data joined by a shared customer ID. What they found: three different identifiers across three systems, with no reliable way to match them. Manual reconciliation took days and produced results nobody trusted.

AI Cash Flow Forecasting Feature

Eighteen months of clean, labelled transaction history was required. The data science team found three rounds of re-categorisation, incompatible schema versions, and no migration logic connecting them. The feature slipped nine months before anyone admitted why.

Support Triage Agent

By 2025, agentic AI was gaining real traction. ClearLedger's attempt ran straight into four years of support history with a rebuilt taxonomy, inconsistent free-text fields, and resolution notes logged in an unarchived Slack channel.

‍

By the end of 2025, the AI roadmap had delivered one internal dashboard and a great many lessons. Roughly 18 months of engineering time had been absorbed by three initiatives that each made sense individually - and shared a common root cause that no one had been given permission to address.

2026: Pressure from two directions

Then the SaaSpocalypse arrived. By early 2026, buyer behaviour had shifted. CFOs across ClearLedger's customer base had audited their SaaS estates and reached the same conclusion: too many tools, too little proof of value. Renewal conversations that had previously been rubber-stamped became adversarial. Net revenue retention started to compress.

ClearLedger now faced both pressures simultaneously. Its AI roadmap had stalled on data problems it hadn't resolved. Its customers were demanding more capability or lower cost. And the AI-native competitors it had dismissed in 2024 were beginning to win deals.

IDC estimates that companies with strong data foundations will generate twice the AI-driven revenue of those without. McKinsey consistently finds data and integration challenges - not model quality - as the primary reason enterprise AI initiatives fail to scale.

This is not unusual. It is almost universal.

ClearLedger's story is fictionalised. The pattern is not. Across industries - particularly in businesses that have scaled through a decade of SaaS accumulation - the same structural problem repeats. Data lives in silos. Systems were chosen for departmental convenience rather than enterprise coherence. Integration was always the thing that would get sorted later.

The answer is not the AI. The answer is what the AI is being asked to work with.

What 'foundation first' actually means

Saying 'fix your data' is easy. Knowing what that means in practice - and sequencing it to deliver value quickly while building for the long term - is the hard part. A genuine data foundation has three properties.

Clean

Data is accurate, consistently structured, and trustworthy. Duplicates are resolved. Schema changes are documented. Historical records are coherent. This is not a one-time exercise - it requires ongoing governance.

Connected

Records and signals can be joined reliably across systems. There is a common identity layer: one way of saying 'this customer in Salesforce is the same account in the product database.' Without it, every model works with a partial picture.

Governed

Someone owns data quality. There are clear rules about what data means and how it changes. Access controls exist. Lineage is traceable. When something goes wrong, there is a way to find it before a model fails silently.

‍

None of this is as exciting as an AI demo. All of it is what turns AI demos into AI businesses.

The opportunity inside the pressure

The SaaSpocalypse is not just a cost story for boards to manage. It is a data architecture moment dressed up as a procurement exercise. The businesses that use this moment to understand and consolidate their data estate - rather than just cutting line items - will emerge with something their AI-native competitors will take years to build: depth. Real transaction history. Real behavioural signals. A clean, connected, governed foundation.

At exit, this shows up in due diligence. It shows up in the multiple.

Data first. AI second. Outcoms always.

Where to start.

Vantage builds the data foundations that make AI work in practice, not just in pilots. We start with a 2–4 week fixed-price diagnostic - a clear picture of where you stand and what to do next, with every output an artefact you can act on.

Weeks 1-2: AI & Data Readiness Diagnostic

An 8-dimension framework scoring your maturity across strategy, data, technology and people. We identify where AI will move the needle and what data work needs to happen first.

Weeks 2-3: Strategic Roadmap

Prioritised use case shortlist with quantified ROI, a 90-day execution plan, and a data gap register. Every output is an artefact you can act on - not a deck of recommendations.

What you walk away with

A prioritised roadmap, a quantified business case, and a clear view of what to build first. Every output is something your leadership team can act on immediately.