A pattern is becoming hard to ignore at leadership level.
Most organisations have now invested heavily in AI. Agent frameworks, large language models, retrieval systems and internal copilots are everywhere. The ambition is high. The budgets are real. And yet, when senior leaders look at actual business outcomes, the results are still underwhelming.
The models work well in demos. The agents look promising in controlled tests. The moment they move into production, reliability drops, compliance concerns appear, and the expected ROI fails to materialise.
Across our engagements with leadership teams in fintech, healthcare, retail and high-growth SaaS, the same root cause keeps surfacing. The biggest constraint on AI success in 2026 isn't the quality of the models or the sophistication of the agents. It's the state of the underlying data foundations.
This piece explains why, what good foundations actually look like at enterprise scale, and why fixing them has become one of the highest-leverage moves leadership teams can make this year.
The 2026 reality check
AI agents place far greater demands on data than traditional analytics ever did. They need fresh, consistent inputs. They need clear lineage and auditability. They have to operate safely without putting production systems at risk. And they need predictable cost as usage scales.
The numbers tell the story:
These aren't isolated cases. RAND's research into the root causes of AI project failure points to the same conclusion: weak foundations — data, infrastructure and problem framing — are the primary reason most AI initiatives never reach production or deliver meaningful return.
Why traditional approaches don't cut it any more
The shift in 2026 is straightforward.
Traditional data warehouses are too rigid and too expensive for the scale and flexibility AI workloads demand. Older data lake patterns created their own problems — fragmentation, weak governance, and unpredictable performance that quietly degrades as the estate grows. Many teams now call this state "metadata hell": millions of tiny files and bloated metadata silently dragging every query down.
The pragmatic standard that has emerged is the open lakehouse, often built on Apache Iceberg. It combines the cost advantages of object storage with the reliability and governance that serious AI work demands.
What AI-ready foundations actually look like
The capabilities that matter aren't exotic. They're a small, sober list — and they need to be in place before any AI programme tries to scale beyond pilots.
Governance sits on top because it's where most programmes break: data contracts in version control, enforceable quality rules, and access policies enforced at query time. Time travel and branching let you reproduce yesterday's model output for an audit and roll back bad data without panic. Observability — freshness, volume and schema drift — turns silent failure into a paged incident. Automated maintenance keeps Iceberg tables compacted and metadata sane so cloud bills don't quietly balloon — AWS's prescriptive guidance for Iceberg is a good blueprint for what this looks like in practice. And interoperability means the same governed tables are usable from any compute engine, so you're not locked into a single vendor's roadmap.
The business case for fixing foundations now
The cost of weak foundations is rising. We routinely see AI programmes stalled because the data is too inconsistent for an agent to trust. Compliance and risk teams block initiatives because audit trails are missing. Cloud bills creep upward because no one is managing compaction and metadata growth.
The other side of the ledger is just as concrete. Organisations that invest in proper lakehouse foundations see fewer data incidents, materially faster time-to-value on new AI use cases, and lower long-term infrastructure costs once maintenance is automated. They also gain the one thing senior leaders actually need: confidence — from regulators, from internal stakeholders, and from the board.
Where leaders go wrong
Across recent engagements, four mistakes keep recurring:
- Treating data platform work as pure engineering execution rather than strategic infrastructure. The platform decision shapes what AI you can ship for the next five years.
- Choosing tools by vendor relationship or hype instead of fitness for purpose. The right answer for a regulated fintech is rarely the right answer for a high-growth SaaS.
- Underestimating the operational work required to run a lakehouse reliably at scale. Compaction, schema evolution, governance — none of it is glamorous, all of it is load-bearing.
- Launching AI use cases before foundations are stable. This is the most expensive mistake on the list, because the rework lands on top of customer commitments.
A leader's decision framework
If you're trying to decide whether lakehouse foundations belong on your 2026 roadmap, four questions cut through it:
- Can our current data platform support reliable, auditable AI agents at scale?
- Do we have version-controlled, enforceable contracts on our most critical data domains?
- Are we spending more time fighting data issues than building new AI capability?
- Would our risk, compliance or audit teams be comfortable with AI agents operating on the data we have today?
If the honest answer to more than one of these is "no" or "not really", lakehouse foundations are no longer a "next year" decision.
Where this goes
In 2026, competitive advantage in AI is shifting. It's no longer mostly about who has access to the best models — those advantages compress every quarter. It's about who has built the most trustworthy, governable and economically sustainable data foundation underneath them.
The organisations that recognise this now will move faster, with greater confidence, and at lower cost. The ones that keep treating data as an afterthought will keep watching their AI investments under-deliver.
If your AI strategy feels slower, riskier or more expensive than it should be, the problem is almost certainly not the AI. It's the data layer.
Let's fix the foundation.
Share this article
