An AI data foundation is the governed layer underneath an agentic AI system: every signal carries its context (asset, units, validity), quality is checked continuously, any answer traces back to its source measurement, and a written boundary defines what the agent may read and never touch. In industrial operations, the data foundation decides whether an agent reaches production. The model rarely does.
Key takeaways
- Agentic AI demos work on curated samples. Production fails on missing context, poor data quality and undefined access.
- Gartner forecasts over 40 percent of agentic AI projects will be cancelled by the end of 2027, and says organisations will abandon 60 percent of AI projects unsupported by AI-ready data through 2026.
- On a plant, four things carry the weight: an asset model, quality signals, lineage and a hard access boundary.
- Start read-only and advisory. Widen the boundary deliberately, never by drift.
Why do agentic AI pilots work in the demo and fail in production?
A typical story in an Australian industrial business: a team builds an agent that answers production questions, pulls figures from reports and drafts shift summaries. The demo impresses. Then someone asks whether it can be trusted on real decisions, with real money and safety on the line. Months later the honest answer is no. The model reasoned well enough. The data could not carry the weight.
The research points the same way. Gartner (June 2025) forecast that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear value and inadequate risk controls. In February 2025 Gartner also said organisations will abandon 60 percent of AI projects unsupported by AI-ready data through 2026.
| In the demo | In production | Impact | |
|---|---|---|---|
| Data scope | A clean, curated sample | Everything, including the messy tags | Governance |
| Context | Explained by the person demoing | Must live in the data model | Semantics |
| Quality | Known-good period chosen | Gaps, drift, contradictions | Quality gates |
| Freshness | Static snapshot | Live data, measured in minutes | Pipelines |
| Actions | Read-only, watched | May act, unwatched | Hard boundary |
What breaks when the data foundation is weak?
Three foundation problems cause most production failures.
Context. A model handed a number with no idea what it measures, its units, when it was recorded or whether the sensor was healthy will use it confidently and wrongly. A tag called FIC_101.PV means nothing without the asset context that says it is feed flow to mill two, in tonnes per hour, valid only while the mill runs. An operator knows this. An agent knows only what the data model tells it.
Quality. Agents act on what they read. Stale, contradictory or silently backfilled data passes every flaw to the agent, which then acts at speed. The familiar problem of SCADA, historian and ERP numbers that do not reconcile becomes far more dangerous when an agent decides off whichever number it read first.
Access and boundary. An agent that answers questions is useful and low risk. An agent that can act needs a hard boundary, and on operational systems that boundary is a safety requirement.
None of these is fixed by choosing a different model.
What does an AI data foundation mean for operations?
For an operational business, AI-ready data is more demanding than the reporting-grade data most sites already have.
Reporting data can be a day or a week old, reconciled monthly and understood by a few specialists. Data an agent acts on must be governed at the asset level, carry its own meaning, and be quality-checked in hours or minutes, because the agent acts at that cadence.
Four elements carry most of the weight on a plant:
- Tag naming and an asset model. Every signal knows what it is and what it belongs to. This is why PI Asset Framework templates matter: context modelled once and applied everywhere.
- Data quality signals. The agent can tell when a sensor is unhealthy and decline to act on it.
- Lineage. Any answer traces back to its source measurement, which makes the output auditable.
- A defined access boundary. What the agent may read, and what it may never touch, is written down and enforced.
Which of these can you answer today?
- Every tag knows its asset, units and validityContext: ready
- You can tell when a signal is bad in real timeQuality: ready
- Any figure traces to its source measurementLineage: ready
- None of these reliablyFix the foundation first
If most answers are no, the foundation is the next step and the agent waits behind it.
Should an operational AI agent be read-only?
For the first generation, yes. The pressure to let agents act arrives quickly once they prove useful, so decide this early.
A read-only agent that reads plant data and surfaces what a person should look at delivers most of the value with little of the risk. It can spot a correlation someone missed, draft the shift report, or flag a figure that looks wrong before it reaches the board. A person acts on each.
Once an agent can change a setpoint, acknowledge an alarm or write to a control system, the risk profile changes completely. On a plant, a confident wrong action costs equipment, production and potentially safety. The principles in AI agent governance and human override apply with a physical failure mode.
What should you fix first in your data foundation?
The path from demo to trusted agent is a data and governance sequence.
- ScopePick one job the agent should do, and its data
- ModelGive the data context: assets, units, validity
- AssureAdd quality signals and lineage the agent can read
- BoundDefine read-only access and the human override
- DeployRun advisory first, measure, then widen carefully
Start narrow. One well-defined job on well-understood data, with a read-only boundary, survives production. The all-purpose plant assistant impresses in the demo and gets cancelled a year later. Connecting the agent to governed data is mostly system integration work.
How long does it take to build an AI data foundation?
Months, and most of the value arrives before the agent does, because governed, contextualised data improves every report and dashboard the site already runs.
- Weeks 1-3AssessPick the target job, audit the data it needs against readiness
- Weeks 4-8ModelBuild the asset context, units and validity into the data
- Weeks 9-14AssureAdd quality gates, lineage and the access boundary
- OngoingDeployAdvisory agent first, measured, boundary widened deliberately
The assessment alone is worth doing before committing to an agent. It tells you whether your data can carry an agent now or whether a governance project comes first.
- Agentic AI projects Gartner expects cancelled by end of 2027>40%
- AI projects unsupported by AI-ready data Gartner expects abandoned through 202660%
- The variable that separates themThe data foundation
Spend the first portion of any agent budget on the foundation. That spend pays for itself even without the agent. Boards asking what AI advisory should tell them increasingly hear that readiness comes before tooling.
How Solve8 can help
Our consultants bring industrial data, historian and enterprise systems experience gained across their careers, including engagements at BHP, Rio Tinto, Senex Energy and Schneider Electric. We assess your data against what an agent needs before anyone proposes a build. See our mining, energy and manufacturing data work and AI strategy service, or talk to us.
Related reading:
