A modern data platform is not a newer version of the one you have
Modernization
The systems running most industrial operations were built to store readings and draw charts. They do that well. What they were never asked to do is explain what a reading means, and that is the part everything now depends on.
Three things are different now, and only one of them is technology
The questions stopped being known in advance
A historian is built around the reports someone specified when it was installed. Every question after that is a project. A modern platform holds the model of the operation instead, so a question nobody anticipated is answered from what is already there.
Software became a reader
A chart is read by someone who already knows what a tag name means. An AI agent does not know, and will fill the gap confidently. Modern means the meaning is stored beside the measurement rather than carried in the heads of whoever has been there longest.
Storage stopped being the constraint
Keeping ten years of second-level history used to be a budget argument. With columnar storage and the compression it allows, it is a decision you make once. What costs money now is modeling, not disks.
The checklist
Seven capabilities, and none of them optional
Strip away the brochure language and a modern platform has to do seven concrete things: get the data out of closed systems, keep it moving through readable flows, agree on one vocabulary, connect everything into a model that can be reasoned over, remember where every number came from, enforce who may see what, and give AI agents a safe way to work. Each is useful alone. They pay for themselves together.
Data liberation
First, get the data out of the data silos
Decades of readings sit in systems built to collect, not to share. Liberation is the unglamorous first step: copy the data out over standard protocols into storage you own, in formats anything can read. The sources keep running, nothing is migrated, and the integration work that used to gate this is no longer the expensive part. It is the quick win, and everything below builds on it.
You cannot model, govern or reason over data you cannot reach.
Data flows
Data that moves through steps you can read
Liberated data still has to be cleaned, converted and joined, continuously, not in a nightly batch someone prays over. Our flow design is inspired by Apache NiFi: data moves between small, named processors, each with declared inputs and outputs, and every hand-off is recorded as it happens. A flow you can read is a flow you can trust, extend and hand to an agent. The engine itself is being built in the open and coming soon; the technical picture describes it in depth.
The steps are not only plumbing. A processor can read a timeseries and write events: an alarm when a limit is crossed, an anomaly when the pattern breaks, a state change when a machine settles into a new condition. Those derived events land in the model attached to the asset that produced them, with lineage back to the signal window they came from, so what happened becomes as queryable as what was measured.
Defined by your engineers
A processor is a function: readings in, readings out. Your people write them for unit conversion, resampling, quality flags and business rules, in code that is tested and versioned like any other code.
Increasingly defined by agents
A transform step is contract-to-contract work, the same shape as an integration, which is exactly what agents draft well. A person reviews and merges. The flow records who wrote the step, and does not care.
Observable while it runs
Every step reports what came in, what went out and what was dropped. When a flow misbehaves, the question is which step, never which system.
Ontology
Agree what things are called, once
An ontology is the quiet agreement underneath everything else: what a pump is, what it can be connected to, which properties it must carry. Write it down once and every flow, report and agent means the same thing by the same word. Skip it, and every integration invents its own dialect, which is how a company ends up with five names for one valve.
A shared vocabulary is cheaper than a thousand translations.
Contextualization
A knowledge graph turns numbers into answers
Contextualization attaches every signal, event and document to the asset it describes, and the ontology's relationships join those assets into one connected model. That is what reasoning runs on: a question follows real relationships from symptom to cause, instead of a join across five systems that only one person knows how to write.
Reasoning is cheap when the relationships already exist.
Provenance and lineage
Every number can show where it came from
Provenance records where a value entered from: which instrument, which system, which moment. Lineage records every step that shaped it afterwards, and because the flows record hand-offs as they happen, neither is reconstructed after the fact. When a regulator, an auditor or your own engineer asks why a figure looks wrong, the answer is a walk backwards through the model, not an archaeology project. It is also what tells an agent whether its own work held.
A number that cannot show its inputs is an opinion.
Data governance
Who may see what, decided in one place
Governance is not a binder on a shelf; it is enforced at the door. Every request carries the identity of whoever is asking, person or agent, and policy decides what that identity may read, what it may change, and what it will never see. A denial is logged the same as an answer, which is what makes opening the data safe rather than brave.
One door, one set of rules, whoever knocks.
AI agents
Agents do the work. Governance draws the fence.
This is what the other six capabilities were for. Agents that read the same model as your engineers can take on real work, alone or orchestrated as a fleet, because the platform already knows what they may touch and when to stop.
They work from the model
An agent's context is the graph, the ontology and the lineage, the same picture your engineers see, through the same door. It does not guess what a tag means, it looks.
The fence is policy
What an agent may read, what it may change, and where it must stop and ask a person is written as policy and enforced by the platform, not by hoping a prompt holds.
Their work is on the record
Everything an agent writes lands with provenance attached, so its work is reviewable like any colleague's, and reversible when review says no.
Deliberately science fiction
Three postcards from the near future
None of this has happened. Each story is fiction on purpose, built only from the seven capabilities above, and worth a minute exactly because it sounds like a shift report. If one of them bothers you, the useful question is which capability it would fail without.
The pump that never made the morning meeting
03:12. An agent watching vibration trends flags a bearing signature on P-204. It walks the graph: P-204 feeds the separator train, the spare was overhauled in March, the seal kit is in stock. It drafts the work order, attaches the trend and its lineage, and books the lifting crew's first free slot. At 07:04 the supervisor reads for three minutes and approves. The pump never failed, and nobody was woken up.
The audit that fit in an afternoon
Two years after the fact, a regulator questions a reported emissions figure. An agent replays its lineage: the sensors it came from, the calibration that drifted that spring, the correction applied in the June flow change. The figure turns out to be off by 0.4 percent. The evidence bundle, every step with provenance, is filed the same afternoon, together with the corrected number. The meeting about it is short.
The question that made the model smarter
A new engineer asks, in plain words, why line 3 uses more steam every October. An agent reasons across ten years of liberated history, finds the seasonal product change nobody had connected to the steam bill, and proposes a new derived signal so the pattern stays visible: a processor, drafted and tested, waiting in review. A senior engineer reads the function, corrects one assumption, and merges it. The question became part of the platform.
Nothing above needs new physics. Every piece is one of the seven capabilities, pointed at the same model.
Seven capabilities, one argument
They are not a menu. Liberation feeds the flows, the flows fill the model, the ontology gives it one vocabulary, the graph makes it answer, lineage makes it defensible, governance makes it safe to open, and agents are why all of it is suddenly urgent. A platform with five of the seven is a platform with a hole in it, and the first fleet of agents you run will find it.
Most of it is not replacing anything
Your control systems are not the problem and should not be touched. Neither, usually, is your historian: it is a good archive and it can stay one. What is missing sits above them, the layer that says what an asset is and how the measurements relate to it. Adding that is a smaller job than a migration, and it can be undone, which a migration cannot.
Common questions
Asked and answered
What makes a data platform modern?
Not its release date. A platform is modern when a question nobody predicted, asked by a person or an AI agent, can be answered from what is already there. Concretely that takes seven capabilities: liberated data, readable flows, a shared ontology, a connected knowledge graph, lineage on every number, governance enforced at the door, and a safe, fenced way for agents to work.
What is data liberation?
Copying data out of closed systems, over standard protocols, into storage you own and formats anything can read. The source systems keep running and nothing is migrated, which is why it is the first step rather than a project to fear: it is additive, reversible, and everything else builds on it.
What is the difference between data provenance and data lineage?
Provenance is where a value entered from: which instrument, which system, which moment. Lineage is everything that happened to it afterwards: each processing step between the source and the number in the report. Recorded together, while the data moves, they let any figure show its inputs on demand.
Why do AI agents need data governance?
Because an agent asks with an identity, exactly like a person, and policy has to decide what that identity may read, what it may change, and where it must stop and ask. Without an enforced fence you either lock agents out, and get nothing, or open everything, and hope. With one, a denial is logged the same as an answer and every action is attributable.
Do we have to replace our historian to modernize?
No. Control systems should not be touched, and a historian is a good archive that can stay one. The modern layer sits above them: liberated copies, flows, one model, lineage and governance. Adding that layer is a smaller job than a migration, and unlike a migration it can be undone.
Not sure which part of this you already have?
Tell us what your operation runs on today and what you have been asked to answer with it. We will say honestly whether the gap is a platform or something smaller.