Ask an operations director at a refinery what they want from AI and the answer is rarely a chatbot. It is one trusted view of the plant: what is happening now, what changed overnight, where the margin is leaking, and what the next shift needs to know. A conversational interface is useful only if the numbers behind it are right.
The trap: letting the language model do the maths
Large language models are excellent at understanding questions and summarising answers. They are not reliable calculators, and they do not know your plant's tag naming conventions, unit conversions or mass-balance rules. A system that lets the model infer a throughput figure or an inventory position will eventually produce a confident, wrong number — and in a refinery, one wrong number destroys trust in the whole tool.
The design principle we insist on is simple: every number is computed, never inferred by the language model. The model's job is to understand the question, route it to the right calculation and explain the result.
Layer 1 — the data backbone
An industrial twin starts with connections to the systems that already hold the truth: process historians, laboratory systems, ERP, maintenance management, planning and scheduling tools. At the petrochemical complexes where dDriven's platform runs, that meant more than twenty IT and OT sources. The backbone reconciles tags, units and time bases so that a barrel means the same thing everywhere.
Layer 2 — the deterministic twin
On top of the backbone sits a model of the operation: units, streams, tanks, constraints and the business rules that connect them. This is where KPIs are computed, mass balances are closed and deviations are detected. It is deterministic and auditable — an engineer can trace any figure back to its source tags.
Layer 3 — reasoning
Refinery-GPT is a reasoning layer over the twin. An operator asks, in plain language, why a unit's yield dropped yesterday. The layer identifies the relevant calculations, retrieves computed values, and assembles an explanation with links to the underlying data. It cannot invent a number because it never produces one.
Layer 4 — agents with human sign-off
Agents automate recurring operational work: reconciling inventory across tanks and systems, drafting shift handover notes, flagging exceptions for the morning meeting. Each output is a draft for a named person to review and approve. The agent saves hours; the human keeps authority.
What it takes to build
The common assumption is that a twin of a large refinery requires a large programme. At MRPL, an enterprise digital twin for a 15 MMTPA refinery was built by a team of five in six months (Partner-reported). That pace comes from a no-code platform with industrial connectors and templates, rather than custom engineering for every unit.
How to measure it
Across dDriven deployments, operational performance improvements of 10–15% have been reported (Partner-reported). For your plant, the pilot should set specific targets against your own baseline: hours spent on reconciliation, time to produce daily reports, deviations caught before they cost margin.
Questions to ask any vendor
- Which numbers does the language model compute, and which does it retrieve?
- Can an engineer trace any figure back to source tags?
- Which historians and ERPs are supported out of the box?
- How are agent outputs reviewed and approved?
- Where is the data hosted, and does any of it leave the site?
Get clear answers to those five questions and you will know whether you are buying a demo or an operational system.