The Investment Surge and the Accountability Gap
Across manufacturing, energy, and construction, AI adoption has accelerated sharply over the last three years. Boards are applying pressure. Budgets are being allocated. Pilots are launching across procurement, operations, quality control, and predictive maintenance.
Yet the 2026 Grant Thornton AI Impact Survey of 950 senior business leaders across ten industries reveals a widening accountability gap. Despite significant AI investment, the majority of organizations cannot demonstrate that their AI systems are working, cannot measure the impact of individual initiatives, and are not prepared to respond when something fails.
- 78% of business executives lack full confidence they could pass an independent AI governance audit within 90 days
- 22% of operations leaders report having a fully developed and implemented AI strategy — despite AI already running in their operations
The survey's conclusion is pointed: organizations are expanding AI across more pilots, use cases, and functions — but without consistent measurement, feedback loops, or clarity on where value is created. In manufacturing specifically, companies are responding to board pressure by deploying AI solutions before choosing use cases — a pattern Grant Thornton describes as "putting a square peg in a round hole."
Source: Grant Thornton AI Impact Survey, 2026 — n=950
Why the Process Industry Is a Harder Case Than Most
The generic AI adoption challenges above are acute across all industries. In the process and EPC industry, they are compounded by a structural characteristic that most AI deployment frameworks do not account for: a single project is never owned, designed, and executed by a single organization.
A refinery expansion, a petrochemical plant, an LNG terminal — each involves an asset owner who commissions the project, an engineering contractor who designs it, procurement firms who source equipment across global supply chains, and field construction contractors who build it. Within each of those layers, there are further subcontractors, specialist vendors, and discipline engineers, each working from their own view of the same underlying project documentation.
This fragmentation has a specific consequence for AI strategy: you cannot impose a unified data platform across organizations that do not share ownership, systems, or incentives. The push toward a fully data-driven, platform-integrated EPC industry is real — but it is a decade-long infrastructure transition, not a deployment decision.
What exists today, in every project, at every organizational boundary, is documents. Piping and instrumentation diagrams. Equipment datasheets. Instrument specifications. Vendor drawings. Inspection reports. These documents are the connective tissue across the entire project lifecycle — from FEED through commissioning — and they are almost entirely read by humans, manually, one at a time.
The Document Layer: Where Every Task Starts
Every substantive task in an EPC project begins with an engineer reading a document. Not navigating a digital twin. Not querying a structured database. Reading a document — often a PDF, often a scanned drawing, often one created by a different organization using different conventions and symbology.
This is not a temporary condition that will resolve as digitization matures. It reflects the contractual, legal, and multi-organizational reality of how capital projects are delivered. Documents are the audit trail. Documents are the handoff mechanism. Documents are what every organization can send to every other organization regardless of which software stack they use.
The research bears this out. Knowledge workers across industries spend an average of 9.3 hours per week searching and gathering information — equivalent to hiring five employees and having one spend the entire week looking for answers instead of working.
Source: McKinsey Global Institute, The Social Economy, 2012
For engineers working from multi-organization, multi-revision document sets, this figure is likely conservative.
The RFI as the clearest expression of the document problem
| Metric | Figure |
|---|---|
| RFIs per $1M of project value | 9.9 |
| Average days to close an RFI | 9.7 |
| Cost per RFI to process | ~$1,080 |
| RFIs that receive no response | 21.9% |
A Request for Information is, almost always, a question that could be answered by someone reading the right document — but the person asking either cannot find it, cannot interpret it confidently, or cannot do so quickly enough to avoid disrupting the field schedule. A crew standing idle for three days while an RFI works its way through the system represents lost productivity, potential out-of-sequence work, and material substitutions made under deadline pressure that have to be revisited later.
The Chain Reaction From Faster Document Reading
What happens when an engineer can find the right answer in 30 seconds instead of 30 minutes — with a direct link back to the exact region of the drawing or specification it came from?
RFI cycle time drops. A question that previously required formal submission, routing, and a 7–10 day wait gets answered at the point of need. Idle crew time is eliminated or dramatically reduced.
Deadline-driven shortcuts decrease. Engineers who cannot find the right specification under time pressure make judgment calls. Better information access under the same deadlines means fewer field decisions made without documentation.
Safety and reliability improve. Many near-misses and reliability issues in process plants trace back to a decision made with incomplete information. Reducing the friction of finding the correct specification reduces the frequency of this class of error.
Auditability improves without additional effort. When every answer is traceable to a source document and region, the audit trail is a byproduct of the workflow rather than a separate documentation burden.
The efficiency gain compounds across every organization on the project. Because documents are the universal handoff mechanism across organizational boundaries, improving document readability benefits every party — owner, engineer, contractor, and subcontractor — without requiring any of them to change their systems.
Why This Approach Does Not Require Trusting AI to Make Decisions
The approach described here is not AI autonomy. It is AI-assisted search with human verification — using AI the way a highly capable analyst who has read every document on the project would be used: to surface the relevant passage, identify the relevant drawing region, and hand it to the engineer for a decision.
This matters for adoption in a regulatory and safety-critical industry. The process industry does not need AI to be right by itself. It needs AI to reduce the time it takes a qualified engineer to be right. Every extracted answer should point back to its source — the exact drawing, the exact revision, the exact region — so that the engineer can confirm it in seconds rather than re-derive it from scratch.
Traceability is not a feature in this model. It is the mechanism by which the entire workflow remains human-accountable — which is the precondition for adoption in any environment where a wrong answer has safety or legal consequences.
The Technology Precondition That Now Exists
Template-based and rule-based document extraction systems for engineering drawings have been attempted repeatedly since the early 2000s. They worked in controlled conditions and failed in production: custom symbol sets, varying drawing conventions, scan quality, and project-specific tag formats each required re-training or re-templating — and a wrong extraction still required a human to re-derive the answer from scratch.
The shift that makes the current moment different is not incremental improvement in object detection. It is the availability of models that can reason about an unfamiliar symbol from context — inferring relationships rather than only matching memorized patterns:
- Which line connects to which equipment
- Whether an instrument reads the process line or carries only a signal
- Whether nearby text is a tag, a size, a specification class, or an unrelated label
This capability generalizes across drawing conventions and symbol sets in a way that template-matching cannot. Combined with source traceability — pointing every extracted fact back to the exact region of the drawing it came from — it addresses both of the failure modes that stalled earlier systems: generalization and correction cost.
What Novek-AI Is Building
Novek-AI is building relationship-aware extraction for engineering drawings — P&IDs, datasheets, and instrument specifications — with full source traceability. Our extraction model reasons about relationships between entities on a drawing, not just the entities themselves, and every output is linked back to its exact source location so that engineering review is a confirmation rather than a re-derivation.
We are publishing our accuracy results against ground truth, segmented by industry, scan quality, drawing density, and entity type — including where we are weaker — at novek.ai/accuracy.
References
- Grant Thornton, AI Impact Survey 2026, n=950 senior business leaders across 10 industries. grantthornton.com/insights/survey-reports/2026/ai-impact-survey
- McKinsey Global Institute, The Social Economy: Unlocking value and productivity through social technologies, 2012. mckinsey.com
- Navigant Construction Forum, Impact and Control of RFIs on Construction Projects, 2013. Study of 1,362 projects containing over 1 million RFIs. Note: cost figures are from 2013 and likely conservative by today's standards. procore.com/library/rfi-construction