AI Accuracy Benchmarks: Text vs. Tables vs. Technical Drawings

Standard AI is more accurate on plain text than on tables or drawings. And the accuracy gap is bigger than most buyers realize. This post breaks down the benchmark numbers by content type and explains what they mean for enterprise AI accuracy.

TL;DR: Standard RAG and Copilot-class AI systems achieve roughly 80–85% accuracy on plain text, ~60% on tables, and ~35% on technical drawings. However, that collapse stems back to the architecture, and it maps directly to cost. For a team handling 400 visual support tickets a month, the difference between 60% and 95% accuracy works out to roughly $48,000 a year. The figures here draw on AECV-Bench 2026 (UNSW), Docopilot (CVPR 2025, Peking University), and an ACL 2025 table-understanding benchmark, and they explain why a single headline “AI accuracy” number is one of the most misleading things a vendor can show you.

So when you ask how accurate an AI system is, the honest answer is a question: accurate on what? AI accuracy is not a single number. It should be at least three, because real documentation is not one kind of content. The same PDF that is comprised of text paragraphs also holds compatibility tables, performance curves, and annotated drawings. Standard systems perform very differently on each. If you take the average of them together, you’ll get a number that describes nothing your team will actually experience.

This post pulls the benchmark data apart by content type, shows where and why accuracy falls off a cliff, and translates the gap into the cost to the buyer.

Why benchmark accuracy by content type?

A single “accuracy %” figure that most vendors quote is usually measured on text because that’s the content where systems do best.

Enterprise documentation is rarely uniform. A single maintenance manual might combine prose, spec tables, wiring diagrams, and schematics, sometimes all on the same page. And the questions your team asks aren’t neatly inside the text. They want the force value labeled on a drawing, the correct cell in a three-condition compatibility matrix, or a number that can only be read off a curve.

An AI system can score 85% on a text-heavy benchmark but only 35% on the questions your technicians actually ask. The headline number isn’t exactly lying. It’s just answering a different question than the one you truly care about. The only useful accuracy figure is the one broken out by the kind of content your business depends on.

How accurate are LLMs on plain text? (~85%)

On plain text, modern AI systems are genuinely strong. Standard retrieval-augmented generation (RAG) on FAQ-style questions clusters in the 80–85% range across the major platforms, including Microsoft Copilot, Gemini, and ChatGPT.

This is the part the field has solved well. Text chunking is mature, vector search over text is well understood, and semantic retrieval reliably surfaces the right passage. If your knowledge base is clean, structured prose, standard AI tooling will serve you most of the time.

The catch is the handoff. Accuracy on text drops the moment the words point to something visual: “see Figure 3.2,” “refer to the wiring diagram,” or the “values per the table below.” The response comes back fine, but the thing it points to doesn’t. And in technical documentation, cross-references are everywhere.

How accurate are LLMs on tables? (~60%)

Tables are where the first real drop appears in AI accuracy. On cell-lookup queries against genuine product-compatibility tables, standard ingestion lands around 60%.

The failure mode is structural. When a table is ingested as text, a row like Pump HP-200 + Motor XZ-3000 + 70°C gets flattened into a stream of tokens. The relationship that gives those values meaning is destroyed in the process. The system retains the words but loses the grid.

An ACL 2025 table-understanding benchmark captured this well. Even when the underlying data is fully structured, standard LLMs produce table outputs that are disorganized and hard to trace back to the source. There’s a difference between formatting something that looks like a table and returning the correct cell content for a specific combination of conditions. That difference is exactly where accuracy declines. The more a question depends on filtering across several conditions at once, the worse it gets.

How accurate are LLMs on technical drawings? (~35%)

Technical drawings are where accuracy really starts to collapse. AECV-Bench 2026, a benchmark from UNSW, tested the strongest available models at the time, Gemini 3 Pro, GPT-5.2, Claude Opus 4.5, and others, on architectural and engineering drawings.

Results on AI accuracy for drawings

• 0.39 mean accuracy on door-symbol counting
• 0.34 mean accuracy on window-symbol counting
• 0.51 mean accuracy on symbol-heavy counting overall
• 0.95 accuracy on OCR within those same drawings

That last line is the one that really demonstrates this point. These models read the text on a drawing almost perfectly at 95% but then fail roughly two times in three at understanding what the drawing actually depicts. The gap between recognizing characters and reasoning about the space between them is the single most important finding in the field right now.

These are some of the best models money can buy, but on spatial tasks they are wrong more often than they are right.

Why does the accuracy collapse on visual content?

The Docopilot research (CVPR 2025, Peking University) points to two compounding root causes.

1. Fragmented retrieval. Structured documents get broken into chunks for retrieval, and chunking severs cross-referential meaning. A callout that only makes sense in relation to the component it points to is split from that component. The relationships between pieces don’t survive the chunking.
2. Multi-stage error accumulation. A small retrieval error early in the pipeline doesn’t stay small. It feeds the next stage, which compounds it, until the system produces a confidently wrong generation built on a quietly wrong foundation.

Underneath both is the simpler truth that visual information lives in the layout. The meaning of a drawing is in its spatial positions, its connections, and its mapping of callouts to components. OCR-first pipelines strip all of that out and keep only the characters. A force value entered as a labeled arrow or a temperature that exists solely as a point on a curve doesn’t appear in a text extraction. The system isn’t answering badly. It just never had access to the information.

Quick-reference AI accuracy table

What AI architectures perform better on visual content?

The systems that hold their accuracy on tables and drawings tend to share a few design choices. The first is native multimodal processing. They treat the page as an image rather than running an OCR-first pipeline, so layout, position, and visual annotation survive into the reasoning step. The second is hierarchical chunking, which preserves chapter, section, and figure context instead of slicing documents into context-free fragments, so cross-references stay intact. The third is citation-based generation with confidence thresholds. This means every answer is tied to a source, and the system declines to answer when confidence is too low.

This is exactly the pattern that shows up in enterprise testing, too. An independent March 2026 Tolly Group benchmark scored a purpose-built multimodal system at 96% on complex industrial documentation, while standard systems running on the same frontier models scored as low as 26%. This gap is driven entirely by architecture, not by model choice. The detail can be found in octonomy’s 2026 report.

What this means for enterprise buyers

There are three things to take into any AI evaluation:

• Don’t accept a single accuracy number from the vendor. Request the breakdown by content type. A vendor that only shows you a blended figure is showing you their best case, not yours.
• Run a proof-of-concept with your own documentation that’s deliberately weighted toward your hardest content, including drawings, tables, and questions whose answers span several documents. The performance there tells you more than any pitch.
• Quantify the gap. Every wrong answer on a visual ticket is an escalation, a re-work, or a site visit. At 400 visual tickets a month, moving from 60% to 95% accuracy is worth roughly $48,000 a year.

Closing

Here are the three AI accuracy benchmarks one more time: standard systems hit roughly 85% on text, 60% on tables, and 35% on technical drawings. The collapse comes back to whether the architecture around the model can see the parts of your documentation that aren’t text, not the model itself. Most of your business knowledge lives in that type of documentation.

If you want to see what closing that gap looks like on real industrial documentation, the full benchmark and the architecture behind the 96% result are in octonomy’s 2026 AI Accuracy report.