In ship design, the most dangerous promise AI can make is that it will "do the calculation for you." A ship is not a problem with a single generated answer; it is a system in which countless assumptions, standards, revisions, and lines of accountability are bound together. AI becomes genuinely useful not by hiding the calculation, but by opening up its basis, explaining how far a change propagates, and surfacing the connections engineers are most likely to miss.

Engineers Have Rational Reasons to Distrust AI

In weight and stability, structure, resistance and propulsion, piping and electrical design, a small difference in input can produce a large difference in outcome. If a unit is read as kilograms instead of tonnes, if the coordinate datum is swapped between AP and FP, or if sea water and fresh water densities are confused, the calculation can complete cleanly and still reach the wrong conclusion.

Generative AI is excellent at reading context and explaining it, but it produces sentences probabilistically. It may appear to cite formulas and standards precisely while in fact omitting or blending them. Handing final responsibility for a calculation to an LLM is therefore closer to concealing uncertainty than to automating work.

Trust in engineering AI is decided not by how fluent the answer is, but by whether the answer can be traced back to the source data and the calculation path.

Draw a Clear Line Between AI and the Calculation Engine

In a safe architecture, calculation and analysis belong to a deterministic engine. Excel formulas, validated Python functions, dedicated stability and structural programs, and unit and rule checkers all produce the same result from the same input. AI connects what comes before and after. It locates candidate inputs in drawings and datasheets, maps calculation sheet cells to their meaning, and explains differences in results along with the evidence behind them.

This division of roles does not constrain AI; it widens where AI can be applied. Without ever producing a calculated value itself, an LLM can find the relevant values across hundreds of documents, reconcile differences in notation, and explain complex calculation dependencies in the language a reviewer actually uses. It removes the checking and cross-referencing that consumes most of an engineer's time.

Deterministic Calculation and Rule Engine

Produces the same result from the same input, with standards and units under explicit control.

Runs weight, stability, capacity, and strength calculations through approved formulas and modulesChecks coordinate systems, units, signs, mandatory inputs, and allowable rangesValidates engine changes against reference cases and regression testsLocks the calculation state to failed whenever there is an error, an omission, or an out-of-range value

LLM and VLM Layer

Reads the context of documents and drawings, then connects evidence and change impact in a form people can follow.

Finds candidate objects and input values in GA drawings, 3D models, and vendor documentsLinks naming, tag, and abbreviation differences for the same equipment into one relationshipExplains the cause of a change in result, in the order of input, formula, and revisionPoints to where the relevant class rules and internal guidelines can be foundTurns uncertain items into questions and review tasks
Figure 1. When AI takes on explanation and linkage while the calculation engine owns the numbers, automation and verification become possible at the same time.

Every Answer Must Be an Evidence Package

A good AI answer does not end with a single concluding paragraph. It must also state which object the answer concerns, which document and revision the value came from, which calculation sheet and formula were used, what changed relative to the previous result, and which conflicts remain unresolved.

Call this an Evidence Package. At minimum, an answer should carry the coordinates of the source region, the document revision, the object ID, the value and its unit, the calculation run ID, the applied rules, the confidence state, and whether approval is required. The user does not trust the AI's prose; the user verifies it directly through this package.

Does the P-2101 weight change breach a stability criterion?

Under the current loading condition, the approved limits for GM and the GZ curve are still met. However, the rise in VCG reduced the minimum margin for Damage Condition 04 from 6.2% to 3.8%. The calculation was re-run on Stability Engine v4.8; before class submission, sign-off from the stability engineer and an update of the lightship baseline are required.
Vendor Datasheet Rev.03P-2101 · Weight / C.G.
Lightship Baseline Rev.CWeight Item W-8831
Stability Run STB-2026-071Damage Condition 04
Class Rule SetProject-approved rule revision
Figure 2. Only when the conclusion, the numbers, the source evidence, the calculation run, and the approval state arrive as one bundle does an AI answer translate into an actual review action.

Fail-closed: A System That Stops When It Does Not Know

In industrial AI, the most important capability is sometimes not accuracy but knowing when to stop. If tags do not match, units are ambiguous, document revisions are mixed, or an input falls outside the allowable range, the system must not fill the blank with an estimate. It must leave the calculation pending review rather than marking it "complete."

Fail-closed does not simply mean a conservative UX. It requires classifying the error type precisely and offering the user a next action. Rather than "please check the weight unit," it should say "the vendor document uses kg and the weight calculation sheet uses t, and the conversion has not been approved." Stopping must be designed as the starting point of problem-solving.

01Evidence CollectionLinks objects, documents, tables, and cells, and confirms whether the revision is current.
02Consistency CheckFinds conflicts across tags, units, coordinates, states, and the governing source.
03Calculation SandboxRuns the calculation with the changed inputs without touching the official baseline.
04Quality GateMoves to review only when mandatory inputs, rules, and regression results all pass.
05Human ApprovalAn authorized reviewer confirms the impact and evidence, then approves the official revision.
06Audit and RollbackKeeps inputs, engine, results, approvals, and application history reproducible.
Figure 3. Automation should be a controlled pipeline passing through several quality gates, not a single button that produces a result.

What AI Should Be Best At: Explaining Why Something Changed

The questions that arise in practice are less "what is the result" than "why did it change." When lightship weight increases by 12.4 tonnes, you can only act on it if you know whether it came from confirmed vendor values for three pieces of equipment, from an increase in structural quantities, or from a revised margin policy. Comparing totals is not enough.

AI can follow the calculation dependency graph and the document revisions to explain each contribution to the change. It ranks the changed inputs by significance and shows which calculations were affected and which documents have yet to be updated. The explanation should read as natural language, but every sentence must lead back, on click, to the actual cell and drawing region.

Main Engine Package+5.2 t from confirmed vendor values
Accommodation Outfit+2.7 t from detailed quantities
Cable Estimate+1.9 t from confirmed routing
Lightship weight +12.4 tBaseline Rev.C → Rev.D
Design Margin-1.3 t from the stage transition
Tank Coating+0.8 t from a revised area
Foundation Steel+2.1 t from reinforced support structure
Unmapped Items+1.0 t with no confirmed source
Affected Documents2 stability · 1 estimate · 4 reports
Figure 4. Change Intelligence does not stop at the total delta; it separates each contribution from the unresolved items so decisions can be made.

A Good Copilot Screen Looks More Like a Workbench Than a Chat Window

Build a ship design copilot as a single chat window and you multiply context switching, because the engineer must read the answer and then reopen the drawings and Excel files to check it. A better screen places drawings and 3D objects, calculation sheets, the evidence panel, and the change history on one workbench.

When a question is entered, the AI opens the relevant objects and calculation sheets and highlights the evidence regions. The user approves or corrects the inputs, runs a sandbox calculation, and compares against the previous revision. The approve button becomes active only after every check has passed. Conversation is one part of the interface; the center of gravity is linked evidence and work state.

Chat-first Tool

A plausible answer arrives in the chat window
The user goes back to the drawings and Excel to verify it
The conversation stays detached from the official revision
Hard to record who applied what, and on what basis
→

Evidence-first Workbench

The question links to objects, documents, and calculation runs
Evidence regions and input cells are visible on one screen
Before and after results are compared against quality gates
Approval, application, and rollback stay in the project history
Figure 5. An industrial copilot should resemble a review workbench where evidence is confirmed and work is completed, rather than a chat tool that generates answers.

Accuracy Alone Cannot Explain Trust

Describing product performance as "95% accurate" hides the real risk. Equipment tag recognition, table structure parsing, unit interpretation, object matching, calculation results, and criteria verdicts all have different error characteristics. Some items may be good enough for automatic approval; for others, a single error carries enough impact that a person must review it.

Risk grade and automation level therefore have to be set item by item. Low-risk, high-confidence items can be candidates for automatic application, while high-risk items must require review no matter how confident the AI is. Records of reviewers correcting AI suggestions should feed back into training and rule improvement, with model versions and evaluation criteria managed alongside them.

AI Confidence Is Not the Same as Business Approval

Even at 98% model confidence, high-impact items such as the stability baseline or values submitted to class still need an approver. Conversely, a low-risk formatting error can be corrected automatically at around 90% confidence once it passes the rule check. The scope of automation should be set by business risk, not by the model score.

The Path to a Verified Ship Calculation Workbench

The first step is not making AI explain the calculation sheet. It is the foundation: executing the original Excel and reproducing its results, mapping objects and cells to their meaning, and managing sources and revisions. Only then do you attach AI that proposes input values from documents and models and explains change impact with evidence. Limited edits and approval automation come last.

0Read-only EvidenceLinks documents, models, and Excel to show where a value came from and whether the revision is current.
1Assisted MappingAI proposes the semantic mapping of objects, cells, and tables; people confirm it.
2Verified RecalculationRuns deterministic calculations in a sandbox and compares them against the original.
3Change IntelligenceProposes each contribution to a change along with the affected documents, calculations, and owners.
4Controlled UpdateUpdates inputs and states within an approved scope and keeps the full history.
5Portfolio LearningAnonymizes and structures review cases across projects to improve rules and templates.
Figure 6. AI capabilities must be stacked step by step on a verifiable calculation foundation. Reverse the order and the demo comes quickly, but product trust arrives late.

AI Does Not Design the Ship for You

Final responsibility for ship design rests with engineers and their organizations. The claim that AI can take on that responsibility is neither realistic nor desirable. What can be reduced, substantially, is the daily burden of searching, cross-checking, tracing, and explaining.

Good ship AI does not hide its answers. It shows which data was used, what changed, and which calculations and standards were affected. And it stops when it does not know. With that measure of humility and rigor, AI can earn the trust of the shipyard and take its place inside the real design workflow.

doAZ Point of View

The competitiveness of shipbuilding AI is decided less by model size than by how naturally the Evidence Package, deterministic calculation, fail-closed gates, the change graph, and human approval are bound into a single workflow.

YT
Youngtae Kim · CEO · doAZ

Working across construction and engineering sites, R&D, and the industrial AI business, and thinking about how to connect the knowledge scattered through drawings, documents, and calculation sheets. This series is a public record of the product philosophy and field direction doAZ is pursuing.