# Draw the Line on Responsibility, Not Capability: A Law Firm and a Design Office Found the Same AI Boundary
Canonical URL: https://theunclej.com/blog/draw-the-line-on-responsibility-not-capability
Markdown URL: https://theunclej.com/blog/draw-the-line-on-responsibility-not-capability.md
Description: The judgment of this piece: the first cut of an AI project is not a technology cut — it is a cut between people and machines. That separation line looks like it is drawn on "can AI do this," but it is
Category: ai-organization
Tags: ai-organization, piece-00000176
Published: 2026-10-19T10:32:03.947Z
Updated: 2026-10-19T10:32:05.659Z

---

![og-cover](https://assets.theunclej.com/uploads/c0fc4581f290fe53220d5bddda8ea0ec.png)

# Draw the Line on Responsibility, Not Capability: A Law Firm and a Design Office Found the Same AI Boundary

> **The judgment of this piece: the first cut of an AI project is not a technology cut — it is a cut between people and machines. That separation line looks like it is drawn on "can AI do this," but it is actually drawn on "who signs when something goes wrong." Professional judgment can't be unloaded — not because machines can't carry the judgment, but because responsibility has no stand-in. Information handling can be split off, because it produces no direct external liability and the next checkpoint backstops it. How deep the line cuts is a function of liability institutions, not a function of model capability.**
>
> What this piece solves: when you scope AI for a professional role whose output exits under a name — a lawyer's report, an engineer's drawings — the question everyone asks first, "can AI do this work," is the wrong question, and it quietly misdirects the whole project. This piece replaces it with a test you can run before any build: does this work's output have to exit under a named accountability to leave the building? Using two live deployments, it takes both sides of the line apart — which half goes to the machine whole, which half keeps a human in the loop with real knobs — and shows what has to move before AI can take on one more stretch.
>
> Boundary first: both cases come from the first 24 cases of the *FDE Cases 100* collection, published September 2026 — material assembled by one interview editorial project, not 24 independent causal experiments, and every effect stops at the participant's own account. "The line is drawn on responsibility" and the "named-accountability exit test" are our cross-case decomposition framework — not the case companies' own claims, and not customer-validated results. The test's induction range and its failure conditions are stated item by item at the end.

## Two Sites, One Line

Everybody says AI takes the simple work first.

We took two cases apart and reached the opposite conclusion: **AI never takes the simple work first. It takes the work nobody signs for.**

One case sits in a law firm: a commercial litigation team handling complex matters — contract disputes, equity disputes. A single case can involve dozens of material types and often more than a thousand pages, and collation alone can eat three to five workdays (NO-03 txt L40-43, print p. 25 — the interviewed FDE's industry-observation caliber; an FDE, forward deployed engineer, is the delivery role that walks into a client's site and lands AI inside the actual business. The source gives no sample and no provenance for this figure, so we use it as background tension only).

The other case sits on a building fire-protection construction-drawings site. What the client company is called, and what kind of business it runs, the source does not say — below, we call it "the design site." One ordinary commercial project can involve facility placement for hundreds of spaces on the fire-protection sheets alone. The case's interviewed author puts it plainly: "the checking process often takes more time than the drawing itself" (NO-21 txt L46-48, print p. 211). One building can hold dozens of functionally similar spaces, the same placement routine repeated over and over — "the time truly spent on creative design is severely squeezed" (NO-21 txt L41-43, print p. 210).

One whole industry apart, two sites cut out the same line.

The evidence floor, laid on the table first:

| | Law-firm case (NO-03) | Design site (NO-21) |
| --- | --- | --- |
| Site and deliverable | Commercial litigation team; reports go to clients and into court | Building fire-protection construction drawings; sheets must pass fire acceptance |
| The pain at the neck | Dozens of material types, often 1,000+ pages; collation alone can eat 3–5 workdays (interviewed FDE's industry observation — no sample, no provenance) | Checking takes more time than drawing; creative time squeezed (the case's interviewed author, self-report) |
| System status | Live in normal operation; two months after launch, users spread across multiple cities nationwide (Figure-2 caption caliber — proves spread, not effect; no activity data in source) | Live in use; user count, project count, deployment scope all undisclosed — single-client caliber only |
| Effect evidence | No comparison detail (single-case self-reports); this piece cites no benefit figures | "About 30 minutes to a preliminary result" = feedback from the project side, relayed by the case author; preliminary, not final; no same-caliber before/after comparison |

Both cases come from the same interview editorial project — not two independent experiments. We cite structure, not report cards.

## Two Sentences, With the Reason Stated Outright

In the law-firm case, the interviewed FDE puts the bottleneck bluntly: "the capacity bottleneck of lawyers is not in professional judgment; it is in the efficiency of information handling" (NO-03 txt L52, print p. 25).

Note what he actually said: "not in professional judgment." Not that lawyers can't judge — that judgment is besieged by information handling.

Once the system was built, the division of labor was written down like this: "In our design, AI handles the bulk of information processing, analysis, organization and repetitive work; the professional system carries case context and knowledge flow; the lawyer holds final judgment, strategy choice, and professional responsibility" (NO-03 txt L110-113, print p. 27).

The design-site case draws the line harder. When the system first reached users, the expectation was "one-click generation of complete construction drawings" (NO-21 txt L142, print p. 215) — and the professional designers objected. Why? The interviewed author's own words, with the word "reason" pre-filled for you: "The reason is that fire-protection design involves liability. Designers bear responsibility for the compliance and safety of the final drawings" (NO-21 txt L144-146, print p. 215; the speaker is the case's interviewed author — their identity and commissioning relationship are undisclosed in the source, and the account is first-person "we" throughout).

Put the two sentences side by side.

Law firm: AI handles information processing; the lawyer holds final judgment and strategy choice — **and professional responsibility**.

Design site: AI generates; designers answer for the compliance and safety of the final drawings.

Two groups of people, industries apart, cut out the same line — and both hang the word "responsibility" on the human side.

One more observation, and this one is ours, not theirs. We ran a five-breakpoint scan over the design-site case: role, process, knowledge, responsibility, governance. Of the five, only "responsibility" is actively and repeatedly argued in the source — three self-accounts on print pages 213 and 215, all about it. The participants talk efficiency with their mouths and return to responsibility with their pens. **That site's organizational center of gravity is not efficiency; it is the boundary of liability.** (Our decomposition, not the case's own claim.)

That is not coincidence. That is structure. The structure we see is this: **the separation line looks like it is drawn on "can AI do this," but it is actually drawn on "who signs when something goes wrong."**

## The Blade Isn't on Capability

We first read the line as a capability split: judgment is complex, machines can't do it; information handling is simple, machines can.

Reread both cases — that reading doesn't stand.

**Professional judgment can't be unloaded. Not because machines can't carry the judgment — because responsibility has no stand-in.**

A lawyer's report goes to the client and into court; a designer's sheets must pass fire acceptance — "the position deviation of a single sprinkler head can affect fire acceptance" (NO-21 txt L152-153, print p. 215). What these deliverables share: they exit under a name. Once out, someone must answer for them — and to answer, you need a name.

Behind the name sits an entire accountability apparatus: licensure, signing authority, the capacity to compensate when things go wrong. However strong AI gets, it holds none of the three. It cannot be disbarred. It cannot sit as a defendant. So "can AI make this judgment" is a fake question. The real threshold reads: when this judgment turns out wrong, a name must own it.

Why does information handling split off? Reading, searching, transcribing, sorting, checking — none of it directly produces external liability. If a case summary skews, the next checkpoint catches it; if a sprinkler spacing is miscomputed, the designers' review-and-confirm gate still stands — "final schemes must pass the designers' review and confirmation, and the system sets manual confirmation at key nodes" (NO-21 txt L155-156, print p. 215). And the backstop is not a slogan — it needs traces to stand on. The law-firm case keeps "all analysis results, strategy suggestions and document drafts deposited in the case space, forming a traceable work record" (NO-03 txt L174-179, print pp. 30-31); that is where review has somewhere to land. When process errors have a backstop, the whole step can go to the machine. When result errors must exit under a name, they stay with people.

One more detail nails this reading down harder.

Within these two cases' texts, the only place where "capability falls short" appears is on the information-handling side: an architectural drawing is a professional carrier of spatial relations and graphic symbols, and "the general large model cannot directly understand it" (NO-21 txt L77-78, print p. 212) — so that team built its own dedicated model to parse CAD drawings (NO-21 txt L86-87, print p. 212). The capability gap was closed with engineering. The nearest thing to a counterexample anywhere else in the text: some people held doubts about the reliability of AI-generated results (NO-21 txt L161-162, print p. 216) — that is an attitude toward use, not a statement about capability; same direction, no damage to the claim.

And on the other side — why judgment stays with people — not one sentence in either case's text says "AI didn't judge well enough, so it stayed with people." The reason the source gives for keeping people is responsibility.

Before writing this, we set ourselves a falsification clause: if the source attributed the kept-for-human steps to "we tested AI first and it failed, so we kept people," this mainline would have to be torn up and rewritten. We read both cases end to end. No such passage exists — the direction runs exactly the other way: the design-site case attributes it to liability with the word "reason" spelled out, and the law-firm case writes "professional responsibility" into the division-of-labor sentence itself.

Strictness demands one narrowing: information handling is not risk-free. If AI-organized material bypasses review and exits directly, liability arises all the same. So the strict statement is — **only intermediate steps that do not directly produce external liability, and are backstopped by the next checkpoint, can be split off.** Both cases' architectures built that checkpoint in: the lawyer's final judgment; the designers' review and confirmation.

Within the material these two cases disclose, we found no counterexample: not one step where AI output exits directly under a name with nobody signing for it.

## What Sits on Each Side of the Line

The line is drawn. What sits on each side? The two cases' answers look like they copied from each other.

**Intermediate steps that carry no direct liability go to the machine, whole.**

The law-firm case's approach: "We did not change the lawyers' original way of working; we looked for where AI could embed inside the original process" (NO-03 txt L97, print p. 27). Reading a thousand pages of material, drafting case summaries, converting colloquial descriptions into legal language, mapping the relationships between pieces of evidence — these steps go to AI assistance. And they don't live in a chat box. The previous generation of tools died exactly there: lawyers switching across five or six systems, case context torn to pieces (NO-03 txt L69-72, print p. 26). They live in a case workbench: "We moved the product entry from a chat window to a case workbench. The lawyer uploads the case materials first; the system automatically builds a case knowledge graph" (NO-03 txt L127-128, print p. 28). Context sinks into a space, not into a conversation.

The design site is blunter still. Construction drawings are "highly rule-bound design work that allows no free play" (NO-21 txt L90, print p. 212), so the team converted "the code clauses into a rule engine the system can execute" (NO-21 txt L95-96, print p. 213) and let AI generate by rule. In our decomposition: checklist-type judgment moved out of human heads and into deterministic rules — the machine handles certainty, not inspiration.

**Steps that exit under a name keep a human in the loop — with real knobs in hand.**

The design site opened parameter entries for special spaces: hazard class, K-factor, sprinkler spacing, layout method — designers can adjust them (NO-21 txt L102-105, print p. 213). AI's output is positioned as a "professionally usable first draft," and "final decision authority always rests with the designers" (NO-21 txt L104-105, print p. 213). The line was written into the product as a principle: "From the very beginning we fixed one principle: AI is responsible for generation; humans are responsible for decisions" (NO-21 txt L99, print p. 213). Even AI's circle of competence was declared in advance: no "creative judgment" beyond the scope of the rules (NO-21 txt L154-155, print p. 215).

And what the machine hands back is pinned in place by the source: a first draft. The author relays the project side's feedback — for large projects, about 30 minutes to a preliminary result. State the full caliber: a relayed account, preliminary not final, no same-caliber before/after comparison (NO-21 txt L110-113, print p. 213). The work after those 30 minutes carries no number: manual adjustment, review, confirmation. Precisely that numberless stretch is the named, liable stretch.

The law-firm case mirrors it from the other end: AI drafts; the lawyer judges and answers — the very half of the division sentence that reads "the lawyer holds final judgment, strategy choice, and professional responsibility." Where the line falls follows the site's needs: "which decisions stay with people, and which AI finishes and delivers directly — all of it follows real needs" (NO-03 txt L85-86, print p. 27). The interviewed FDE's field research came back plainer still: where they want time saved, and where they wouldn't rest easy even with the time saved (NO-03 txt L82-86, print pp. 26-27). Where they don't rest easy — that is where the liability line sits.

The two ready-made templates, side by side:

| Side of the line | Law-firm case (NO-03) | Design site (NO-21) |
| --- | --- | --- |
| No direct liability, intermediate → machine, whole | Case workbench (context container): material reading, summary drafting, colloquial-to-legal conversion, evidence mapping; AI drafts | Rule engine: code clauses converted to executable rules; routine spaces auto-generated with default parameters |
| Exits under a name → human in the loop, with knobs | Lawyer: final judgment, strategy choice, professional responsibility | Designers: parameter entries (hazard class / K-factor / sprinkler spacing / layout) + manual confirmation at key nodes + review with sign-off |

One more word on those parameter entries: the human is not a rubber stamp, and the knobs must be real. If you can turn the parameters, you can carry the responsibility. If a confirmation step offers nothing but an "agree" button, that isn't keeping a human in the loop — that's keeping a human to take the fall.

An aside, on a blade we know unusually well — we come from HR. Drawing this line is, in effect, the timesheet audit that professional roles never got. The JD says judgment; the hours ledger mostly says hauling. The design site's own people reported the split before AI arrived: "70% execution work such as drawing and code checking; 30% professional judgment and scheme adjustment" (NO-21 txt L116-117, print p. 213) — seventy execution, thirty judgment. Mind the caliber: the participant's own estimate, not a measurement. The most expensive heads doing the cheapest work — that sentence is about exactly this.

Earlier pieces in this series handed out two other blades: costs split into build-time and run-time, and replacement depth sorted into tiers. This piece's blade sits further upstream than both — the first cut of a project, between people and machines.

## The Ceiling Is the Institution, Not the Model

Compress the structure above into one sentence: **how deep the separation line can cut is a function of liability institutions, not a function of model capability.**

Model capability reinvents itself yearly. The lawyer's report still needs the lawyer's signature; the construction sheets still need a designer to answer. Within what these two cases disclose, the liability line has not moved a millimeter.

So the answer to "how do we let AI take one more stretch" is not in the tech stack. Write the review nodes into the process; assign a named signatory to every deliverable that exits; settle in advance who compensates whom when an error lands. Where the liability institution moves, the line moves. If the institution doesn't move, the model can be as strong as it likes — AI stays parked at "assistant." Not because the model can't. Because the institution hasn't yielded.

One fact has to be stated just as clearly: both cases only preserved this liability line. Neither moved it. In the design-site material — whether sign-off and licensure arrangements should be redrawn once AI entered, and how an identification error splits from a judgment error — the source answers none of it. In the law-firm case — when AI misreads a material and the strategy follows the misread, how the firm pursues remedy internally — likewise no answer. So "change the liability arrangement first" remains an inference from the criterion today, not a conclusion with a live site behind it. The day some organization actually rearranges "who pays when it's wrong" — that is the day this line starts cutting deeper.

Last, the criterion you can carry out of here. When you make the first cut for any professional role's AI project, don't ask "can AI do this work." Ask —

**Does this work's output have to exit under a named accountability to leave the building? That is the named-accountability exit test.**

If yes: human in the loop, three things in place — parameter entries, manual confirmation, professional review with sign-off. The design site's rule engine plus parameter entries plus manual confirmation; the law firm's context container plus AI drafting plus lawyer sign-off — the two templates in the table above, ready to copy.

If no — if the step is intermediate and produces no direct external liability: hand it to the machine whole — context container, rule engine.

And drawing the line has a by-product: technology selection demotes itself into a downstream procurement question. Which model, buy or build — all of it comes after the line is drawn. One case states that ordering out loud: the design site wrote the division principle before the architecture (NO-21 txt L99, print p. 213) and built the architecture to fit the principle. The law-firm case never states such an ordering; its division of labor was ground out step by step against site feedback (NO-03 txt L119, print p. 28).

If you want AI to take on more, change the liability arrangement first — not the model.

## Sources and Calibers

- Case sources: the first 24 cases of Datawhale's *FDE Cases 100*, published September 2026 — write-ups from one interview editorial project (L1), not 24 independent causal experiments. This piece uses two cases only — NO-03 (commercial litigation) and NO-21 (building fire-protection construction drawings); each case's facts come only from its own source text and decomposition card; nothing is borrowed across cases.
- Quote anchors: (NO-XX txt L…) = physical line of the source-text extraction file, each one re-verified against the source txts during this draft; page numbers are print pages (print page = PDF physical page − 1). The main direct quotes:

| Direct quote | Anchor | Caliber |
| --- | --- | --- |
| "the checking process often takes more time than the drawing itself" | NO-21 txt L46-48 (p211) | The case's interviewed author, self-report |
| "the time truly spent on creative design is severely squeezed" | NO-21 txt L41-43 (p210) | The case's interviewed author, self-report |
| "the capacity bottleneck of lawyers is not in professional judgment; it is in the efficiency of information handling" | NO-03 txt L52 (p25) | Interviewed FDE, self-report; opening tension only |
| "In our design, AI handles the bulk of information processing … the lawyer holds final judgment, strategy choice, and professional responsibility" | NO-03 txt L110-113 (p27) | Division-of-labor sentence; spans the p27/28 footer, pinned to p27 |
| "one-click generation of complete construction drawings" | NO-21 txt L142 (p215) | Users' expectation, in the source's words |
| "The reason is that fire-protection design involves liability. Designers bear responsibility for the compliance and safety of the final drawings" | NO-21 txt L144-146 (p215) | Speaker is the case's interviewed author; identity and commissioning relationship undisclosed |
| "the position deviation of a single sprinkler head can affect fire acceptance" | NO-21 txt L152-153 (p215) | Named-accountability background sentence |
| "final schemes must pass the designers' review and confirmation, and the system sets manual confirmation at key nodes" | NO-21 txt L155-156 (p215) | Review-backstop face |
| "the general large model cannot directly understand it" | NO-21 txt L77-78 (p212) | Capability statement, information-handling side |
| Built a dedicated model to parse CAD drawings (paraphrase) | NO-21 txt L86-87 (p212) | Paraphrase, same direction as source |
| "We did not change the lawyers' original way of working; we looked for where AI could embed inside the original process" | NO-03 txt L97 (p27) | Division form |
| "lawyers switching across five or six systems, case context torn to pieces" | NO-03 txt L69-72 (p26) | Why the previous tool generation failed |
| "We moved the product entry from a chat window to a case workbench. The lawyer uploads the case materials first; the system automatically builds a case knowledge graph" | NO-03 txt L127-128 (p28) | Context-container form |
| "highly rule-bound design work that allows no free play" | NO-21 txt L90 (p212) | The case's interviewed author, self-report |
| "converted the code clauses into a rule engine the system can execute" | NO-21 txt L95-96 (p213) | This piece takes the division-form face only |
| "From the very beginning we fixed one principle: AI is responsible for generation; humans are responsible for decisions" | NO-21 txt L99 (p213) | Product principle |
| "professionally usable first draft"; "final decision authority always rests with the designers" | NO-21 txt L104-105 (p213) | Output positioning and human-side decision authority |
| About 30 minutes to a preliminary result on large projects | NO-21 txt L110-113 (p213) | Relayed from the project side; preliminary, not final; no same-caliber before/after comparison |
| No "creative judgment" beyond the scope of the rules | NO-21 txt L154-155 (p215) | Declared circle of AI competence |
| Doubts about the reliability of AI-generated results | NO-21 txt L161-162 (p216) | Attitude toward use, not a capability statement |
| "which decisions stay with people, and which AI finishes and delivers directly — all of it follows real needs" | NO-03 txt L85-86 (p27) | The line follows site needs |
| "where they want time saved, and where they wouldn't rest easy even with the time saved" | NO-03 txt L82-86 (pp26-27, spans footer) | Field-research finding, in the source's words |
| "70% execution work such as drawing and code checking; 30% professional judgment and scheme adjustment" | NO-21 txt L116-117 (p213) | Participant's own estimate, pre-AI; aside only |
| "all analysis results, strategy suggestions and document drafts deposited in the case space, forming a traceable work record" | NO-03 txt L174-179 (pp30-31, spans footer) | Traceable backstop |
| "Two months after launch, Fasuo AI's users had spread across multiple cities nationwide" (Figure-2 caption) | NO-03 txt L140 (p29) | Proves spread, not effect |

- Number calibers, marked inline: "three to five workdays," "1,000+ pages," "dozens of material types" = the interviewed FDE's industry observation, no sample, no provenance — background tension only. NO-03 is live in normal operation, with no activity data (the source's efficiency figures are single-case self-reports, uncited here); this piece cites no benefit figure for it. "About 30 minutes" = relayed from the project side, preliminary result not a final deliverable, never converted into a delivery-cycle reduction. "70/30" = the participant's own estimate, pre-AI; used only as the timesheet-audit aside, never as efficiency evidence. NO-21 is live in use at a single client; user count, project count, and evidence of scale are all undisclosed in the source. The full text carries no "verified" or "data proves" phrasing.
- Framework, signed as ours: the line drawn on responsibility; "responsibility has no stand-in"; the five-breakpoint scan observation; the named-accountability exit test; "a function of liability institutions"; the timesheet-audit aside; "checklist-type judgment moved into deterministic rules" — all our cross-case decomposition and naming, not quotes from the cases, not customer-validated results.
- Falsification clause and range limits, written into the body: the attribution countercheck (no passage in either case attributes kept-human steps to "AI failed first"); the universal claim scoped to "within the material these two cases disclose"; the narrowed half ("only intermediate steps that do not directly produce external liability, and are backstopped by the next checkpoint").
- Geography and entity caliber: the NO-21 client company's name and nature are undisclosed in the source; the body says "the design site / the case's interviewed author" throughout, and "design office" in the title is the industry-generic sense. NO-03 is described as the source describes it — a commercial litigation team — with no country attributed and no flattening into "Chinese companies."
- Same-case division: NO-21 is shared with the investment-line piece in this series — that piece takes the investment-scale face (how big, which tax), this piece takes the responsibility-division face (who signs). Where one quote serves both, each piece takes one face, and neither retells the other's judgment.

---

The liability lines in these two cases are now drawn, and one question in this series is still open: the system got built, the process got changed — but inside the organization, whose is this thing? Who owns the tool, who owns the demand, where does the role sit? The three checkpoints of capability ownership are the next piece's work.

Related reading: the two blades this piece points back to each have a piece of their own — "costs split into build-time and run-time" stands in [AI Builds Your Tool. The Tool Needs AI to Run. Those Are Two Different Things.](/blog/building-with-ai-is-not-running-on-ai), and "replacement depth sorted into tiers" stands in [AI Adoption Has Six Levels: From H0 to H5, Where Does Your Company Stand?](/blog/h0-to-h5-ai-adoption-ladder). And the role that ran back and forth through both sites — the FDE who embeds in the business front line — has its business model told separately in [Your FDE Moved Into the Client's Building. How Do You Get Them Back Out?](/blog/fde-never-comes-out).

I'm Uncle J. Eighteen years in HR taught me to read exactly this line — who signs, who answers, whose name goes on what exits. I run a one-person AI organization now, so every line I draw lands on my own signature.

**The model doesn't decide how far AI can go. The signature does.**