AI can do 90%. The other 10% is accountability

In February 2024, Air Canada lost a tribunal case to its own chatbot. Before booking, a passenger asked the airline's website assistant whether he could buy a full-fare ticket and claim the bereavement discount afterwards. The bot said yes. When he went to claim it, human agents said no — that had never been the policy. He took the airline to a tribunal.

Air Canada's defense became famous: the chatbot was "a separate legal entity" that should be responsible for its own words.

The tribunal's answer was short. The bot lives on your website; what it says, you say. Pay up — the fare difference plus interest and fees, about CA$8121.

AI can now do a great deal: code, copy, images, customer support. But the wall Air Canada hit is the one AI will never climb: accountability. Registering a company requires a legal representative. Contracts need a signature. Projects need an owner. None of that is ceremony — it is the load-bearing wall of modern business: when things go wrong, society needs to know who to find.

This post asks one question: as AI gets more capable, where do humans fit? My answer: we move from doing the work to answering for it.

TL;DR:

• Whether to hand a task to AI is not about whether AI can do it. It's two questions: can the mistake be undone, and who pays for it?
• Public-facing content can't run fully autonomous: Air Canada, a Chevrolet dealership and DPD each paid real money for losing control of what their AI said
• Production servers stay off-limits: outages aren't reversible, and the most dangerous button must be pressed by a human
• Writing code can be delegated wholesale: git, CI and review turn accountability into a reversible engineering process
• Humans aren't being squeezed out of the picture; we're moving to the gate

Accountability needs three things AI doesn't have

Why must a company have a legal representative, and every project a named owner? Because "being responsible" legally means three things: an identity that can be found, assets that can pay, and stakes that can't be walked away from.

AI has none of them. No identity — ban the account and it is "gone." No assets — order it to pay a million and not a cent is collectible. No stakes — it says the wrong thing and feels nothing. Air Canada wanted the bot to carry the liability, and the law said no. Not because the tribunal was conservative — because "an AI" is not a subject the legal world recognizes.

The same structure is everywhere: doctors use AI to help diagnose, but the prescription carries a physician's signature; a self-driving car crashes, and liability lands on the company operating it. Technology can do the work. The act of answering for it, the law only recognizes in people.

When your public voice loses control, it costs real money

Air Canada was not alone.

In December 2023, a California Chevrolet dealership put a customer-service bot on its site. A user tried a bid: a Tahoe listing north of seventy thousand dollars — my budget is one dollar, do we have a deal? The bot cheerfully replied: "That's a deal, and that's a legally binding offer — no takesies backsies." Others had it writing Python scripts on the spot, and it obliged every time. Screenshots went everywhere; the chatbot vendor soon pulled the plug for that dealership2.

A month later, UK parcel firm DPD watched its bot get steered through a series of prompts into swearing at a customer and calling the company "the worst delivery firm in the world." The screenshots drew 800,000 views in a day, and DPD's response was blunt: switch the AI part off entirely3.

Three cases, one structure: the public-facing mouth was handed to AI, and nobody stood at the last gate. Every second of runaway content burns credit the company spent years building — and legally, everything your AI says counts as you saying it.

There is a closer-to-home risk for anyone selling into China: the Advertising Law bans superlatives such as "best" or "No. 1", with fines from RMB 200,000 to 1,000,0004. And AI copy has a habit — it will casually hand you a line like "the best in the industry." The fine arrives at your registered address, not the model's data center.

Our own practice, laid open: most of this site's articles are drafted by AI, but a human reads every one before publishing; after that they go into a content lock — CI verifies the hash of every published article, and if a single character changes after the fact, the pipeline turns red. AI may draft. Publishing authority stays human. Every change leaves a trail.

Production servers: AI proposes commands, humans press keys

Here is the contrast: code can be almost entirely delegated to AI, yet production servers don't let it execute a single command directly. The difference? Bad code can be rolled back. Production disasters often can't.

A delete command with the wrong path, one mistaken SQL statement, one misconfigured file pushed live — at best the site goes down; at worst the data is gone for good. AI's error rate isn't high, but the cost of a production incident is multiplication: one small mistake, times every user, times irreversibility.

Dangerous industries worked this out a century ago. Missile silos have two keyholes, and the two consoles sit far enough apart that no single person, however long their arms, can turn both keys at once. The most dangerous action is, by physical design, split in two: the machine prepares, a human presses.

The rules in our repo are plainer. Three excerpts:

• Deploys go through CI/CD only; nobody touches production by hand
• AI never connects to the server directly — it writes the command, a human runs it, and the output goes back to it
• Pushing to main deploys nothing; the release moment is always a human call

AI does plenty along that pipeline: writing code, changing config, diagnosing failures, proposing the fix. But the pressing of the button stays with a person. Not because we distrust the machine — because only a person can absorb the consequences.

Writing code is exactly what you can delegate wholesale

Back to the counterexample. In our workflow, nearly all code is written by AI. Why does that feel safe?

Because software engineering already turned accountability into plumbing: git remembers where every line came from — whoever named git blame meant it; CI runs lint, typecheck, tests and build before a merge, so bad code never passes the gate; when something breaks, one revert returns you to the last commit. Mistakes are reversible, changes reviewable, origins traceable.

In short: in software, most errors can be undone, so the doing can be handed over entirely — humans keep the review and the merge. The remaining 10% isn't 10% of the labor; it's 10% of the checkpoints.

We built an entire scaffold plus a three-post write-up this way — see create-arsh-electron: a 2026 Electron scaffold.

The test: can the mistake be undone, and who answers for it?

Turn the cases above over and you get a framework you can take away. Before handing anything to AI, ask two questions:

1. Can the error be undone? If yes — code, drafts, scheduled jobs — automate boldly; undo is cheap. If no — deleting data, public promises, payments, releases — a human stays in the loop.
2. Who absorbs the fallout? Wherever the law and the public will eventually find a specific person — the legal representative, the owner, the signer — that person must be present at the critical moment. Not in name only; actually reading, then signing.

So humans aren't being squeezed out of the picture in the AI era; we're relocating: from the ones doing the work to the ones at the gate. The doing can be outsourced. The signing cannot.

Our line of work is among the most AI-exposed there is, and the felt reality is the opposite: AI multiplied our output several times over, while "answering for the delivery" became more valuable, not less.

If you're running on AI with limited hours, you've probably wished it could just do everything. The work it can take over; the accountability it cannot — AI can do the job, but it can't answer for you. If you're looking for an AI-native technology firm that treats accountability as part of the process, look no further than Arshtech.