Why AI Needs Expert Oversight: Lessons From Typewriters to Spreadsheets

TL;DR

AI is the most useful business tool since the spreadsheet, but it cannot tell you whether its own output is right.
Every new tool, from writing to the typewriter to VisiCalc, removed labour but not judgement. The spreadsheet ended up creating more accountants, not fewer.
Lawyers, an airline, a newspaper and a Big Four firm have all been caught out by unchecked AI output. None of them lacked expertise; they skipped the check.
Good prompting improves the first draft. The real guard rail is a person who knows what right looks like.
At Saigon Digital, AI does the legwork and our team owns the output: we publish, design, interpret the data and sign it off.

AI is the most useful tool businesses have been handed since the spreadsheet. But a tool cannot tell you whether its own output is right, and AI is unusually good at producing work that looks right when it is not. The value has moved from producing the answer to knowing whether the answer is correct. That is the part you still need an expert for.

Almost every client conversation I have now ends up on AI. "Can we use it for content?" "Can it run our ads?" "Do we still need an agency if ChatGPT can write?" Fair questions. Here is my honest answer, starting about 2,400 years ago.

Has a new tool ever made the expert unnecessary?

Not yet. Each one has been sold that way, and each one has ended up raising the value of the people who knew how to use it properly.

Writing. In Plato's Phaedrus, Socrates tells the story of the god Theuth presenting writing to the Egyptian king Thamus as a gift for memory and wisdom. Thamus is not convinced. People who learn from written marks, he warns, will have "the show of wisdom without the reality". Swap "written marks" for "a chatbot" and that line could run in a tech newsletter this week. Writing did not make people stupid. It also did not turn everyone who could read into a philosopher.

The typewriter. Mark Twain bought a Remington typewriter in 1874, and Life on the Mississippi (1883) is widely regarded as the first book manuscript submitted to a publisher in typed form. The machine made text faster, cleaner and easier to copy. It did not make anyone Mark Twain. (Twain did not even type it himself. He had a typist. Even the early adopter delegated.)

The spreadsheet. When VisiCalc went on sale in 1979, it could do in seconds what used to take an accountant a full day of erasing, re-adding and cross-adding columns. You would expect accountants to disappear. As NPR's Planet Money tells it, lots of bookkeepers and accounting clerks were replaced by spreadsheet software, but the number of accountant jobs actually increased. Once calculation got cheap, businesses started playing the "what if" game: what if we hired more people, what if we charged a little less? They needed people who knew which questions mattered and which answers to trust.

The pattern is consistent. The tool removes the labour. It does not remove the judgement. It usually increases the demand for it.

What are you really paying an agency, or an accountant, for?

Two things: time and expertise.

I run a business. I do not do my own accounting. I could, in theory. I am a developer, and with AI I could build a forecasting model in an afternoon. It would have tidy tabs, sensible-looking formulas and a confident summary at the top.

Would it look right? Absolutely.

Would I know whether it is right? No. I would not know which assumption was quietly wrong, which tax rule had changed, or which number the model had filled in because it sounded plausible. And if I cannot validate it, the speed is worthless. I have just produced a wrong answer faster.

That is why I pay an accountant. Not because I cannot open a spreadsheet, but because I want someone who knows what right looks like to put their name to it.

Your marketing, your website and your data work exactly the same way. AI can produce the draft, the design concept, the keyword list, the report. The question every business owner should ask is: who checks it, and do they know enough to catch what is wrong?

What happens when nobody checks AI output?

These are not hypotheticals. They are on the record, and the people involved were professionals.

The lawyers. In Mata v. Avianca (2023), two New York lawyers filed a brief citing court cases that ChatGPT had invented. Judge P. Kevin Castel fined them and their firm $5,000 and ordered them to write to each judge falsely named as the author of the fake opinions.
The airline. In February 2024, a Canadian tribunal ordered Air Canada to compensate a passenger after its website chatbot invented a bereavement refund policy. The airline argued the chatbot was responsible for its own words. The tribunal disagreed: it is your website, so it is your answer.
The newspaper. In May 2025, the Chicago Sun-Times ran a syndicated summer reading list in which only five of the 15 books existed. The freelance writer had used AI to compile it, and nobody between him and the printing press checked.
The consultancy. In October 2025, Deloitte agreed to partially refund the Australian government for a report worth almost A$440,000 after reviewers found fabricated references and a made-up legal quote. The corrected version added a disclosure that generative AI had been used.

Look at who is on that list: lawyers, a national airline, a newspaper and one of the Big Four. None of them lacked expertise. They skipped the step where the expertise gets applied.

To be fair to AI, this is not a new failure. Spreadsheets have done it for decades. In 2013, a graduate student found that an influential Harvard economics paper, widely cited in the austerity debate, had left five countries out of a key average because a formula covered the wrong rows. Professor Ray Panko's reviews of spreadsheet audits found errors in the large majority of real-world spreadsheets examined. The lesson is the same: unchecked output is the risk, whatever produced it.

Why does AI output need a different kind of checking?

Because AI errors do not look like errors.

A broken spreadsheet formula often shows itself: a #REF!, a total that is obviously off. A fabricated citation from an AI model is perfectly formatted, plausibly titled and delivered in exactly the same confident tone as the correct ones. You only catch it if you know the subject well enough to be suspicious, or you check every claim against a source.

Even our sense of whether AI is helping can be off. In a July 2025 randomised trial by the research group METR, 16 experienced open-source developers worked on real tasks with and without AI tools. With AI, they took 19% longer. Afterwards, they believed AI had made them about 20% faster. That is one study, with early-2025 tools, on a specific type of work, so I would not generalise the number. But the gap between how productive it felt and how productive it was should make every business owner pause.

This is also where prompting comes in. Good prompting matters: giving the model the right context, the real source material, clear constraints and a definition of "done". Skilled prompting gets you much better first drafts. But prompting is not a guard rail. A well-prompted model can still be wrong, and it will still sound sure. The guard rail is a person who can tell.

What do good AI guard rails look like?

If you are bringing AI into your business, in-house or through a partner, these are the questions I would ask:

Does someone know what "right" looks like before the AI starts? If nobody could have done the task without AI, nobody can check it either.
Is the AI working from real sources? Ask it to work from your data, your documents and named references, not from what it "remembers".
Is every fact, figure and citation checked against a primary source? Not spot-checked. Checked.
Does a named human own the output? Someone who signs it off and is accountable if it is wrong. Air Canada learned that "the chatbot said it" is not a defence.
Does the level of checking match the stakes? A brainstorm needs a glance. Anything touching money, legal claims, health, customers or your search visibility needs a proper review.
Are you open about how it was made? Google's own guidance says it rewards "high-quality content, however it is produced", and that disclosure helps where a reader would reasonably ask how something was made. Our guide to AI content disclosure covers the practical side.

How does Saigon Digital use AI?

We use it every day, and we tell clients that up front. Hiding it would be silly, and pretending we do not use it would be worse.

Research and data. AI helps us pull and explore large data sets from Search Console, analytics and SEO tools, and test how brands show up in ChatGPT, Perplexity and Google's AI answers. A person then checks the patterns against the raw data before anything reaches a client report. (More on this in our guide to AI for data analysis.)
Content. AI helps with research, outlines and first drafts. Our editors decide what gets said, add the first-hand experience, and check every number against a source. Our internal writing rules put it plainly: "One invented fact destroys trust with a sceptical reader."
Design. AI helps us explore ideas and variations quickly. Our designers decide what fits the brand and the user, and what actually goes live.
Development. AI speeds up how our developers write and test code. Every change is still reviewed by our developers before it ships, because a confident bug is still a bug.
How we run the agency. We build AI into our own tools and workflows, from audits to internal assistants, so the team spends less time on admin and more on the work clients pay for.

The rule underneath all of it: AI does the legwork, our team owns the output. We are the ones who publish the content, sign off the design, interpret the data and put our name to the recommendation. We even have an internal rule against "slop grenades": dropping a wall of unedited AI text where one human sentence would do the job.

That is what you are buying from an agency in 2026. Not access to the tools. You already have those. You are buying the time you do not have, and the expertise to know when the tool is wrong.

So should your business use AI?

Yes. Use it to draft, to research, to explore your data and to move faster. Just do not confuse a fast answer with a right one.

The typewriter did not replace the writer. The spreadsheet did not replace the accountant; it created more of them. AI will not replace expertise either. It makes answers cheap, which makes knowing which answer is right the valuable part.

If you want to bring AI into your marketing, website or reporting without handing over control, read how we approach AI implementation, or talk to us directly. We will show you where AI helps, where it needs a human, and who on our team checks it.

Frequently asked questions

Can AI replace an agency or an accountant?

AI can replace a lot of the manual work, such as first drafts, data pulls and calculations. It cannot be accountable for the result, and it cannot reliably tell you when it is wrong. You still need someone who knows the subject to validate the output and sign it off.

What is an AI hallucination?

An AI hallucination is when a model produces information that sounds plausible but is false, such as a book, court case, statistic or policy that does not exist. It is delivered with the same confidence as correct information, which is why expert review matters.

Does Google penalise AI-generated content?

No, not for being AI-generated. Google's guidance says it rewards high-quality content "however it is produced", and penalises content made primarily to manipulate rankings. Unchecked, low-value AI content at scale is the risk. See our post on whether AI content hurts rankings and our guide to E-E-A-T for AI content.

Is prompt engineering enough to make AI output accurate?

Good prompts produce much better first drafts, especially when the model is given real source material and clear constraints. But a well-prompted model can still be confidently wrong. Prompting improves the input. Human validation protects the output.

Was this article written with AI?

Yes, in part. AI helped with research and a first draft. I rewrote it, checked every source and signed it off. Which is rather the point.

Sources

Checked on 11 November 2026.

Plato, Phaedrus 274c to 275b, Fowler translation via Perseus
Mark Twain and the typewriter: Open Culture and Vanderbilt University Library
VisiCalc and accounting jobs: NPR Planet Money, "How The Electronic Spreadsheet Revolutionized Business" (2015)
Mata v. Avianca sanctions: LawNext and the American Bar Association
Moffatt v. Air Canada, 2024 BCCRT 149: CBS News
Chicago Sun-Times reading list: Axios
Deloitte Australia refund: AP via CTV News
Reinhart and Rogoff spreadsheet error: Retraction Watch
Spreadsheet error research: Raymond Panko, "Spreadsheet Errors: What We Know"
METR developer productivity study: The Register
Google on AI-generated content: Google Search Central