Why Documentation Is More Important Than Ever (Thanks to AI)

If you thought your business no longer needs documentation because AI will fulfill that need and more, you're about to face a potentially expensive disappointment. AI tools based on LLMs (Large Language Models) can turbo-charge your existing business practices, including documentation, but they're not ready to replace them. In fact, good documentation has become more critical than ever - because AI amplifies both the value of good information and the cost of bad information. Here's why AI fails without good documentation and how to make it work.

"Garbage In, Garbage Out"

A more accurate but less glamorous name for Artificial Intelligence would be 'Artificial Pattern Matching and Repetition'. Ask an AI agent built into your work communication tool for an answer and it will synthesise confident-sounding nonsense from whatever bad information it finds in your public channels. No 'Intelligence' will be applied to the pile of content it uses to answer your question - it will simply regurgitate whatever it comes across, blended with mediocre sources from generic data outside your business. Worst of all, it will make its answers sound very authoritative and plausible, even when they're based on complete "garbage".

Instead, you want to tightly control the business-specific context provided to your AI helper and ensure it is correct, specific and representative of reality in your business.

For example, instead of asking an AI assistant: 'Based on this meeting transcript, write an email to my boss pitching a switch from payment gateway A to payment gateway B' - use documentation of why payment gateway A was selected, what the current challenges are, and how it fits into your business and architecture as context for the prompt: 'Act as a top-tier business consultant. Based on the documentation attached, write 10 bullet points to convince an exec to switch from payment provider A to payment provider B.' Then write a simple email in your own voice: 'Here's my thinking based on our current situation,' paste the bullet points, double-check their accuracy, and clean up. You will get a pitch any exec will want to read provided you have invested in quality documentation before that can be used in the prompt.

While the email from the first prompt will likely go straight in the bin - people, especially busy leaders, are great at spotting and filtering generic AI slop and the vague weak points often expressed in meetings - the second email will have a high concentration of specific, quality data crafted into effective language (which LLMs excel at). It will be hard to ignore.

AI Helps Analyse "What", Documentation Captures "Why"

AI tools excel at analysing and summarising large volumes of information and can be very helpful answering questions that often start with "What", for example, "What programming language are our biggest services written in?". But it it won't be able to definitively tell you Why that is the case unless you explicitly documented it.

The documentation doesn't have to be complex. For example: "We decided to implement backend services in Python because of limited seed funding and lower contractor rates among Python developers. Subject to business performance and additional funding rounds, we would prefer to eventually rewrite services in Go for better performance." The "Why" can be as simple as "We had to decide so we flipped a coin". This note stops teams guessing in the future.

Remember your AI helpers can only base their answers on what they read, hear or see. They can see the state of the code/product, they can't see the thinking that went into it unless you document it.

Reliable Source to Check AI Homework

If your business is alive in 2025, you are either providing something that is not a commodity, or providing a commodity in a way no one else does. Either way, you operate in a context specific to your business - you may even have unique know-how.

When a teenager ChatGPTs their history homework, they can cross-check against Wikipedia. When you use ChatGPT for a client pitch that includes competitive metrics, you can only verify the AI's work if you have a reliable source - your own documentation - in the first place.

Throughout my career, I've witnessed hundreds of hours wasted by teams gathering information for RFPs and RFCs from scratch, every single time. Without sufficient documentation, every few weeks they'd struggle with answering the same questions in different orders.

Thinking AI will solve this by analysing previous papers to draft new versions is rather naive as it assumes the product remains the same between RFPs. Even if it did, the teams would still have to check the drafts and go back to hundreds of hours gathering information to check AI's homework and struggle with deciding who is right, because the tool sounds so convincing.

AI doesn't eliminate the need for reliable sources, it makes them more critical than ever. By keeping an up to date library of concisely stated facts about your business or product, you are making validating critical AI-assisted artifacts easy and stress-free.

Median AI Answers Don't Bring Competitive Advantage

General purpose AI models thrive on consensus and naturally gravitate towards median responses. For competitive businesses, consensus and averages are a recipe for failure.

Assume that 5% of publicly available information represents breakthrough approaches, while 95% represents standard practice. General purpose language models are trained on all of it - so when you ask for advice, instead of the best answer, you get something average.

Regardless of the size and maturity of your organisation, documentation of your specific business knowledge, practices and decisions is a large part of your competitive advantage. General purpose AI models are an equaliser - available to Amazon and the baker next door alike. You can invest in training your own models, build a RAG (Retrieval-Augmented Generation) capabilities or, if you can't yet afford either, simply ensure your team uses your internal knowledge base. Either way, you must populate and maintain that knowledge base - get the valuable business knowledge out of people's heads and meeting conversations, and into the company data storage.

No need for vast amounts of documentation, short searchable notes of business decisions and the thinking and experience behind them go a long way. For example: "The product colour filter only displays items which contain ALL colours selected by the user. After reviewing 100 user sessions, we have found that those who selected 2 or more colours were mostly looking for items containing both." Such insights about a specific marketplace product and its users cannot be replaced by generic advice from a general purpose model - and they become the foundation of your competitive advantage.

Documenting Trains Clarity

Generating artifacts has become accessible and fast, and as a result, a large proportion of the global workforce now struggles to write an email when major AI services go down. Writing without AI assistance forces us to retrieve, organise, and present thoughts with impact.

Writing useful documentation forces us to consider its future use cases and empathise with its reader - likely ourselves in the future - the same way we empathise with our future selves when paying into a private pension. Like any mindful investment, documentation helps exercise strategic thinking and focus on future goals.

No wonder many successful businesses insist on writing a lot, including Amazon with its strong writing culture and Basecamp with its written pitches. Writing thoughts down highlights logical conflicts, lack of substance, unsubstantiated claims, and contradictions - and if you want to perform well, it forces you to fix them.

If you want to find practical examples of clear written communication, common mistakes and how to avoid them, I strongly recommend finding publicly available posts by current and former Amazon employees or at least give this proompt to your favourite AI chat (although I still recommend the former):

(Results may vary depending on how much you've documented about your own business context first.)