How to Validate Knowledge Work
Code can be validated by its output; knowledge work can't. You have to examine the process, the inputs, and the reasoning. Products need to be designed for that.
There's a fundamental difference between code and knowledge work that decides how you validate AI output: code is highly output-oriented; knowledge work is not. With coding tasks you can verify correctness with tests; with knowledge work you can't trust the result just by looking at the final artifact.seshan
What you'll run into:
- You asked AI for an analysis, the numbers look great, and you have no idea how it got them
- You asked AI to write a plan, it reads smoothly, and you suspect it missed a critical input
- Everything AI produces looks right, yet you can't shake the feeling something's off — and you can't say where
Why code can be judged by output
Coding is output-oriented. Give AI a coding task and you can verify whether it did it correctly with tests — try it, see if it works, validation based on output. If the tests pass, it's very likely correct, because the intermediate details don't matter: you don't need to know its architectural decisions, only that it passes every test.
Output-oriented validation is cheap, objective, and fast. That's why coding agents progressed fastest — they have a natural referee: the test. Every change gets judged correct or not immediately, and the feedback loop is astonishingly short.
Why knowledge work can't be judged by results
Knowledge work is different. You can't look at the final number in the deck — say "90% success rate" — and believe it. You have to examine the process, the inputs, the reasoning, and how it got there step by step. A financial model can look complete while sitting on a wrong assumption; a competitive analysis can read as thorough while missing a key source.
The reason is that knowledge work output is inherently faked-able: a pretty deck doesn't prove the reasoning is sound, only that it's well formatted. The final artifact carries almost no signal about whether the process behind it was trustworthy — and the process is the only thing worth trusting. The prettier the number, the more you should ask where it came from.
Products need to show proof of work
This sets a clear design requirement for AI knowledge-work products: let users see the work in progress, the citations and inputs, and walk the whole journey with the model to the final output. Users need to be able to convince themselves that "this thing is right, good, and useful" — not be handed a finished artifact.
This is the same thing as an executive meeting pitch. When you argue "this is what we should do," a large part is helping them see the work you did and every step of it — AI needs to show its proof of work too. The more it does, and the more transparent the process, the more you can trust its conclusion.
How to check an AI knowledge-work product
To judge whether an AI product is trustworthy, don't just look at the demo — check whether it gives you three things:
- Visible process. Does it show the work in progress and intermediate steps, instead of only dropping the conclusion?
- Traceable citations and inputs. Does it tell you which source each number and claim comes from? Which key input might be missing?
- Followable reasoning. Can you walk through its thinking to see how it got from input to conclusion? Or are you stuck with a black-box answer?
If all three are there, you have a chance to verify its work; if any is missing, you're back to "trust the numbers." One related check is context: does the agent have the context it needs to do what you asked — can it see your email, all your documents? An agent that starts working without context is more suspicious the prettier its output gets.
When to look at output, when at process
Not all knowledge work needs the full proof process. The dividing line is how irreversible the consequences are: if a wrong conclusion causes real loss, is hard to detect, and affects others, you must look at process — financial models, recommendations, external copy, risk assessments. If a wrong conclusion is easy to spot, fast to correct, and only affects you, output is enough — drafts, brainstorming, summaries, format conversion.
A practical habit: separate "needs to be believed" from "just reference." For what needs to be believed, require proof of work; for what's just reference, output suffices. Mix the two together and you'll be stuck in a false choice between trusting everything and trusting nothing.