Operations: AI-First Snack

Unit 3, Professor Frances Frei's Operations & Technology course, AI-First Snack.

This time, the prep material isn't a long case about a company. The professor gave us three questions: how I like to use AI, how it will change business over the next 6–18 months, and how I want employees to use it.

Something I've been working on lately happens to make a good starting point: using AI to connect my Harvard readings, the professor's questions, bilingual Q&A, my own thinking, the articles and their illustrations. The process has become more complete, and it has also made me increasingly clear about which steps can't be skipped just because an answer appears quickly.

Here goes — enjoy:

Case Summary

This is an in-class exercise. The professor asked us to download the data files handed out during the course onto our laptops before class, and to use them in class together with our preferred LLM. I don't have the files yet, so at this point there are no data findings to report; preparation should start from the questions, my experience of using AI, and how to test things.

The first question is about personal and professional use. What it asks for isn't a list of tool names, but an account of what AI actually helped with in real tasks. For me, learning and writing have already become one connected process: understand the source text, respond to the questions, then express my judgement.

The second question stretches the timeframe to the next 6–18 months. Technology changes fast, but changes in business only become visible through specific work. Faster first drafts and easier search — whether they really improve delivery still depends on checking, rework and handoffs between people.

The third question turns to teams. One person occasionally getting a good result and many people reliably completing the same kind of task are two different capabilities. Leaders need to provide suitable tasks, time to practise, quality standards and support when exceptions come up.

So my pre-class preparation falls into two parts: sorting through the personal experience of using AI that has already happened, and putting forward testable ideas about the future. The latter are not yet actual business results, and shouldn't be taken as plans already carried out just because they're written in the first person.

My Thinking

A complete process is worth more than one beautiful answer

Turning a reading into a summary is convenient; combining it with the professor's questions, discovering where the evidence is thin, and then practising explaining it in another language — that's where the value gets concrete.

In this process, AI can offer candidate structures, compare explanations, help with expression, and also help generate illustrations and previews. Once the different steps are connected, the outputs are easier to keep using.

But being connected doesn't mean being automatically right. If a denominator is misread early on, the error may be written up more clearly and more beautifully later. The longer the process, the more the key facts need to be traceable back to the original material.

Checking time is part of efficiency too

If a first draft goes from an hour to ten minutes, but you then spend an extra hour finding sources and fixing errors, you can't judge the overall improvement by those ten minutes alone.

The times here are just an example to show how the calculation works, not my measured results. What I want to measure is the full cycle: drafting, checking, revising, communicating, until the recipient can actually use it.

The quality bar has to be the same as well. You can't compare the speed of unreviewed output with work that used to be delivered complete. A truly meaningful improvement is getting the whole thing done faster without any drop in reliability, or doing it better for similar effort.

Even the most fluent answer can't decide for me what to believe

When preparing a case, AI can supply arguments for several options at once. When they all sound plausible, personal judgement actually matters more.

I still have to ask myself: which arguments come from the source text, and which are inferences? If one assumption changes, does the conclusion change? Which cost am I willing to bear? When challenged in class, can I explain rather than just repeat?

The outcome of learning can't be judged by the thickness of the document. If the article is finished but I can't clearly state the most important disagreement, the understanding still isn't complete.

Employee adoption needs room for practice and feedback

The change I'd like to see is people being able to use AI around real tasks: state the goal first, provide the right context, check the output, and know how to proceed when unsure.

That takes deliberate arrangement, not just an instruction that "everyone has to use it." You can start with one repetitive task, provide examples, have people who know the business practise together with people willing to experiment, and then compare quality and total time on similar tasks.

Failure cases are just as worth showing. How an error was caught helps a team build judgement more than a pretty screenshot does. If employees worry that admitting a problem will make them look like they can't use the tool, learning may stay on the surface.

For the next six to eighteen months, start with testable hypotheses

I don't want to write my expectations of AI as firm predictions. It's more appropriate to pick a few working hypotheses: which repetitive processing can be reduced, which explanations can come sooner, and which exceptions still need a person to catch them.

Start with one process that has clear boundaries, measure the baseline first, then look at the results with AI assistance. Keep the instructions, examples and checks that work; if the effect falls short, adjust or stop.

Scaling should happen after the method has been proven useful. The number of times a tool gets used can rise quickly, but service capability has to be built up one delivery at a time.

Takeaways for Arcus42

Define value from specific tasks. First say whose problem you're solving and what it is, then decide which part the tool takes part in. With a clear scope, results are easier to evaluate.

Keep evidence and judgement separable. Mark in the output which parts are factual basis, which are inference, and which are not yet verified, so the recipient can understand how strong the conclusion is.

Keep checking and exception handling in the process. Even after speed improves, it still needs to be clear who confirms which results, and what situations need escalating.

Form a shared method through small-scale trials. Record the practices that work, the conditions under which they fail, and the fixes, so experience can be reused by others instead of living only in one person's chat history.

Direct people's time towards more valuable judgement. The point of reducing repetitive work is to leave more attention for understanding, relationships, explanation and taking responsibility. That needs to be validated by actual results.

Before Class

When I bring the data files into class, I want to first confirm what each row represents and how each metric is calculated, and only then look at what conclusions the tool gives.

This also corresponds to one expectation I have of using AI: form candidates faster, check the basis more clearly, and decide more carefully what I ultimately want to say and do.

This piece is my pre-class thinking for an in-class exercise; the data files have not yet been obtained, and it contains no actual in-class analysis results. The content about business and employees consists of general methods and ideas about the future, and does not represent processes or results already implemented by any organisation.