Building an AI-First Snack Company

You will receive a data file for this session during the program. Please download the file to your laptop before class. We will use it throughout the session, along with your preferred LLM.

The Professor's Questions

1. What are your favorite personal and professional uses of AI?
2. How do you expect AI to change your business over the next 6-18 months?
3. How would you like your employees' use of AI to change? What could you do to accelerate that change?

My Answers

Q1. What are your favorite personal and professional uses of AI?

My most useful personal use is turning learning material into something I can understand, discuss, and revisit. For OPM, I use AI to connect the assigned readings with the professor's questions, prepare answers in English and Chinese, and organize my reflections into an article. I also use it to develop illustrations and a preview that I can review. This is a connected learning workflow, not only a search for a quick summary.

I value the ability to compare interpretations and practice explaining an idea in simple English. But I still need to return to the source, check numbers, distinguish a case fact from an inference, and decide what I actually think. A polished answer is not proof that I have learned the material.

For professional work, my preferred direction is assistance with knowledge organization, research synthesis, drafts, and comparing alternatives. Those are proposed priorities here, not a claim about current staff practices. The goal is to free attention for judgment, relationships, and difficult exceptions, while retaining responsibility for what is delivered.

My test is whether the entire task becomes more useful and reliable. I would count checking and revision time, not only generation speed. I would also be explicit when input is missing—as with the class data file—rather than ask the tool to produce a plausible-looking result.

Q2. How do you expect AI to change your business over the next 6-18 months?

Over the next 6–18 months, I expect the most useful change to come from redesigning selected workflows around AI assistance. This is a working hypothesis, not a technology forecast or an internal operating plan. Drafting, retrieving information, comparing documents, and preparing explanations may become faster. Whether that improves the business depends on the quality of inputs, verification, and handoffs.

I would start with one repeated, bounded task using material appropriate for the tool and the task. First define the output standard and measure a manual baseline. Then compare AI-assisted work on similar tasks, including review and rework. If it performs well, turn the useful instructions, examples, and checks into a shared method. Only then expand its scope.

In a trust-based service, speed is valuable only when the explanation is accurate and the promise can be delivered. I would expect people's work to move toward framing problems, interpreting evidence, handling exceptions, and maintaining relationships. I would not assume a specific headcount reduction, revenue increase, or autonomous service capability without evidence.

I would track cycle time, important errors, first-pass acceptance, and recipient usefulness. If drafts are faster but verification consumes the gain, I would narrow the task or improve the inputs. If evidence remains weak, I would stop that use case. The business change I want is more dependable judgment and service, not merely more generated content.

Q3. How would you like your employees' use of AI to change? What could you do to accelerate that change?

I would like employees to move from occasional prompting to accountable use within a real task. A good user can explain the objective, provide relevant context, check the result, disclose uncertainty, and know when to ask a person. This is my desired direction; I do not have a verified baseline for current employee use in this preparation.

To accelerate learning, I would select one task with the team, give protected practice time, and provide an approved set of tools and suitable sample materials. Pair a domain expert with a curious user. Show both a strong result and a failure, including how the error was found. A short weekly demonstration should focus on what changed in the work, not who wrote the cleverest prompt.

For a four-week pilot, week one sets the baseline and quality rubric; week two tries assisted drafts; week three tests unfamiliar but comparable examples and exceptions; week four compares total effort and quality. These are suggested stages, not commitments already made. One named owner should decide whether to expand, revise, or stop.

I would reward useful outcomes and honest reporting of problems. I would not make message counts or tool usage a performance target. Employees need to know they can reject an output and escalate an exception without being treated as resistant to change. Keep reference material current, make successful methods reusable, and retain human approval for consequential commitments. The cost is time for learning and review; leadership must actually allocate that time.

Case Summary

This session asks students to bring their own experience of using AI into the classroom, then practise together using a designated data file and their preferred LLM. Preparing well means more than knowing the names of tools: it also means being able to explain what work a tool has changed, and which judgements you still have to own yourself.

This process links input, analysis, expression and delivery — and it also shows that checking against the original, consistency in how numbers are defined, personal stance and final review cannot be skipped.

The professor's second question shifts the lens to how the business will change over the next 6–18 months. This calls for conditional scenarios and experiments, not claims that the organisation has already transformed. Gathering materials, producing first drafts and repetitive processing may be the first parts of work to change; judgement standards, knowledge updates, accountability and exception handling determine whether that change can become a reliable service.

The third question is about employees' usage habits and how leaders can encourage them. There is a gap between one person occasionally getting a good answer and a team consistently delivering verifiable results. Bridging it takes use cases, time, reference examples, feedback and clear accountability for judgement.

The data file hasn't arrived yet, so I can't choose fields, aggregate metrics or announce classroom conclusions in advance. I've prepared the workflow below for checking the file, clarifying the question, verifying calculations and interpreting results; once the file is in hand, I can go straight through it for the actual exercise.

Key Data

Type	Content	Boundary
Professor's requirement	Download the designated data file to a laptop and use it in class with a preferred LLM	The data file has not yet been obtained; other cases' spreadsheets are not to be used as substitutes
Professor's questions	Personal/professional uses; business change over the next 6–18 months; employee use and how to encourage it	Answer each of the three original questions in turn
Confirmed personal use	This preparation, bilingual expression, articles, illustrations and previews	An observable workflow; does not represent all personal preferences
Unknown internal status	Employee adoption rate, time spent, quality baseline, customer satisfaction, data permissions	No percentages or performance figures that haven't been provided
Suggested pilot period	First evaluate one repetitive task over about 4 weeks	A management suggestion — not the professor's requirement, nor a project the company has implemented
Suggested efficiency metrics	Full cycle time = drafting + checking + rework + communication; net saving = baseline time − pilot time	Compare tasks of the same complexity and quality bar; don't count generation time alone
Suggested quality metrics	Rate of significant errors, source traceability, first-pass acceptance rate, escalation/rework	Grade errors by severity; don't let average accuracy hide critical mistakes
Future judgement	Gradual process restructuring as the working hypothesis for 6–18 months	Not a firm prediction about model capabilities, the market or staffing changes

Key People

Role	Responsibilities	Key Connection
Employee as task owner	Provides context, uses the tools, checks and delivers	Has resources to learn, and is responsible for flagging uncertainty
Reviewer/domain expert	Tests key conclusions and exceptions	Does not mean redoing every word from scratch
AI tool	Organises, generates candidates, assists with calculation and expression	Output is material to be verified, not independent authorisation
Recipient/customer	Uses the result and gives feedback on whether it helped	Success should be judged by whether their problem was solved
Professor/program	Provides the data and the exercise context	The actual analysis must wait for the designated input

Core Tensions

1. The trade-off between faster drafting and the time needed to verify and absorb understanding.
2. The distance between individual fluency and repeatable team delivery.
3. Fuller context may improve output, but the scope of input must fit the task and the authorisation.
4. Tools can generate candidate positions, but individuals still have to decide what to believe and what they are willing to own.
5. The trade-off between seeing many possibilities and choosing one measurable pilot.