Rotman Full-Time MBA AI Camp
Assistant Professor of Finance, Teaching Stream
Director of Teaching Innovation, FinHub
Rotman School of Management
August 24, 2026
Kevin Mott · Rotman School of Management
Mathematics
Northeastern University
Financial economics
My Ph.D. work used deep learning to solve hard math problems that model the macroeconomy or price interest-rate derivatives.
Teaching at Rotman
Director of Teaching Innovation, FinHub
The math matters. For most managers, intuition matters more.
You will not need to understand the mathematics inside the model or build one yourself. You will need to decide what to ask it to do, whether to trust the answer, and what you still need to understand yourself.
To understand AI’s role in our work and learning, we need a simple way to describe its relationship with us. Economics gives us two familiar categories: complements and substitutes.
Complement
Using more of one makes the other more useful.
Substitute
Using more of one reduces the need for the other.
Which one is AI?
Goods we tend to use together.
Use AI to generate alternatives, test your reasoning, and improve the work while you remain responsible for the result.
For one trip, either can take the place of the other.
Have AI finish the homework, then submit the answer without understanding it.
So which one is it? What have we seen in the news?
For one person, AI can add to the work or replace some of it. Across firms and occupations, the same distinction becomes a labour-market question.
Is AI more likely to complement or substitute for white-collar workers?
The case for complement
The case for substitute
Both stories are economically coherent. Which one dominates is an empirical question.
The WSJ asked economists how AI will affect work, including Rotman's Ajay Agrawal and Joshua Gans. Their answers split between complement and replace.
Read the article ↗
AI can substitute for you on a well-defined task while complementing you at the job or project level. The next question is which tasks AI can complete from start to finish.
Imagine that your career after Rotman goes exceptionally well. You hire an experienced wealth manager to invest a substantial sum on your behalf. You then learn that the manager delegates some client work to an intern.
Raise your hand. Keep it raised for as long as you would be comfortable allowing the intern to complete each revealed task without supervision during the work.
AI puts you in a manager’s role sooner, whether or not anyone has given you the title. Once you delegate work to AI, you take on the manager’s job: choose what to hand over, define what good looks like, review the work, and decide when it is ready to use.
The partner reviews the finished result
The partner reviews each step
At the rote end, inspect the finished task. Toward the creative end, inspect and approve each step.
Let’s make this concrete with examples of each type of task from work I’ve been doing over the past few days.
Closer to rote
The output was cheap to inspect, so I could delegate the mechanics.
Closer to creative
Each transaction affected later journal entries and statements. An error early in the case could change every output that followed.
My first semester teaching
Option A
The one on the left.
Option B
The one you are looking at now.
What I delegated to AI
What remained mine
Moving from PowerPoint to HTML files used to require more web-development work than I could justify. Delegating the web mechanics was safe because this was a slide deck, not a web product with unseen systems behind it, and I could inspect every rendered slide quickly.
The accounting case was different: one mistake could change every later record and statement.
I teach an introductory accounting and finance course for engineering undergraduates. I wanted a case that felt like an actual business, not a sequence of disconnected textbook transactions. So I wrote my own. The tedious part was checking every journal entry and financial statement after each change.
My decision
Choose the next business stage
Month 1: set up and finance the company.
Approved inputs
Define the accounting for that stage
Business events, journal entries, and assumptions.
Automated rebuild
Rebuild every linked accounting output
Trial balances, statements, schedules, and validation results.
My review
Review, then choose the next layer
Month 2: buy inventory and sell products.
I did not ask AI to build the whole case at once. I broke it into smaller stages with a repeatable sequence: choose the next part of the business, define its accounting, rebuild every downstream output, then review the result. Only after the accounting was right did I decide which transactions came next, so the case became more difficult one layer at a time.
The script does not invent the accounting. It does the repetitive work after I approve the events, journal entries, and assumptions for that stage.
Every financial-statement number comes from these journal entries. The script checks that the September entries balance before producing the statements. Once they balance, the next job is to turn them into financial statements.
| Account | Debit | Credit |
|---|---|---|
| Sep. 1 · Owners invest | ||
| Cash | 40,000 | |
| Share capital | 40,000 | |
| Sep. 2 · Bank lends money | ||
| Cash | 20,000 | |
| Bank loan | 20,000 | |
| Sep. 3 · Pear buys equipment | ||
| Equipment | 55,000 | |
| Cash | 55,000 | |
| Sep. 12 · Customers pay for setup services | ||
| Cash | 15,000 | |
| Service revenue | 15,000 | |
| Account | Debit | Credit |
|---|---|---|
| Sep. 20 · A university will pay later | ||
| Accounts receivable | 40,000 | |
| Service revenue | 40,000 | |
| Sep. 25 · The university pays part | ||
| Cash | 5,000 | |
| Accounts receivable | 5,000 | |
| Sep. 29 · Pear pays wages | ||
| Wage expense | 10,000 | |
| Cash | 10,000 | |
| Sep. 30 · Pear pays rent | ||
| Rent expense | 10,000 | |
| Cash | 10,000 | |
| Account | Debit | Credit |
|---|---|---|
| Sep. 30 · Record depreciation | ||
| Depreciation expense | 900 | |
| Accumulated depreciation | 900 | |
| Sep. 30 · Accrue interest | ||
| Interest expense | 200 | |
| Interest payable | 200 | |
Once the journal entries balance, the script posts each debit and credit to its account, builds the trial balance, and produces all four financial statements. Change one entry and it rebuilds the whole chain.
Validated inputs
Start with balanced journal entries
Every debit has an equal credit.
Posting
Move each line to its account
The ledger accumulates each balance.
Statement build
Build all four financial statements
Income, changes in equity, balance sheet, and cash flow.
Reconciliation
Check that every statement agrees
One failed check stops the build.
I know what the entries and statements should look like.
What the system handles
What I still decide
Automated checks catch entries and statements that do not reconcile. My accounting knowledge catches transactions that are wrong even when every total still balances. That makes verification cheap.
Unit economics: Without AI, I would have checked the balance by hand after every change. I never would have built reusable transaction logs and verification scripts that make it easier to teach the course again.
Research on the Jagged Frontier shows that AI’s abilities can be uneven across tasks that people might judge as similar.
Fabrizio Dell’Acqua et al. (2026) Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science 37(2):403-423. https://doi.org/10.1287/orsc.2025.21838
Creative work stays high-touch
If I have to define every step, inspect every handoff, and correct the process, how can this possibly save time?
The first pass may not save much time. The gain comes when the work repeats: preserve the process you already debugged instead of rebuilding it from scratch.
Skill
Reusable instructions, examples, templates, and checks for a particular kind of task.
Agent
Works across multiple steps, uses tools, inspects results, and stops or asks when it reaches a boundary.
A skill is a set of reusable instructions. An agent can follow those instructions, use approved tools, inspect its work, and complete a job within limits you define.
An LLM produces a response. An agent can also use approved tools, inspect what happened, and continue through several steps until it finishes the job or reaches a decision that requires you. A skill can provide the instructions it follows.
Before delegating the whole workflow
Can I name the inputs, the result I expect, the checks it must pass, the tools it may use, and the decisions that still require me?
A good candidate
Develop the workflow first
An agent can follow instructions written for a single job. If the same method will be useful again, save the instructions, examples, and checks as a skill.
Broad assignments create scope creep: the model starts taking on adjacent work before you can inspect the part you actually asked for.
Define one specific output
Provide the inputs and constraints
Inspect and correct the current result
Approve the next step only when it is right
For complex work, give AI one clearly defined step, inspect the result, and only then continue. If that sequence will repeat, package it as a skill an agent can use.
After today, choose one workflow you already know and use AI to improve it. Start with work you have done often enough to recognize a bad result yourself.
New model
When a model appears in the news, try it on work whose answer you can already judge.
Different provider
Try tools from OpenAI, Anthropic, Google, and others. Also try open-weight models—models that you or your organization can run directly—when they fit the task.
New tool
Check whether a new tool improves the result, reduces the time, or makes verification easier.
Computer work is unusually good for experimentation because the feedback is immediate: a script runs, a page renders, a spreadsheet balances—or it does not.
Build the habit of testing new capabilities on work you already understand well enough to verify.
You already have ChatGPT Edu and Codex. Start there. Use only personal, public, or authorized data.
Kevin Mott
I host course materials on my personal site, including this deck. Come back whenever you need a refresher, and stay in touch if I can help.
Assistant Professor, Teaching Stream
Director of Teaching Innovation, FinHub
Managing and Leading with AI
