• Acquisition: collecting and measuring data
• Storage: could just be excel, but hopefully not!
• Cleaning: ensuring the quality of the data, audits
• Combination: combining multiple data sources
• Analysis: modeling the economics
• Communication: conveying accounting insights
• Data driven decisions: using analytical conclusions to inform decision-makers

• Data-driven decision-making
• Improved operations
• Better customer understanding
• Risk assessment & management
• Regulatory compliance

• Regardless of your analytics chops, your domain expertise will always be valuable
• Don’t underestimate how impactful knowing accounting can be in DA conversations
• Don’t overestimate how much others know about accounting, you’re the expert!
• Play to your strengths

• There’s a lot of it
• The 5 Vs:
• Volume
• Velocity
• Variety
• Veracity
• Value

• Only pay for storage (bits) and usage (CPU cycles)
• Server hardware maintained by third party
• Shifting CapEx to OpEx with EBITDA impact
• Provider shares resources amongst clients (or not)
• Server software maintained by third party
• Firm avoids dealing with software updates, fixes, etc.
• Customization often more difficult (platform lock-in)
• Leverage external analytics expertise

• Cloud systems solve problems (tech assets, access, scale) but create others (security, complexity)
• ROI on moving to the cloud is unclear
• 66% of execs say cloud hasn’t lowered IT cost (KPMG 2022)
• Can create audit-trail concerns (right to audit)

• Data Lake: centralized hub of unstructured data
• Stores data in “raw” unprocessed format (as-is)
• Because storage is “cheap,” firms now often keep all data in its raw form, and process later when needed
• Data Warehouse: centralized hub of structured data
• Stores pre-processed data (schema-on-write)
• Governance and quality controls ensure datasets are properly formatted, normalized, ready to use

• Text, images, etc.
• Structure is missing, not-consistent, or complex
• 80-90% of collected data is unstructured
• Needs processing before it can be used
• AI is making processing this data far easier
• Examples:
• Invoices, receipts
• Regulations, FAQs, instructions, guidance

• Think of a spreadsheet with columns (variables) and rows (observations)
• Often quantitative, easily processed for analytics
• Can be broken into multiple, related tables (RDB)
• Examples:
• General ledger
• Chart of accounts

• “Wide” format: add data in columns
• Each column is a different variable
• “Long” format: add data in rows
• Repeat IDs, columns for variable and value

• Concisely conveying data
• Telling stories for retention
• Relevant XKCD

Relevant XKCD

• Leveraging strengths of human cognition
• Visual bandwidth1: About 20 Mbps (1Gbps total)
• Pre-attentive processing2: Instinctive processing
• Graphical perception3: Process some designs better
• Mitigate constraints of human cognition
• Avoid overloading limited “working memory”4
• Leverage external recognition

1

2

3

4

• Median is “middle” of data
• Mean (expected value) is weighted by values
• Data: https://www.cbo.gov/publication/60706#_idTextAnchor039

• Credit Tyler Vigen

• “You can see a lot, just by looking” ~ Yogi Berra
• EDA establishes understanding of the data before assumptions are made or complex modeling techniques are applied
• More EDA now, fewer mistakes later
• Finds “unknown unknowns”
• Be curious and judgmental
• Skepticism is important, look for errors, bias, misinformation, etc.

• Combining to get variables from multiple sources
• External data (marketing data, economic data, etc.)
• Various internal systems (invoices, production, etc.)
• Aggregation
• Aggregation can require combining data with itself
• Analysis, cleaning, and verification
• Comparing observations, interpolating, etc.

• Database

• Table 1: GL_Detail

• Table 2: Chart_of_Accounts

• SQL is the language used to interact with RDBs
• SQL playground: sqlbolt.com
• Do you need to know SQL?
• You should understand conceptually the moving parts
• You should understand what it can and can’t do
• Which you can ask LLMs
• Learn it if you need it, are interested, or taking the ISC section of the CPA

sqlbolt.com

• Just keep rows that match (join keys are in both tables)
• Most common join (JOIN keyword in SQL is inner)

• Keeps all rows in one table
• Used to add columns from another table without altering original rows
• Left and right join are the same with swapped table orders
• A left join B = B right join A

• Calculations across multiple rows
• Counting, summing, averaging, regressions, etc.
• Aggregations collapse multiple rows into one
• Often combined with merges
• E.g. total expenses by vendor within date range
• Often done in “subquery” and merged

• Dashboards, KPIs for decision makers
• Continual process analysis, internal audit
• Robotic Process Automation (RPA) to achieve specific tasks
• Small one-off tasks (filing PDFs, scraping data, etc.)

• RPA can’t handle un-known circumstances
• Well it will, but you probably won’t like it
• Human in the loop is when your process calls a human for help
• Important decision oversight
• When something goes wrong

• Accounts Payable (procure-to-pay): automate invoice processing (OCR) and vendor payments
• Accounts Receivable (order-to-cash): automate invoicing, billing, and collections
• Data Acquisition: automate downloading data, client files, etc.
• Account Reconciliations: automatically reconcile financial accounts and ledgers, 100% testing, flagging errors for review
• Report Generation: automate pulling, processing, and formatting of data into regular reports
• Payroll & HR expenses: validate timesheets, initiate direct deposits, employee expense reimbursements

• Once we have data, what then?
• Use it!
• Model the world
• Descriptive: What happened?
• Diagnostic: Why did it happen?
• Predictive: What will happen?
• Prescriptive: What should we make happen?

• Analytical models can be very powerful
• But that doesn’t obviate your professional skepticism
• Apply your domain knowledge
• Question the assumptions
• Question the model
• Question the data

• Random errors (called noise) results in worse measures and outcomes, but randomly
• E.g. predicting default risk inaccurately
• Errors all in the same direction (called bias) can result in consistently wrong results
• E.g. always underestimating default risk
• Minimize noise, beware bias

• Important to differentiate when making decisions
• Causal inference is designed to identify causation
• Has its own caveats, but better than nothing

• Machine learning is great at capturing correlation
• ML models are often impossible to understand
• Understanding a model is important to understanding when it will work and when it won’t
• E.g. ChatGPT seems to be able to “think” until you come across non-existent citations
• Accountants must justify their work
• To auditors, investors, regulators, etc.

• Statistical significance: high certainty in the model
• Economic significance: how many dollar signs?
• Remember that models don’t have business sense
• Important to translate model conclusions into economic impact on costs or revenues

• How well does the model describe the data you have
• Called in sample fit, based on training data
• How well does the model predict new data
• Called out of sample fit, based on test data

• Some models can just memorize the training data
• Does not mean they have learnt the “truth”

• Each observation is related to the next
• Order matters
• Example: Revenue data over time

• Training data could contain random economic events that aren’t indicative of the future (overfitting)
• New economic events that aren’t modeled can occur (underfitting)
• Uncertainty increases when forecasting further out

• Compare outcome to “counterfactual”
• Anything that is different between treatment and control could potentially cause the estimated effect
• Called omitted variables or measurement error
• Solutions to identify causality limit generalizability
• If you can’t define how treatment results in outcome (the mechanism), be very skeptical
• If you can, then that’s potentially testable

• Regression Approach

• Classification Approach

• Positive and negative don’t mean anything other than the two outcomes
• You could reverse them, nothing would change

• A bag of skittles is dumped on the floor, and all measurements taken (color, size, location, etc.)
• Clustering – separate skittles into colors
• Dimensionality Reduction – first “factor” is color, then location, no difference in size

• Still just fitting a line
• A very very very complex line
• LLMs predict next-word probability (e.g., logistic regression)
• Fits the line with a neural network
• Different types of NNs based on connection of neurons
• Transformers changed the game

• Text is broken into “tokens”
• E.g., shareholder equity = assets - liabilities
• Network predicts probabilities of next word
• Next word is randomly selected

• next word = m * X
• m is the weights (learned coefficients)
• X is the context (input data)
• Human analogy: long vs short term memory
• Weights are everything you know and how you think
• Context is what you’re thinking now (applying that knowledge)

• The entire context window is included in processing
• A prompt may need refining with some back & forth
• Consider ending with “write a new prompt that captures all the updates over our conversation”, try in new chat
• Use files or other resource to provide necessary info
• E.g. The HTML file from Lab 6 with table details
• Attempt different prompts, require & check citations
• Assistants have pre-set context, data, or examples

• Preloaded context (GPTs / Gems)
• Customized system prompt for specific functionality
• Retrieval Augmented Generation (RAG)
• LLM does a google search (of just your documents)

• MCP describes tools in English, simple syntax to call
• MCP client takes LLM output, talks to MCP server

• Tool using AI: an employee told to do a specific task
• Agentic AI: a manager that decides what tasks to do
• Example: tax agent to analyze tax positions, evaluate adherence to existing regulations and estimate risk
• Incorporates feedback/looping, executing until it observes and decides its objective is met
• Can be hierarchical, high-level agents orchestrating a series of specialized sub-agents

• 44% of CFOs report using AI, only 33% at scale
• 88% of employees report using AI
• 37% worry reliance on AI will erode skills & expertise
• 64% of employees report higher workload
• Only 5% using AI to significantly alter their workflow
• Q1 2026: 54% of firms deploying AI agents
• 60% expect humans to manage agent “teams”

• Desire to implement AI / data analytics is ubiquitous
• Ability to do so is not, creating the value gap
• Common frictions to adoption:
• Technology learning curve
• Employee adoption
• C-level acceptance
• Implementation difficulty (trust & governance)
• Data availability

• Data security & privacy
• 36% employees admit to using non-approved LLM
• Accuracy & data
• Mitigating hallucinations
• Acquiring useful, clean, and reliable data

• Source: KPMG AI in financial reporting and audit, 2024

• ACCA report finds half of accounting and finance leadership in AI adoption roles, 20% strategic owners
• Natural extension for the business data roles
• Moving from data preparation to advisory roles

• Data engineering: clean data, ETL, SQL, EDA
• Data visualization: histograms, line charts, trends
• Data analysis: regressions, classifiers
• AI: vibe coding

• “The greatest enemy of knowledge is not ignorance, it is the illusion of knowledge.” ~ Daniel J. Boorstin
• “If you can't explain something to a six-year-old, you really don't understand it yourself.” ~ not Albert Einstein
• As the data expert, you are the most valuable and necessary part of the analysis. ~ Me

• Analytical models can be very powerful
• But that doesn’t obviate your professional skepticism
• Apply your domain knowledge
• Question the assumptions
• Question the model
• Question the data

• Wednesday: Lab 12 & Project 4 work-session
• Next week (week 14):
• Monday: Live lab / demo
• Wednesday: KPMG AI guest speakers
• Week 15: Presentations
• Sign up on spreadsheet in announcement
• Load presentation slides before class
• Hard stop at 10 minutes