Subscribe
Share
Share
Embed
An AI agent can add the numbers correctly and still misunderstand what they mean. Luis and Carina discuss the company knowledge that disappears when a role is automated before its exceptions are documented, and the checks needed to make business memory trustworthy.
Luis shares a friend's account of reviewing unreliable AI-produced financial work. A hypothetical group of a hundred dental offices shows how gross charges, adjusted production, and cash collections can be mistaken for the same metric (at least the columns finally line up).
The episode covers a practical operating recipe:
The upcoming MemroOS v2 direction is business memory that improves through validated corrections and work outcomes. The discussion describes product intent. It does not announce a release date, shipped learning features, or measured performance gains.
Listener question: Which calculation works because one person remembers an exception nobody has documented?
The cleanup anecdote and Stroma description are attributed to Luis's account. They have not been independently verified. Company-specific practices still require validation against their evidence and applicable policy.
Opening: the full 45-second audio file supplied by Luis, SNL-The-Apprentice-3m36s-4m21s.mp3, from The Apprentice — Saturday Night Live. It plays at time zero before the full show jingle. The previous ASDF sample has been replaced.
Read the full article: You Fired Your Company's Memory.
What product managers can actually build with AI today—and where it still breaks.
LUIS: The spreadsheet is wrong. The AI is confident. And the person who knows why is now billing by the hour. Welcome to Contextually Aware. I'm Luis.
CARINA: And I'm Carina. Today: you fired the person who understood the spreadsheet. Then hired them back to explain the spreadsheet. (A bold new subscription model.)
LUIS: A friend of mine is helping companies figure out whether they can trust their own books. He described AI-produced financial work with incorrect calculations and unreliable results. People are going back through it, including people who understood the calculations before the automation arrived.
CARINA: Before we declare the entire economy a spreadsheet crime scene... is that your friend's account, or something established across the market?
LUIS: His account. It isn't an audited study. There is separate survey evidence of employers restoring roles. Careerminds surveyed six hundred HR professionals who had conducted layoffs. About thirty-six percent of companies in that selected sample reported restoring more than half the eliminated roles.
CARINA: Restoring a role doesn't necessarily mean bringing back the same person. And a selected sample can't tell us what every company is doing.
LUIS: Right. Klarna also described substantial AI customer support alongside investment in human support. AI and people can share the work. The management question is whether you've captured enough knowledge to change the job safely.
CARINA: Okay. A hundred dental offices. One consolidated spreadsheet. Where does the trouble start?
LUIS: Imagine one office uses production to mean gross charges. Another reports production after contractual adjustments. A third highlights cash collected. An agent puts all three into a revenue column. The arithmetic works. The meanings don't.
CARINA: But the columns line up. So, obviously, finance is solved.
LUIS: And somebody now has to explain why the consolidated number is useless. Different dentists don't get different accounting standards. The contracts and treatment timing can change the facts those standards apply to.
CARINA: So the employee's knowledge includes which definition belongs to which office, and when it applies.
LUIS: Exactly. But we still need to check that knowledge. A long-standing practice can also be a long-standing mistake. The evidence might be a contract or an approved policy. Someone remembering it confidently isn't enough.
CARINA: Couldn't we just give the agent more memory? That seems to be everyone's answer. More memory. Bigger context window. Perhaps a nicer dashboard.
LUIS: We could preserve the wrong answer. One agent saves a bad report. Another cites it. A third summarizes that citation. A reader could mistake those repetitions for independent evidence. We need to keep track of where each claim actually came from.
CARINA: So we've automated the office rumor. And given it citations. Start with definitions, then?
LUIS: Yes. Ontology sounds academic, but here it means naming the things in the business and their relationships. A charge is different from a payment. An office belongs to a legal entity. A contract has effective dates. Those distinctions need to survive consolidation.
CARINA: And mapping connects the source columns to those definitions?
LUIS: That's the work my friend describes his startup, Stroma, doing with AI. I'm relaying his description. Flatfile offers a public example of AI-assisted field mapping with validation. But the business still owns the destination definitions. A consistent label can consistently hide a mistake.
CARINA: Consistently wrong. Finally, a metric we can hit every quarter. What checks would you put around it?
LUIS: First, map the process with the people who know it. Capture exceptions and verify them. Second, give important metrics an owner, a definition, and effective dates. Resolve conflicts before agents treat them as settled.
CARINA: And where do you actually let AI do the work? Because the answer can't be six people manually copying cells forever.
LUIS: Extracting information, proposing mappings, explaining discrepancies. Tested code should perform arithmetic using approved rules. Then enforce golden rules outside the prompt. If required evidence is missing, the system can't automatically post the result.
CARINA: So the prompt saying please be accurate isn't the internal control. *Devastating.* How do you test the real thing?
LUIS: Build expert-reviewed cases, including awkward exceptions and cases that should stop for review. Hold some back for evaluation. Check results by office and financial impact. Run alongside the existing process before granting more autonomy. Keep sampling afterward.
CARINA: Where does the upcoming MemroOS version two fit?
LUIS: We're building toward memory that learns how your business operates and becomes more useful over time. Shared access to company knowledge is the foundation. The next work is learning from validated corrections, including which definitions apply and which sources have authority.
CARINA: Say the accountant corrects that production-versus-revenue mistake. What should the memory retain?
LUIS: The distinction, its supporting evidence, and where it applies. Future agents should use that reviewed knowledge without making the accountant explain it again. That's the direction we're building toward. Improvement still has to show up in evaluations. We're not claiming a measured result today.
CARINA: So it learns from a reviewed correction, not whoever sounded most confident in the meeting. That alone would be a significant workplace innovation. And when policy changes, keep the history, but say which rule applies now.
LUIS: Yes. Use AI broadly where the process supports it. Let demonstrated results determine how much work it handles. Before removing a role, have the person doing it help build the difficult cases its replacement must pass.
CARINA: Here's a question to take back to your team: which calculation works because one person remembers an exception nobody has documented?
LUIS: Start there. Capture the explanation and check it against the evidence. Thanks for listening to Contextually Aware.
CARINA: The article and research links are in the show notes. And maybe check on the person who knows why that one cell is yellow. We'll see you next time.