1. What is "prompt injection" as a risk for small businesses using AI tools?
✓ Correct. Prompt injection occurs when external (often customer-provided) text is structured to manipulate AI system behaviour — a risk when customer input is unsanitised before reaching the AI.
✗ Prompt injection is a security risk where customer-submitted text is deliberately crafted to override or manipulate the AI system's instructions.
2. When an AI expense monitoring tool flags "expense creep," what pattern is it identifying?
✓ Correct. Expense creep is the accumulation of small costs — each individually defensible — that collectively erode margins significantly. It is particularly hard to detect without AI because no single expense triggers a review threshold.
✗ Expense creep refers to the gradual accumulation of small costs that individually seem justified but together represent significant margin erosion. AI detects this by analyzing category trends and velocity, not just individual transaction sizes.
3. Amy Edmondson's term "stage setting" refers to:
✓ Stage setting is Edmondson's term for what leaders do to create psychological safety — particularly by demonstrating their own fallibility and framing failure as learning rather than incompetence.
✗ Stage setting is Edmondson's specific term for leader behaviors that create psychological safety — the most important of which is modeling vulnerability: publicly sharing your own failures, uncertainties, and learning process.
4. What does the Stanford HAI 2024 report identify as the most common source of AI-related operational errors in small business deployments?
✓ Correct. The Stanford HAI 2024 report found that removing human review steps — not model hallucination or training failures — was the most common source of AI operational errors in SMB deployments.
✗ Incorrect. Stanford HAI 2024 found that the removal of human review steps in the name of efficiency was the most common cause of AI-related operational errors in small business settings.
5. HubSpot's 2023 internal data on AI content assistant adoption found that teams with an internal champion had what adoption advantage over teams without one?
✓ HubSpot documented 2.4× higher 90-day adoption rates in teams with internal champions, confirming Rogers' insight that peer credibility is the most efficient adoption accelerant.
✗ HubSpot's figure was 2.4×. This is a significant multiplier, illustrating why seeding teams with internal champions — identified and prepared before full rollout — is one of the highest-leverage adoption strategies available.
6. What is the primary benefit of using a structured brief before prompting AI to write a job description?
✓ Correct. A structured brief containing hard requirements, schedule, pay, and culture context produces a targeted draft that filters candidates before a human reads a single résumé.
✗ The primary benefit is specificity: a detailed brief produces output that self-selects qualified candidates and filters those who won't fit — before any human review time is spent.
7. Google's Project Aristotle study (2012–2016) found that the #1 differentiator of high-performing teams was:
✓ Correct. Project Aristotle's headline finding was psychological safety as the top predictor — above talent, compensation, or management style. This finding has since been replicated in AI adoption contexts specifically.
✗ Psychological safety was Google's #1 finding. The study looked at 180 teams over four years. Individual talent and other commonly cited factors ranked lower than the interpersonal risk-taking environment the team had collectively established.
8. What is concentration risk in the context of small business financial management, and what threshold was cited as material in Lesson 4?
✓ Correct. Concentration risk from excessive single-entity exposure is flagged when a customer exceeds 20% of revenue or a supplier exceeds 30% of COGS — thresholds AI tools can continuously monitor against your transaction data.
✗ Concentration risk is the risk from excessive single-entity exposure. The lesson cited 20% of revenue for a single customer and 30% of COGS for a single supplier as material thresholds — levels AI tools can track automatically.
9. Shopify's merchant success team found that teams doing weekly error hunting sessions caught what percentage more consequential errors than teams that didn't?
✓ 68% more consequential errors caught by teams doing weekly error hunting. This is why critical calibration training — not just feature training — is essential to AI deployment safety.
✗ Shopify's figure was 68%. Error hunting sessions train the specific skill of recognizing AI failure patterns — a capability that feature training alone does not develop, and that is increasingly the highest-value human competency in AI-assisted workflows.
10. Gate 1 of the Three-Gate Tool Test asks whether the tool:
✓ Correct. Gate 1 is fit — does the tool directly address a task on your opportunity list? If not, it fails Gate 1 regardless of its other merits.
✗ Incorrect. Gate 1 tests fit: does the tool directly address a task on your pre-built opportunity list? Integration is Gate 3.
11. What is the single highest-leverage prerequisite before deploying AI personalization tools for most small businesses?
✓ Correct. Clean, unified customer data — merging POS, email, and web session records — is the essential input for any downstream AI personalization model.
✗ Data unification is the critical prerequisite — AI personalization is only as good as the customer data it ingests.
12. According to Simon-Kucher & Partners, what percentage of companies actually conduct systematic price testing?
✓ Correct. Only 5% of companies conduct systematic price testing — despite the fact that a 1% improvement in price realization produces 8–11% improvement in operating profit, far more than comparable improvements in volume or cost.
✗ Only 5% of companies systematically price-test, according to Simon-Kucher & Partners. This represents a massive missed opportunity given that a 1% pricing improvement yields 8–11% operating profit improvement.
13. What is the correct first step when implementing AI budget forecasting, according to Lesson 1's practical framework?
✓ Correct. Clean data first. No AI forecasting tool can compensate for miscategorized expenses or missing months of history. At least 12 months of clean, consistently categorized P&L data is the prerequisite for meaningful AI-assisted forecasting.
✗ The first step is always data quality: audit your chart of accounts and ensure at least 12 months of clean, consistently categorized P&L data. AI models produce output only as reliable as their input data — "garbage in, garbage out" applies fully.
✓ Correct. Sweetgreen's shift from inconsistent manager-written postings to AI-assisted standardized drafts with embedded screening criteria reduced time-to-fill for crew positions — because specific postings attract specific, qualified candidates.
✗ Sweetgreen's case showed that standardized, specific AI-assisted postings attracted more fitting candidates and reduced time-to-fill. Specificity drives applicant self-selection.
15. Zone 2 in the Three-Zone Framework covers which type of business work?
✓ Right. Zone 2 is decision-heavy work — pricing adjustments, supplier selection, staff scheduling under variable demand. AI can support these with better information but replacing judgment here is premature for most small operations.
✗ Zone 2 is decision-heavy work — tasks where evaluation and judgment are required. Zone 1 is repetitive tasks; Zone 3 is customer-facing interactions. This distinction drives very different AI strategies for each zone.
16. What does MAPE stand for and what threshold is considered "good" performance in retail demand forecasting?
✓ Correct. Mean Absolute Percentage Error (MAPE) is the standard accuracy metric for demand forecasts. A MAPE below 20% is generally considered good for retail; many businesses using AI forecasting tools report 30–50% MAPE improvement over manual methods within six months.
✗ MAPE is Mean Absolute Percentage Error — the average percentage by which forecasts deviate from actual outcomes. Below 20% is the accepted "good" threshold for retail demand forecasting. Businesses switching to AI tools commonly report 30–50% MAPE improvements within six months.
17. What percentage of B2B invoices in North America were paid late according to Atradius's 2023 study?
✓ Correct. 48% — nearly half of all B2B invoices in North America were paid late in 2023. AI-powered AR tools that predict late payers in advance allow businesses to intervene before invoices become delinquent.
✗ Atradius found that 48% of B2B North American invoices were paid late in 2023, disproportionately affecting small businesses that lack dedicated collections infrastructure.
18. The MIT study on AI-assisted decision-making found that workers who received only feature training were more likely than untrained workers to:
✓ Feature training creates false confidence. Workers who understand how the tool works assume its outputs are reliable, reducing critical evaluation. This is why training must address failure modes alongside features.
✗ Over-reliance is the counterintuitive finding. Feature literacy without critical calibration training produces workers who trust AI outputs more than they should, because they understand the mechanism but not the failure modes.
19. T-Mobile's ML-driven churn prediction models ingested how many behavioral variables per subscriber on a daily scoring basis?
✓ Correct. T-Mobile's documented churn prediction architecture ingested over 200 behavioral variables per subscriber including call quality, data usage trends, and support contact patterns.
✗ T-Mobile's model used over 200 variables — a scale of feature engineering that drove its industry-leading churn reduction.
20. A small food delivery business uses OptimoRoute. What type of constraint can the software incorporate when generating optimized routes?
✓ Correct. OptimoRoute incorporates real operational constraints: time windows (customer availability), vehicle capacity (volume, weight, temperature zones for food), and driver start/end points. These constraints make the output executable, not just theoretically optimal.
✗ OptimoRoute handles operationally relevant constraints: customer delivery time windows, vehicle capacity limits, and driver start/end locations. This produces routes that are actually feasible given real-world constraints — not just the shortest path on a map.