AI Literacy for Business Find business applications

First Questions Before AI Implementation

AI doesn’t solve problems—it amplifies solutions. Before launching a pilot, ask yourself six key questions: What exactly do you want to change? Why is this needed now? And what happens if you *don’t* implement AI?

Discuss implementation All Academy Topics

What are “first questions,” and why are they needed?

This is not a technical checklist but a strategic filter: it helps separate hype from value. Questions focus on business processes, not technology. For example: “Which step in the current process consumes the most time *and* mental energy?” matters more than “Which LLM should we choose?”

Example: Order processing workflow in CRM

“Tepliy Dom” (appliance repair) faces rising requests but refuses to hire more staff due to low margins. Manual sorting of requests by type (warranty/paid repair/consultation) consumes up to 40% of operators’ time. Solution: AI classifier for incoming emails/messages + simple routing to CRM.

No Solution: “Let’s deploy an agent who will communicate with the client as a manager.” This is the next step, not the first.

Three Main Traps

The “AI as a panacea” trap: Implementing AI into a process that already works (but slowly), without measuring current metrics.
The “Technology First” Trap: selecting a model/tool before defining the task (e.g., “we need an agent” → but why?).
The “all or nothing” trap: attempting to automate the entire cycle at once instead of isolating a single repetitive step.

What to check before the first step

Criterion	Question for verification	Minimum to get started
Repeatability	Does this happen 5+ times per week?	Yes, and the data is structured (text, table, JSON).
Clarity of rules	Can the step be described in 3–5 points?	Yes, without “depends on the client”
Measurability	Is there a metric (time, error, cost)?	Yes, even approximately
Data readiness	Are the data (history, examples) available without complex approvals?	Yes, over the last 3 months

Path from question to pilot

flowchart TD  
    A[Question: What’s blocking? Where are we spending time?] --> B[Process Selection: Repeatable, measurable]  
    B --> C[Step Breakdown: Where does AI *actually* accelerate? ]  
    C --> D[Metric Definition: Time/errors/cost before]  
    D --> E[Pilot: 1 step, 1 process, 2 weeks]  
    E --> F[Measurement: Compare against baseline]  
    F --> G{Result?}  
    G -->|Yes| H[Scale Up]  
    G -->|No| I[Return to questions: Where’s the error?]

How to formulate a task for AI (example)

Bad: “I need AI for analyzing reviews.” — Immeasurable, not specific.

Okay: “20–50 reviews arrive daily in the Telegram channel. We spend ~15 minutes/day manually categorizing them (positive/neutral/negative + topic). Goal: reduce time to ≤2 minutes/day with ≥80% accuracy on the last 100 reviews.”

Such wording allows:

Measure the current “cost” of time
Calculate ROI: (13 min/day × 22 days × ₽500/hour) ≈ ₽24,200/month
Choose a simple model (e.g., GPT-3.5-turbo + few-shot) instead of a custom one.

Pseudocode: Simple Classifier for a Pilot

// 1. Load examples (from history)
const examples = [
  { text: "Urgent! Water heater isn’t heating, warranty is still valid", label: "negative", theme: "repair" },
  { text: "Thanks, installed the kettle in 10 minutes", label: "positive", theme: "installation" }
];

// 2. AI prompt (via API)
function classify(review) {
  const prompt = `
    Classify the review by 2 parameters:
    - sentiment: positive / neutral / negative
    - theme: installation / repair / consultation / other
    
    Examples:
    ${examples.map(e => `Text: "${e.text}" → ${e.label}, ${e.theme}`).join('\n')}
    
    Review: "${review}"
    Output in JSON: {"tonality": "...", "theme": "..."}
  `;
  
  return callLLM(prompt); // returns JSON
}

// 3. Result → CRM + time metric
const result = classify("Kettle works, but it’s very noisy");
// → { tonality: "neutral", theme: "repair" }

Next step: 3 actions for today

Select one process — where there is data, repeatability, and measurable pain (time/errors/satisfaction).
Write the “baseline metric” — How much time/money/nerve does this process currently consume (even roughly)?
Formulate the task in the format “AI as a step accelerator” —not “implement AI,” but “automate step X in process Y.”

After this—transition to AI hype vs. useful automation or process map.