Dmitri Tsizov — your partner in AI automation

What I built: A data-driven customer segmentation framework for UrbanStyle.ltd, generated using Claude as an AI strategic consultant.

Five distinct customer segments (Eco-Kadri, Klubi-Maret, Trendi-Liisa, Narva-Olga, Ühekordne-Andres) with full demographic, behavioral, and value profiles. Each segment includes a marketing recommendation and a clear AI-automation priority. The analysis ends with a 6-month strategic roadmap for marketing director Anna Mets.

Key insight

Klubi-Maret = only 15% of customers but 35% of revenue. Churn prevention here is the highest-ROI AI use case.

Business impact

~60,000€ additional annual revenue + risk mitigation on 1.4M€ of revenue from the loyalty segment.

Consultant move

Flagging that Narva's 2.1% conversion is not an AI problem — it's a localization/staffing issue.

What I built: A production-grade content drafting workflow — three input triggers feeding one AI logic, generating five channel-specific drafts from a single brief, with a Telegram feedback loop.

Marketed internally as the "1 brief = 5 channels" workflow. Form + Telegram (/reklaam, PARANDA:) + Schedule triggers converge into one Normalize node. Brand voice rules live in code, not PDF — every campaign update changes AI behavior automatically.

Architecture

3 triggers → Normalize → CSV lookup (products + campaigns + segments + inventory) → AI (Groq llama-3.3-70b) → 5 channel outputs.

Output channels

Telegram / Instagram / Facebook / Email / Product description — all from a single brief in ~10 seconds.

Time saved

~75 minutes per execution, tracked automatically by n8n. Brand voice consistency guaranteed by code-level constraints.

Operational lesson

Credentials are the interface between the workflow and reality — an exported JSON cannot carry that interface across environments.

What I built: A weekly sales-data report workflow that separates deterministic logic (Code node calculates every number) from creative AI (writes the prose around those numbers).

The Code node calculates every number — total revenue, Narva share, stock_risk count, KPI achievement. AI receives these as structured input and only writes prose around them. AI cannot fabricate figures because it never sees raw rows — it sees aggregates.

The Narva discovery

5.96% of revenue against a 12% KPI target — but the problem is not demand. All four stock_risk = high rows came from one combination: Narva × women's clothing × Spring Wardrobe. Supply problem, not demand problem.

Board-level recommendation

Do not finance Riga expansion until Narva supply chain is audited. Not generated content — a finding that survives audit.

Sheets-as-memory

Every report writes a row to the Raportid tab with 11 columns including human_check. Over weeks, the log becomes a searchable decision history.

Business impact

~250 hours/year saved, roughly 12,500€ of director-level cost recovered. Plus compounding value from the decision log.

The brief: UrbanStyle's investor board demanded a defensible AI strategy before approving the Riga expansion (5th store, planned Q4 2026). The CEO had 7 weeks. Weeks 1–5 produced the analysis and three production automations. Week 6 turns all of it into one board-ready deck — and one honest answer.

The answer in one line: AI infrastructure unlocks the Riga expansion by absorbing operational growth without proportional headcount — but only three of ten candidate processes are genuinely AI problems. The rest are either SaaS problems or human-judgment problems, and treating them as "AI" wastes budget.

The three pillars (validated in production)

Pillar 1 — Customer Service

Telegram FAQ bot (Week 4) handles ~75% of repetitive queries in Estonian with hallucination guards. TÄPNE score 20/25. Frees ~120 hours/month across e-shop and store staff. Direct revenue lift via faster response on cart-stage questions.

Pillar 2 — Content Marketing

Hybrid v1.1 (Week 4.3): 3 triggers → 1 logic → 5 channels. Anna's bottleneck — 5 units/week against a need for 20 — closes without 4× headcount. Strategic prerequisite for Riga: brand-consistent content at 4× volume must exist before a new market opens.

Pillar 3 — Data & Decisions

Hybrid v1.0 (Week 5): deterministic Code node computes every number, AI writes the narrative around them. Cuts ~250 hours/year of director-level reporting (~12 500€). Discovered the Narva 5.96% revenue vs 12% KPI gap — a supply-chain problem worth ~60k€/year if fixed.

What AI cannot solve — and why that matters

The discipline of Week 2's TÄPNE scoring carried through. Three high-scoring processes were deliberately not recommended for AI:

• Narva's 2.1% conversion rate — looks like a marketing-AI problem, is actually localization, store hours, and language staffing. AI would mask the real issue.
• Inventory tracking (TÄPNE 22/25) — high score, but solved by a proper ERP, not AI. Labelling it "AI" wastes budget.
• Recruitment & staff scheduling — human-judgment territory. Standard SaaS (Planday) covers scheduling; AI adds nothing.

This is the section investors actually read. A strategy that promises AI will solve everything signals inexperience. A strategy that names what AI won't touch signals operator-level thinking.

The investment case

Annual value unlocked

~60 000€ recovered revenue (Narva supply fix) + ~12 500€ director-time cost recovery = ~72 500€/year recurring.

Protected revenue

~1.4M€ from the Klubi-Maret loyalty segment (15% of customers, 35% of revenue). Churn prevention via AI-personalised retention is the highest-ROI use case identified.

Implementation cost

~15–20 000€ first year: n8n self-hosted, Groq/Claude APIs at SME volumes, setup & documentation. Vendor risk mitigated by n8n abstraction — swap LLM in one node, not a rewrite.

Payback

<4 months, with first measurable savings in month 1 (customer-service hours).

Phased rollout — 9 months to Riga-ready

• Q2 2026 — Foundations: Customer-service bot live in production. Content drafting v1 in human-review mode. Weekly data report runs every Monday.
• Q3 2026 — Scale: Content scaled to 4× output across 5 channels. Daily data reports replace weekly. Klubi-Maret retention campaigns automated.
• Q4 2026 — Riga launch: 5th store opens on AI-ready ops template. Same headcount as 4-store baseline. Strategy validated by year-end metrics.

Risks named honestly

• Customer-facing AI errors — mitigated by human-in-the-loop on outbound content; bot returns "I'll route this to a human" outside its FAQ scope.
• API vendor risk (Groq, Claude rate-limits) — n8n abstracts the model; swap providers without rewriting workflows. Tested as part of Week 4.
• Staff resistance — framed as "freed for high-value work, not replaced." Anna keeps creative direction; bot drafts. Marko keeps analytical authority; AI writes prose around his numbers.

Status: Deck delivered to UrbanStyle's CEO in Week 6 of the 02Signal program. Recommendation: GO on Riga expansion conditional on AI infrastructure milestones meeting the Q2/Q3 gates above. The full executive deck (18 slides, board-ready) is available on request.

The closing moment: a live 30-minute walkthrough to UrbanStyle's CEO Kristi Tamm and the investor board. The strategy from Week 6 is no longer a deck — it's three workflows running on real data, in front of the people who funded the question. This is where seven weeks of work meet the only test that matters: does it actually work?

What's defended: not slides, but production. The board sees a customer submit a question to the Telegram bot and receive an Estonian-language answer in under five seconds. They see a content brief turn into five channel-ready drafts. They see a Monday-morning report arrive in Telegram with the Narva supply finding flagged for review. Three live demos. No screenshots, no "imagine if". Just the workflows.

Watch the demo

The full walkthrough was recorded for three audiences: Estonian for the 02Signal cohort, Russian for the UrbanStyle operations team, English for international review. Same case study, three languages — pick yours and watch the workflows run live.

📚 Source materials & cross-references in NotebookLM

The 30-minute program

0:00 – 0:05 · The question

Recap of the brief: investor board demanded an AI strategy before approving Riga expansion. Seven weeks to deliver one defensible answer.

0:05 – 0:15 · Live demo #1 — Customer service bot

A board member opens @DmitriTsizov_bot on their phone and asks a real question in Estonian. Answer arrives in <5s with citation from the FAQ knowledge base. Then the "hallucination guard" demonstration: ask a question outside scope, bot replies "I'll route this to a human." Trust earned in 90 seconds.

0:15 – 0:22 · Live demo #2 — Content hybrid

Anna Mets fills in a content brief on screen. ~10 seconds later: 5 channel-ready drafts in Telegram (Instagram caption, Facebook post, email subject + body, Instagram Story copy, product description). Then the PARANDA loop — board member suggests a revision, AI rewrites in 30 seconds while keeping brand voice. 75 minutes of work compressed live.

0:22 – 0:27 · Live demo #3 — Data & decisions

Trigger the Monday report manually. AI narrative arrives in Telegram with the headline finding: Narva contributes only 5.96% revenue against a 12% KPI target. Board sees the AI did NOT recommend marketing spend — it flagged it as a supply problem (4 of 4 stock_risk rows from one combination). Decision authority preserved, decision speed multiplied.

0:27 – 0:30 · The recommendation

One slide. One sentence. GO on Riga conditional on the Q2/Q3 AI infrastructure gates from Week 6. No equivocation. Backed by 72 500€/year of unlocked value, <4mo payback, and seven weeks of documented architecture in the GitHub repo behind it.

What the seven weeks proved

• From noise to signal. Week 1 segmentation turned 3000 anonymous customers into 5 actionable segments with one carrying 35% of revenue. The board left knowing where to protect first.
• From manual to automated. Three workflows took ~600 hours/year of repetitive work off Anna, Marko, and the e-shop team. They were not replaced — they were freed to do what only humans can.
• From hype to discipline. Three of ten candidate processes were rejected as "not AI." The discipline of saying no to AI is what makes the yes credible.
• From prototype to production. Every workflow demo'd ran on the same n8n instance Kristi could log into Monday morning. No staging-only theatre. Real credentials, real Sheets, real Telegram. The four production bugs from Week 4's debug log are part of why it survived demo day.

The demo method — why it survived going live

• Live-data demo, not staged. All three workflows ran on real UrbanStyle test data via real Telegram and real Google Sheets. The board could tell.
• Four documented backup plans. One per workflow + a "show execution canvas" fallback. The debug log from Week 4 was the rehearsal — by Week 7 every common failure had a recovery path.
• Honest section before optimistic section. What AI can't solve was presented before the numbers. Investor experience: the people who lead with limits are the people you can trust with capital.
• One-sentence recommendation at the end. Not 18 slides of conclusion — one line, one decision, one direction. Boards remember decisions, not decks.

Outcome & what's next

The 02Signal program closes with pass/fail certification — pass requires 70% session attendance, 6 weekly portfolio artifacts, and the final presentation. Pass achieved. All three production workflows are open-source on github.com/dmitritsizov-cloud/02signal-portfolio-template — JSON exports, documentation, debug logs, and the original prompts. Reproducible, auditable, transferable.

What's next for me: the patterns proven on UrbanStyle.ltd are now a template. Customer-service bot, content hybrid, and data-decision report transfer directly to any Estonian SME with similar bottlenecks. Three pilot clients welcomed at reduced rates — see the services section for how the engagement works.

"From noise to signal. From manual to automated." — 02Signal program motto, made into seven weeks of receipts.