Kavak's Playbook for Rebuilding a Company Around AI

Kavak’s AI-native transformation (business strategy + operating model)

Strategic bet

• Rebuild Kavak into an agent-run company (“agent per customer”), not an AI layer on top of existing workflows.

Core architecture (agent harness)

When a customer enters:

• An agent is spawned specifically for that customer with:
  • its own virtual machine
  • memory of the customer’s history (e.g., prior web visits/calls)
  • a long-term goal: maximize Customer Lifetime Value (LTV) while maximizing satisfaction and conversion into more products over time
• Agents can access company APIs/tools/skills
• Agents coordinate with evals (evaluation harness) and a CLI (command interface)

Organizational design principle

• Build a flat, senior-empowered structure where teams are organized around:
  • Engineering + AI + Operations
  • physical-world execution (e.g., mechanics, field work)
• The “org” becomes collaboration between agents and humans, with tight feedback loops.

---

Frameworks / playbooks implied or explicitly stated

“Evol vs agents” engineering principle (breaks concept)

To move fast:

• Build evals (“brakes”) early and continuously.
• Rule of thumb: spend roughly equal engineering time + tokens + money on building evals as on building agents.

Evaluation & KPI hierarchy (what to measure first)

Start with business outcomes:

• Primary business KPIs (conversion + business value):
  • Did it convert? (buy car, get loan approved, sell car to Kavak)
  • Is the customer happy and willing to re-engage?
• Then optimize the agentic architecture based on those evals.
• Avoid vanity metrics (e.g., number of calls / call minutes).

Success metric shift: transactional → relational

• Old: measure transactions (cars bought/sold, parts volumes, etc.)
• New: measure customer-level relational outcomes using agent assignments (maximize LTV) across millions of customers.

---

Key metrics / KPIs / targets mentioned

Agent coverage at scale

• ~96% of all interactions handled by agents
• ~95% of all transactions handled by agents
• 100–200,000 agents instantiated per day
• Agents handle tasks lasting minutes to days, then go back to sleep

Conversion / performance improvements

• Car-selling agents
  • Tripled NPS and increased customer satisfaction after deployment
  • Initially +50% conversion vs human team
  • Later >2.1x conversion (“over that like 2.1x more”)

Financial product speed

• Car loan approval under ~3 minutes
  • previously could take 2 months or more in Mexico / emerging markets

City experiment (“AI CEO”)

• First month goal: double profits
• Result: 1.5x profits (≈ +50% vs baseline)
• Additional improvements:
  • customer satisfaction
  • better inventory
  • better financing penetration
  • improved daily execution (less micromanagement via forecasts/plans)

Mechanics sidekick outcomes

• Warranties reduced by ~20–26% after launching inspection/repair support (“El Mike”)
• Faster inspections/repairs and cheaper operations while raising quality and satisfaction

Customer economics

• 10 million customers in the database
• Even 1% activation is described as hundreds of millions of dollars, if executed correctly

---

Product + GTM execution examples (what they did differently)

1) Selling expensive, complex used cars (agents as “mega-experts”)

• Problem: Many steps (SKU selection, financing, insurance/coverage, trade-in quoting). Historically required ~15 expert tasks across ~15 teams.
• Agent solution: Build sales agents that can outperform experts across steps and assemble the full purchase package.
• Outcome: Major lift in NPS/satisfaction + >2x conversion.

2) Fast regulated-adjacent financial decisioning (underwriting + servicing experience)

• Example: Car loan product
• Operational advantage: Customers don’t wait weeks/months; approval in under three minutes using relevant data (customer + car)
• Extended strategy for other financial products:
  • customers take 3–4 months to decide
  • a personalized, easy experience improves conversion and retention over that long cycle
  • offers (interest rate, risk pricing, max loan amount) are personalized for:
    • customer conversion timing
    • portfolio-level optimization

3) “AI CEO” as a controlled internal pilot (learning + closed-loop metrics)

• Mechanism: Created a Mexico city pilot (“Guavaka”) and ran an agent as CEO within the agentic harness.
• Learning loop: Agent evaluates decisions, executes plans, receives progress feedback via field-worker voice notes, and iterates.
• Result: +50% profits in first month; improved operational KPIs.

4) Physical-world automation with human collaboration (mechanics “sidekick”)

• Reality: Dexterity/senses are hard to fully automate.
• Tactic: Build “sidekick” agents (“El Mike”) that guide mechanics during inspection/repair.
• Outcome: Faster work, better quality, lower warranties, improved satisfaction.

---

Agent evaluation: how they “get comfortable” (e.g., for money + PII risk)

• Central approach: Tie evals to business outcomes (conversion, happiness/re-engagement) rather than model-process metrics.
• Data capture & training loop:
  • Put agents in front of customers and collect evals/data
  • Agents learn from mistakes, and learning propagates to other agents (shared/continuous improvement loop)

---

Human-in-the-loop design (closing the loop vs “ticketing” and forgetting)

Common failure mode elsewhere

• If an agent hits a wall, it escalates to tier-2 support and the learning loop is not closed—so future performance doesn’t improve.

Kavak’s approach

• A customer-obsessed agent has access to every API/skill
• If stuck, it calls a help API
• Humans intervene to resolve the issue
• The structure is effectively agent-led orchestration with human teams as support, while still generating data for retraining and continuous improvement

---

Advice to founders/operators (business execution recommendations)

• Top-down transformation: AI adoption can fail if only bottom-up (“hackathons” don’t work).
  • Leaders must define a clear 3–5 year target operating model and drive execution vertically.
• Measure what matters, and by agent/evals ROI:
  • They prioritize tokens/compute and agent work by measurable ROI:
    • Tier 3 tokens: direct ROI with measurable outcomes
    • Tier 2 tokens: indirectly measurable improvements (developer/codebase)
    • Tier 1 tokens: “plug-and-play” usage with unclear ROI
  • Don’t just buy tokens—track ROI and iterate.
• Agentic paradigm guidance: Prefer objectives/graphs/functions as tools, and use multi-agent systems only when they don’t constrain new higher-level intelligence.
• “Creative destruction” mindset for incumbents:
  • Incumbents often adopt AI superficially because CEOs can’t justify destroying and rebuilding.
  • New companies can be formed “AI-native,” enabling larger step-changes in productivity/product/service quality.

---

Mentioned presenters / sources

• Ali Masa (Head of AI at Kavak)
• ACC podcast hosts / interviewer (names not provided in the subtitles)