The Ideation Engine: How I Used AI to Generate Truly Novel & Feasible Ideas at GEICO

I joined GEICO as an Engineer and Evangelist to help my old boss Paul Devitt, PhD run an innovation lab, drive engineering productivity using AI tools, and write posts like this spreading the simple news that GEICO Tech is a place where the best engineers are given the best tools available to solve fascinating technical problems.

In Part 1, I built a knowledge base from 147 meeting notes. In Part 2, I used AI to hallucinate responsibly (and discovered 1 product idea worth incubating).

The Problem: Truly Novel Ideas Aren't Cheap

After writing our Claims Lab charter, I embarked on generating some novel ideas based on what I'd built up thus far in my new fancy KB (see Part 1).

Don't hate me, product people - I skirted the traditional brainstorming process - in my defense: it sucks. The outputs either stays too grounded (incremental tweaks) or go too wild (impossible moonshots). For the case of the Lab, I wanted a way to generate ideas that were novel but feasible - that is: grounded in GEICO's reality but not constrained by current thinking.

The Solution: 🔱 3-Phase Ideation Pipeline

Ultimately, I settled on building a three-phase experiment in which I employed AI agents using different AI models in different cognitive modes (dreamer, interpreter, critic) and orchestrate them like an ensemble. The flow: Dream → interpretation→ decision.

This was partly inspired by Margaret Boden's framework for computational creativity: transformational (Phase 1: break conceptual spaces), exploratory (Phase 2: explore within established space), and combinational (Phase 3: combine familiar ideas into practical solutions), plus informed by some idea-to-product questions I learned from my friend Luke Rabin.

The following is a breakdown of each step in the flow! Follow along in your own Cursor IDE, if you dare. 🤠

Phase 1: The Dreamer 🦄

Goal: Generate completely unconstrained ideas in 15 minutes.

I prompted my local knowledge base in a deliberately unhinged prompt:

Forget everything you think you know about auto insurance and claims. Solely based on this KB, generate a wild hallucination from ten conflicting angles that form a holistic view of GEICO's claims tech ecosystem.

For this step, I wanted maximum divergence. Research shows LLMs produce 2-3x more ideas than humans with higher originality scores, making them ideal for the divergence phase. I used **Sonnet 4.5**from Anthropic because it tends toward verbosity and creativity, with a little bit of reasoning. Also, it was the only model I could get API access to during my first month at GEICO (we take security seriously).

What happened: The model went on a hardcore poetic riff about claims as "empathy orchestration systems" and adjusters as "fairness architects", dove into various metaphysical angles (e.g. claims as trust networks, as grief processing, as collaborative storytelling between strangers, etc.), and returned a behemoth 12,000-word document containing gems like this:

What if every claim decision could explain itself in the claimant's own language, not in policy-speak?

I saved the whole thing as WILD_HALLUCINATIONS.md. Time for phase two.

Phase 2: The Interpreter 💡

Goal: Extract structured ideas from the chaos in 20 minutes.

I switched contexts completely. New repo, fresh chat, no grounded knowledge base to distract us. I copied in my WILD_HALLUCINATIONS markdown file and gave Claude Haiku 4.5 a more verbose version of this prompt:

This WILD_HALLUCINATIONS.md document is a prophetic brain dump from GEICO's future CEO, transported back from a world where claims embody peace and harmony. Extract the philosophical struggles. Generate ten business ideas that sound insane today but might be brilliant.

**Why Haiku? **It's fast, cheap, and good at structured extraction. I didn't need creativity here—I needed organization. This detail is important.

💣 What happened: It produced a future-oriented document PHILISOPHICAL_STRUGGLES_IN_FUTURE_OF_CLAIMS_2025.md, which surfaced 3 core tensions:

Automation vs. Empathy: Can AI adjudicate fairly while maintaining human care?
Transparency vs. Privacy: How much should we explain without revealing proprietary logic?
Efficiency vs. Justice: Can we process claims fast without sacrificing fairness?

Then it generated 10 product ideas in IDEAS.md. Here a sample of three:

"Plain Language Claim Explainer": AI that translates policy decisions into conversational language claimants actually understand
"Empathy Training Simulator": LLM-powered role-play scenarios for adjusters to practice difficult conversations
"Blockchain Claims Ledger": Immutable record of every decision for full transparency (this one made me LOL)

Some ideas were wild. Some were practical. The **next step: filtering **them down.

Phase 3: The Critic 🔥

Goal: Filter for feasibility and fit in 30 minutes.

Now came the crucial step. I again loaded my full knowledge base (all 147 meeting notes + AI-produced meta-analysis), copied in IDEAS.md, and prompted**Sonnet 4.5** (best available at critical reasoning in 2025) with a more verbose version of:

You are a CSV generating idea-grading machine. Using this full GEICO Claims Lab knowledge base as reference, evaluate each of these 10 ideas for technical feasibility, organizational fit, and strategic viability. Score 1-10 on each dimension. Be brutal.

The output: CSV that rendered like this (sampling 3 of 10 records):

(An image appears here in the original article on LinkedIn.) **As the human in the loop, I interpreted the results. **My conclusions:

Plain Language Explainer passed because it leverages existing LLMs, aligns with customer transparency initiatives, and addresses real customer pain.
Empathy Simulator passed because LLM role-play is proven, addresses training bottlenecks for new adjusters, and has measurable impact.
Blockchain Ledger failed spectacularly: blockchain adds no value here, regulatory nightmare, solution looking for a problem.

Result: 10 ideas entered. 8 documented failures. 2 candidates. 1 really worth presenting to my boss.

Punchline: That 1 surviving idea is now in our Claims Lab incubation pipeline for Q1 2026. 😎

This multi-agent pattern works: Anthropic's research shows orchestrated agents outperformed single-agent systems by 90.2%, with 90% reduction in research time. Similar critic-generator loops show +7-13% improvements across 3-5 refinement iterations.

Let's recap what just happened!

This wasn't prompt engineering. It was orchestrating three cognitive modes:

The Dreamer (GPT-OSS-120B): Maximum creativity, no constraints. Generate possibilities without judgment.
The Interpreter (Haiku 4.5): Extract structure from chaos. Find patterns, tensions, and concrete ideas.
The Critic (Sonnet 4.5): Apply reality. Score against actual constraints from the knowledge base

Together: dream → interpret → decide. Here's the full flow:

(An image appears here in the original article on LinkedIn.)

Why This Matters

Most companies use AI to accelerate existing workflows: write faster, code faster, analyze faster. We love this! However, there's another interesting opportunity: cognitive expansion. That is: using AI to think in ways you couldn't before, not just execute faster on ideas you already have.

Traditional brainstorming generates either incremental tweaks or impossible moonshots. My fancy new process generates ideas that are novel but feasible, both grounded in reality AND unconstrained by old thinking.

Major consultancies are proving this at scale: McKinsey completed 400+ GenAI projects for clients using their Lilli chatbot (synthesizing 100 years and 100,000+ documents). BCG built 18,000+ custom GPTs used by 33,000 employees for ideation.

Adding Up Cost and Time

The budget on this type of innovation is negligible: 65 minutes (15 + 20 + 30) + ~$12 in API calls (~400K tokens across all phases). This easily beats a buying a few hours with a high-quality product team, which would equate to $2,000+ assuming a HR cost of $1mm/year.

But don't get caught up in the hard costs; the real savings weren't monetary! Instead of scheduling meetings, building consensus, or navigating group dynamics, I just applied structured thinking across three cognitive modes. I went from zero to ideation before my coffee went cold!

If you're still skeptical, these gains align with industry benchmarks: BCG reports 40% productivity improvement on repetitive tasks and nearly 90% of participants experiencing enhanced creative ideation. P&G achieved 40% reduction in product development time with 25% increase in new product launch success using AI-driven ideation.

The Bottom Line

Different AI models have different cognitive styles. Orchestrate them like an ensemble, not solo performers. 🎶

**What do you get? **In my case,two incubation-ready product ideas that wouldn't have emerged from traditional brainstorming (plus eight documented failures that teach us what doesn't work).

Please, let me know how you're using AI to expand thinking! This is a topic I'm very passionate about, so I'd love to meet for coffee to discuss. ☕️

Also, **stay tuned **for more posts about the thorny distributed systems problems we're tackling at GEICO (and of course reach out if you want to learn more about joining the team).

#GEICO #GEICOTech #BuildWithUs