Dave Killeen, Field CPO at Pendo, recently wrote that he had 149 strong ideas on his AI product backlog. He built a custom kanban system with agent-friendly automation just to manage them.

That number isn’t shocking. If you lead a technology team right now, your backlog probably looks similar. Maybe it’s not 149, but it’s more than your team can meaningfully evaluate in a quarter. And the list keeps growing.

The problem isn’t generating ideas. AI solved that. The problem is deciding which ones deserve conviction — the kind of conviction that justifies pulling engineers off something else, spending real budget, and shipping it to customers.

The Cognitive Load Nobody Warned You About

When building was expensive, the cost of building filtered ideas naturally. A feature that took six weeks to build had to justify itself before anyone wrote the first line of code. The economics did the filtering for you.

Now that the development lifecycle has collapsed, building is fast and cheap. An agent can prototype an idea in an afternoon. Which sounds great until you realize what it actually means: every half-formed idea now feels feasible. The filter that cost used to provide is gone.

What replaces it? Judgment. And judgment doesn’t scale the way building does.

A CTO I know described it as “drowning in green lights.” Every initiative on the board has a working prototype. Every prototype looks promising. The dashboards show progress across the board. But progress across the board isn’t the same as progress toward something that matters.

This isn’t a productivity problem you can solve with better project management. It’s a judgment problem. The number of ideas a team can generate with AI now vastly exceeds the number of ideas that team can evaluate with the rigor they deserve.

The Truffle Hunter’s Advantage

Killeen uses an analogy that sticks: in the age of AI, we’ve all become head chefs of our own Michelin-starred restaurants. AI is the kitchen staff. We design the menu, source the ingredients, and take the credit.

But he pushes the analogy further — and this is the part worth paying attention to. He says he’s become less of a head chef and more of a forager. Someone who goes out into the forest, actively looking for truffles. His agents scour GitHub, approaching the frontier from first principles: what is actually being built right now, by people working at the edge?

Anyone can find truffles. A good search agent will surface dozens. But only someone with real experience and genuine taste can bring them back to the kitchen and make the plate sing.

That distinction matters for B2B technology leaders:

More ideas	Better judgment
50 AI use cases identified in a brainstorm	3-5 that survive a rigorous “would a customer pay for this?” test
Prototypes for every department	A ranked shortlist based on actual user need and technical feasibility
A backlog that grows weekly	A backlog that shrinks as low-conviction items get discarded
Broad AI adoption metrics	Concentrated bets that compound

The value of an AI strategy assessment isn’t producing a list of 30 use cases. It’s helping a leadership team walk out of the room with 3 to 5 that they have genuine conviction to execute — and the clarity to say no to the other 25.

A Living Document Instead of a Roadmap

One of the most practical ideas in Killeen’s article is what he calls a Truths Markdown File. It’s a living document organized into three columns: what I believe is true today, what I expect to be true in six months, and what I’m betting on at twelve months out. His AI has access to it every time they work together, grounding every conversation in a coherent worldview instead of disconnected speculation.

It sounds simple. But compare it to how most technology teams plan.

A traditional roadmap gets built during annual planning. It reflects what the team believed in October. By March, the assumptions underneath it have shifted — model capabilities changed, a competitor shipped something unexpected, a key integration became possible or impossible — but the roadmap stays frozen. Budget is allocated against October assumptions. Priorities reflect October conviction.

The Truths File works differently. It’s designed to be updated monthly. When an assumption changes, you change the file. The roadmap stays honest because the document underneath it stays current.

He pushes this further by separating truths into two types. Technical truths — where models, infrastructure, and agent capability are heading. And customer truths — what your users actually need and struggle with right now. The intersection of both is where real conviction lives.

Here’s a concrete example. Eight months ago, fine-tuning a model was a reasonable default for most enterprise AI projects that needed domain-specific behavior. Today, context engineering often replaces fine-tuning at a fraction of the cost. A static roadmap from Q3 2025 would still have fine-tuning budget locked in. A living assumptions document would have flagged the shift by December and redirected the investment.

From Process to Conviction

The traditional product development lifecycle assumed decisions needed committees, approval chains, and sequential gates. Discovery feeds into definition, definition feeds into delivery, delivery feeds into measurement — and at every stage, progress is throttled by meeting cycles and the friction of moving an insight from one part of the organization to another.

That model made sense when building took months and changing direction was expensive. Today it doesn’t.

I’ve already written about what happened to the SDLC. This post is about what fills the vacuum. Not another process — a judgment loop.

Killeen calls it the Conviction Cycle. Still mulling the name, he admits. But the concept is clear: it’s how you orchestrate your thinking, ideate, look ahead, de-risk, and amplify your bets. He compares the role to an orchestra conductor — someone who doesn’t play every instrument but holds the vision, feels the tempo, and brings in the right capability at precisely the right moment.

In practice, the Conviction Cycle is three steps that repeat continuously:

1. Write your assumptions as testable statements. Not strategy documents. Short, specific claims: “Context engineering will replace ~60% of our fine-tuning use cases within 6 months.” “Our largest client segment will ask for agent-based workflows by Q4.” Write them down. Make them falsifiable.

2. Stress-test against real signals. Run a small experiment. Talk to five customers. Check what’s actually shipping on GitHub. Killeen’s agents do this for him — scouring open source repos to answer “where is the puck going?” The point isn’t to validate everything. It’s to kill the ideas that don’t survive contact with reality.

3. Commit or discard. This is the hard part. If the stress test supports your assumption, invest with real conviction — budget, headcount, attention. If it doesn’t, discard the idea without guilt. The discipline to let go of a plausible-sounding idea that failed its stress test is what separates teams that ship from teams that stay busy.

The Cost Curve Nobody’s Ready For

Killeen’s most provocative claim: when agent orchestration for coding reaches real confidence and efficiency, small AI-native teams won’t just compete with large organizations. They’ll undercut them catastrophically.

He has firsthand experience with this pattern. When the Daily Mail launched MailOnline, the paper was generating around £2 million of revenue per day. There was genuine boardroom nervousness about self-cannibalization. What followed is now obvious in hindsight: the era of well-funded, high-budget journalism was quietly ending. The disruption didn’t announce itself. It just made the old cost model untenable.

The same dynamic is coming to software. Large organizations carry enormous cost infrastructure — headcount, real estate, legacy tooling, organizational complexity. A small team with strong context engineering and none of that overhead can deliver equivalent value at a fraction of the price.

In 25 years of building enterprise software, I’ve watched this pattern repeat — mainframes to client-server, client-server to web, on-premise to cloud. Each wave rewarded the teams that built conviction about where to invest before the market forced their hand. What’s different now is the speed. The window between “this is an interesting trend” and “this is the competitive reality” has compressed from years to months.

The question isn’t whether this happens. It’s whether you’re building conviction now or waiting until the cost curve forces the decision for you.

Three Things You Can Do This Month

Write down your assumptions. Create a simple document with three columns: what you believe today, what you expect in six months, what you’re betting on at twelve months. Review it monthly with your technology leadership team. Killeen credits this practice — his Truths Markdown File — as one of the highest-leverage things he does with AI. Every conversation is grounded in a coherent worldview instead of starting from scratch.

Audit your AI backlog for conviction level. If you have more than 10 AI initiatives in flight or planned, force-rank them. Apply the 10x filter: if the next model generation is 10 times better, does this initiative still matter? Investments in clean data, solid integrations, and model-agnostic architecture survive that test. Investments that compensate for current model limitations don’t. Kill the bottom third of your list.

Assign a forager. One person — or one agent workflow — whose job is to scout what’s actually being built at the frontier. Open source repos, new model capabilities, competitor moves. Not a research team producing quarterly reports. A single feed with editorial judgment applied, delivering the 3 most relevant signals each week. The forager who also happens to be a domain expert — that’s the new superpower Killeen describes.

---

If your AI backlog has more ideas than conviction, that’s the gap we close. At IQ Source we run a focused AI strategy sprint — two sessions, one week — where your leadership team writes testable assumptions, ranks current AI bets by durability and alignment, and walks out with a filtered shortlist you actually believe in.

Book an AI strategy sprint