When building got fast, product strategy got skipped. Not because it stopped mattering. Because it had not adapted to the speed. This is what adapting it looks like.
Agentic coding tools (Cursor, Replit, Lovable, Claude Code) have fundamentally changed what it costs to ship software. A solo founder today can build in days what used to require a team and months. The cost of execution has collapsed.
But the cost structure of product decision-making has not moved. A bad strategic assumption still produces wasted build. A vague idea still becomes an expensive wrong turn. The only difference is that now it happens faster, and the mistake is cheaper to ship than it ever was to avoid.
AI does not just accelerate good decisions. It accelerates bad ones at the same rate. Conservative estimates for 2024-2026 put the minimum economic impact of strategic misalignment at 10-30% higher than the pre-AI baseline. Wrong assumptions now get embedded in working code before anyone catches them.
The rational response was to skip strategy and iterate. Build cheap, learn fast, adjust. But iteration at AI-speed without strategic anchors does not produce learning. It produces variance. Features that conflict with each other. Products that do many things for no one in particular.
This is not a niche problem. Strategic thinking is a global industry. It has not yet been restructured by AI.
This is what the world currently pays to think strategically. Most of it goes to consultants, workshops and documents. Very little of it is connected to the tools actually doing the build.
"I'm finding Pabs the best tool I'm using to understand what I need to do. This is different from ChatGPT that tells you what you want to hear. Pabs challenges you. Now that I know the MCP and it can be my strategic companion, Pabs is super valuable."
Sean, Solo Founder · workflow fit 5/5 · would pay · essential score 5/5"I don't think that I will be a great tester for you as I won't be using any of the coding or building aspects. It has just happened to align with a couple of things over the last couple of weeks. I unfortunately don't think I will be enough of a core user."
Hunter, Entrepreneur exploring ideas · sporadic use · would not payBoth found real value. Both gave positive feedback. One converted. One did not. The difference was not the product. It was where they were in their build journey when they used it. Sean was building. Hunter was thinking. That distinction is what the pilot was designed to find.
The pilot surfaced something more useful than engagement metrics. Users were not distinguished by how active they were on Pabs. They were distinguished by where they were in their build journey when they needed it. Three distinct moments. Three different problems. One of them drives real payment intent.
The old value: "I will do discovery and create a strategy doc for you." The new value: "I will make your assumptions visible before they become expensive." It's not about comprehensive plans. It's about surfacing the 3–5 decisions that prevent wrong turns. A Darwinian strategy that stays alive.
A strategy doc that lives in Notion and never gets referenced is worthless. A Decisions View that feeds into Replit via MCP and actually changes what gets built. That's priceless. I watched Pabs agents tell Replit agents not to build a feature because it wasn't in the core strategy. That changed everything.
Pre-build explorers got value from Pabs but rarely converted. Post-build reflectors converted after pain. In-build operators converted immediately. For them, a wrong decision mid-session embeds in code that is expensive to reverse. That is the only moment where strategic thinking competes with speed rather than asking people to slow down for it. Pre-build planning produces a document. Mid-build alignment produces a different product.
Pabs wasn't designed to be an ecosystem. It was an experiment. But solving for workflow fit required agents, an interface, MCP, export artifacts, confidence scoring. Each piece solved a workflow integration problem. Together they became an ecosystem. You can't plan your way to reinvention. You have to build, ship, and learn.
The moat is integrations and data, not the underlying AI. The winners across AI infrastructure aren't competing on AI quality. They're building the orchestration layer that makes AI work together. The right framing isn't "better strategy documents from our AI." It's "orchestrate strategic decision-making across your entire stack."
Fifteen months of questioning the category. One month to build and ship the MVP. Four months of the product learning from real users. Total tool cost: $1,320, against an estimated $53-115k for the same disciplines done traditionally. That is not a marginal productivity gain. That is a structural shift in who can build, what they can build, and what it costs to get it wrong.
For small businesses and solo ventures, the barrier to shipping a product is now effectively gone. The new risk is not "can we build it?" The new risk is "are we building the right thing, fast enough to matter?" For enterprise teams and venture-backed products, the same shift applies at higher stakes: AI-assisted teams are shipping at a pace that outstrips traditional strategy reviews, governance cycles, and product committees.
The category this experiment points to is not another AI productivity tool. It is the layer that keeps strategic intent connected to agentic execution. When the build moves at AI speed, the decisions move with it. That is what Pabs is. And that is the market it is entering.
Next
Fifteen months. No traditional team. No agency. No co-founder. Only the strategic experience to direct AI, challenge what it produced, and decide what not to build. The finding: AI executes almost anything you ask. The expensive part is knowing what not to ask. This is the product that came out of that answer.
Each of these traditionally requires a separate specialist, agency contract, or full-time salary. The model doesn't require owning every discipline. It requires knowing enough about each to direct, evaluate and reject what AI produces.
Discovery · Research · Product strategy · Brand · UX design · Prototyping · Visual design · Frontend · Backend · Infrastructure · QA · Content · Marketing · Product management
Agencies. Studios. UX researchers. Product designers. Strategy consultants. These businesses existed because execution was expensive. Someone had to do the research, the design, the code, the strategy deck. AI moved the cost of that work toward zero. The question is not whether AI is useful. It is where professional value actually lives now.
A document. A workshop before the sprint. A process that assumed the build had a beginning. In a workflow where building never really stops, that model does not hold.
One that stays active through the build, adapts as the work generates evidence, and keeps direction visible to every tool in the workflow. It does not require stopping to update. The build and the strategy run together.
One person. 40x cheaper than the traditional equivalent. One month to build the product. Four months of AI adapting it from real user feedback. The capital cost of building is no longer the constraint. Strategy is. That is what this experiment was designed to test.
Every one of these phases ran alongside a codebase accumulating technical debt at the same pace it was gaining features.
AI builds fast. It does not refactor. At several points the roadmap had to stop, address the foundations, and decide what was safe to build on next.
That judgment, when to pause new work and fix what exists, was always human. Pabs held the context for why architectural decisions were made, which made those calls faster.
But the calls still had to be made.
AI tools will build almost anything you ask. That is the problem. The Pabs MCP connector is what makes them stop and check first. Before building, they read the live strategy. When something conflicts, it surfaces. When a learning emerges mid-session, it travels back. At every decision point, the human decides what moves. Not the AI tool. These are two conversations that prove that shift is real.
Input
Brief or idea
Where every project starts.
Pabs
Strategy layer
Shape. Commit. Adapt. The live source of strategic truth.
Builders
Cursor · Replit · Claude Code
Context flows down. Learnings travel back.
Nine tools to read context. One bidirectional tool to query alignment and report learnings back. When a learning is approved, it writes directly into the strategy. The next time an AI tool reads the context, the strategy has already moved. That is the loop.
Internal coordination
Shared queue
One pipeline, all agents
All agents read and write to the same task queue. A Chat Agent can hand off to the Engagement Agent mid-conversation without losing context.
Visibility
Every action is tracked
What ran, what is pending, what completed. Nothing happens invisibly. The Jobs view shows the full picture in real time.
Human authority
Humans decide what moves
Agents propose. Humans approve, reject or edit. Strategy only updates after a human decision. AI has no unilateral write access.
Execution
Agents that actually act
Agents do real work: web research, strategy edits, email. Not suggestions in a chat window. Actions with outcomes.
Also
Next
The Pabs pilot was designed to find out. Not in theory. In production. 46 real users, 3 months, $17.73 in AI API cost, and every token counted.
Not a curated sample. A real spread of the people Pabs needs to work for. The signal strength varied dramatically by segment.
The most revealing data wasn't project creation. It was what happened after. Who went deeper, what they used, and what they fed back into the system.
Users who generated strategy at least once tended to keep going. That first generation was where a rough idea became structured strategy with risks, hypotheses and decisions attached. For the users who reached it, the product clicked.
Six users found the full loop. One ran 18 market analyses and 28 conversations around a single project, feeding learnings back from real build evidence. Another iterated across 14 projects. A third pasted user interview transcripts directly into the learning engine and described it as a light-bulb moment. That is the product working as intended.
Pilot statuses: 14 completed · 28 active · 4 dropped
The cliff between project creation and first strategy generation was the sharpest drop in the entire funnel.
The pilot hypothesis: in a world where AI makes building instant, a structured strategic pause is worth more than the time it costs. The question was whether users would take it, and whether it changed what they built.
The data shows the pause is almost entirely human thinking time. The AI part is nearly instant. The real cost is 28 minutes of structured input. What follows takes under 2 minutes.
Traditional equivalent
Discovery workshops, stakeholder interviews, competitive research, strategy documentation. Before a line of code is written.
Was the pause worth it?
Users actively building when they used Pabs showed the highest workflow fit, highest ratings, and the strongest payment intent. The pause paid off when there was something to build against.
Where the tokens went revealed what to optimise next.
That is not overhead. That is the product working. 6.2 million tokens of AI sitting alongside a user, reading their live strategy, responding to decisions in real time, adjusting direction on-demand. The conversation layer is where the strategy stays alive between generations. It is the always-on strategic support that used to require a person in the room.
The actual generation events (brief, hypotheses, competitive analysis, decisions) accounted for less than 1% of all token spend. Creating structured strategy is extraordinarily cheap. The cost is in keeping it alive: the conversations, the pulse check-ins, the learning evaluations running alongside an active build. That is where the value is. That is also where the tokens go.
Model selection was itself a strategic decision. Each task type has a different performance requirement and a different cost ceiling. Matching them deliberately kept output quality high and token spend low.
| Model | Calls | Tokens | Cost (USD) | Why this model |
|---|---|---|---|---|
| gpt-5.3-chat-latest | 288 | 7,742,506 | $14.68 | Conversation: highest reasoning quality needed |
| gpt-5.4 | 106 | 852,118 | $2.73 | Strategy, Pulse, Learning eval: structured analytical depth |
| gpt-5.4-mini | 8 | 89,206 | $0.07 | Lightweight Pulse: consistent, lower cost per nudge |
| gpt-4o | 18 | 78,920 | $0.25 | MCP queries: stable API, well-tested for tool calls |
| gpt-4o-mini | 6 | 1,050 | <$0.01 | Actions: fast, cheap, right for simple task execution |
Conversation needs reasoning and memory. Strategy generation needs structured analytical depth. Pulse needs consistency. Actions need speed. Treating every task with the same model would have tripled cost with no quality gain.
The strategic thinking users came for is less than 1% of actual token spend. The real cost is conversation context overhead. Knowing this changed how the product loads and prioritises AI calls.
We tracked the AI token cost of every strategic action across the 3-month pilot: 76 strategies, 62 market analyses, 102 conversations, 40 learnings evaluated, 116 scheduled reviews and 70 agent tasks. 466 outputs in total. Total AI token cost: $17.73. That is $0.04 per output. This is what it cost Pabs in tokens to serve, not what users paid.
Getting from a rough idea to a structured strategy with market analysis, decisions and an action plan cost $0.31. From there, scheduled reviews ran every week and observations from the build fed back in for evaluation. Each action had a measurable cost.
A month of keeping one active project alive ran about $0.20. The first month including setup came to around $0.51.
A freelance product strategist on a monthly retainer to do the equivalent work runs $3,000-5,000 per month.
Users access Pabs through a credit system. Credits are priced based on these token costs plus operating margin. Each action type has a defined credit cost. The first complete setup uses around 16 credits. Keeping one active project alive runs about 10 credits a month. The free tier includes 30 credits.
40 learnings reached the strategy. Pabs evaluated each one, identified what it affected, and proposed what should change. Humans reviewed every proposal. 28 were approved and written back. That 70% approval rate is a trust signal in both directions: humans trusting the AI's judgment, and the AI correctly reading what was critical enough to act on.
How learnings arrived
The system surfaced most of them. Nobody had to remember to submit.
What humans decided
The 10% rejection rate proves the human was genuinely reviewing, not rubber-stamping.
What changed in the product because of those learnings
Activation flow rebuilt
The drop-off before first strategy generation was the sharpest cliff. Onboarding was redesigned around getting users to that first moment faster.
Selective context loading
Token data revealed the chat model was loading the full project context on every turn. Context loading became targeted, not total. Immediate cost and performance improvement.
Multi-plan workflows prioritised
Consultants asked about running multiple client strategies immediately. The first commercial signal shaped the next build cycle.
All participant names are pseudonymised with permission. Token counts and costs are drawn directly from production database records. Learning counts from plan_learnings table. Agent task counts from agent_tasks table. Pulse counts from plan_pulses table. Data sourced May 2026.
Consultants showed early commercial instinct: they asked about client use immediately. But they didn't yet have a workflow that Pabs could anchor to. Innovation and Growth segments showed almost no conversion signal. The data doesn't suggest chasing them next.
The pattern was consistent: value followed workflow fit, workflow fit followed an active build. The product doesn't help people think about building. It helps people who are already building think more clearly.
The model behind the data
Pablo Dunovits
Research published May 2026 · pabs.co.nz