I Wrote the Perfect Spec. Nobody Wanted the Product.

A product manager at a desk planning an AI product: context policy, capabilities as components, and an eval suite

What 20 years in enterprise tech, product and two years building Supernila an agentic AI platform taught me about the one artifact that actually matters.

Josh Elman, a Partner at Andreessen Horowitz (original article: https://x.com/joshelman/status/2099498819914551344) wrote a piece last week called “Product Management is still about telling stories.” In it, he describes an interview with Reid Hoffman who asked him a question he never forgot:

“What is the artifact that a product manager produces?”

Engineers produce code. Designers produce the visual system. Business development produces signed contracts. What does a PM produce?

Elman answered “the spec.” He later decided that was the wrong answer. The artifact is the story.

I read that and felt something between recognition and embarrassment, because I spent most of my career giving the same wrong answer, and I have the documents to prove it.

My 120-page spec

I spent 11 years at a Fortune 500 pharma company, moving from senior principal engineer to solution architecture to director of IT for marketing and sales. That world runs on artifacts. Business cases. Requirements documents. Architecture assessments. Solution designs that go through medical, legal and regulatory review before a single word reaches a customer. I got good at producing them. Some of those programs changed how the company engaged its customers.

Then I left to build my own product, and I brought the habit with me.

The first version of what is now Supernila was an agentic AI platform for small business growth. I specced it properly: multi-platform publishing, a content studio, an analytics dashboard, campaign automation, performance optimization. Java and Spring Boot, because that’s how my whole engineering career was built. The requirements were complete. The architecture was clean.

It had one problem, and it wasn’t technical.

I had built performance optimization features for businesses that had no performance data to optimize. A new customer would sign up, open the dashboard, and find an empty room with very sophisticated furniture. I’d solved the interesting problem and skipped the first five minutes of the user’s life. In the industry we call this the cold start problem. On the ground it just looks like someone logging in once and never coming back.

The spec was right. The story was missing. Nobody ever asked me for a spec after that.

The sentence that broke it

I went and talked to owners. Not a survey. More than 200 conversations with people who run local service businesses: painters, cleaners, spa owners, fitness studios.

Somewhere around conversation forty, a cleaning business owner said the thing that reorganized the entire product:

“I don’t want software to run. I want customers.”

That is the whole story, in nine words. Everything I had built assumed the owner would show up, log in, learn my interface, interpret my charts, and take action. But this is a person who is on a job site at 7am, quoting at noon, invoicing at 9pm. The interface was the tax I was charging for the value.

Once I had that sentence, the roadmap wrote itself. Not “what features do we add,” but “what work do we take off this person’s plate.” Today that’s lead generation and lead management. Next is scoping, estimates, invoicing and payments. After that, the commercial operations of the business.

The spec describes what the system must do. The story describes why someone lets it into their life. They produce very different products.

What AI changed

Elman’s sharpest point is that the cost of making things has collapsed, but the cost of judgment hasn’t moved at all.

I can confirm the first half. I rebuilt the platform greenfield, moved the stack, and now build with Claude Code, Codex and Cursor across a multi-tenant architecture: TypeScript and Python services, LangGraph for agent orchestration, Inngest for durable background jobs and human-in-the-loop approvals, Twilio for messaging, Postgres and Supabase underneath. Work that used to be a quarter is now a long afternoon. The old loop of spec, scope, cost, then build has inverted for me too: build it, play with it, then decide what it actually is.

The pivot that saved the product came out of exactly that. Instead of specifying the fix for the cold start problem, I built a version where the owner states a goal and the agents produce a plan from nothing. Playing with it for an hour told me more than a month of documents would have.

Now the second half, the part I got wrong.

When building is cheap, you build things you shouldn’t. I shipped dedicated pages and positioning for plumbing and HVAC because it took an afternoon and the market logic looked sound. Months later: no customers in either vertical, and a product story diluted across four trades instead of sharpened on the ones that were working. That was AI slop in product form. Not bad code. Bad judgment, executed efficiently.

Deciding what to build stopped being a resourcing debate the moment building got cheap. It became an impact debate: this or that, not this or nothing.

The part that didn’t change

Elman’s framework for product vision is purpose, core actions, and cycle. It’s the most useful thing in his piece, and it’s where agentic products get interesting, because the answers are strange.

Purpose. The owner brings Supernila into their life so that opportunities stop leaking while they’re up a ladder.

Core action. Here’s the strange part. For most products, the core action is the user doing something in your app. For us, the core action is the owner approving something the agents already did, usually from their phone, usually in one tap. An agent captures a lead from the site, qualifies it, responds in under 60 seconds, and pushes the owner a decision that takes three seconds to make.

Cycle. A few times a week. Not daily.

That last one matters more than it sounds. Every instinct in SaaS pushes you toward daily active use, and every instinct is wrong here. If an owner is logging into my product every day, I have failed. I’ve given them another piece of software to operate, which is the exact thing they told me they didn’t want.

So the design goal became: be correct when you’re not being watched, and be unmissable in the three seconds you are. That single decision drives the architecture. It’s why we own our own orchestration layer rather than renting a generic one. Tiered autonomy, approval gates, escalation rules and per-business memory are not features bolted onto a model. They are the product.

“Are people really using it?”

Elman has a test he applies to founders: are people really using your product? Most answer with signups, or ARR, or tokens per day. None of those answer the question.

Here is my honest scoreboard, from production as of this week, across paying customers in cleaning, painting, spa, fitness and retail:

And the most instructive number in the table isn’t any of those.

One customer shows 41 leads and zero conversions. On paper that’s my worst account. In reality that business converts people in person, at a trial class, and nobody was ever going to type the outcome back into a system. The product captured the leads. My data model never asked the right question.

If I had read that row as a dashboard, I’d have concluded the product doesn’t work for that vertical. Reading it as a story tells me something different: my instrumentation doesn’t match how that business actually closes. Exclude that account and the same product converts at 31%.

This is the part of the job AI has made easier and more dangerous at the same time. Because our product talks to people, I have transcripts of real homeowners describing what they want, in their own words, at the exact moment they want it. It’s the richest research input I’ve ever had. It’s also very easy to hand to a model and ask for a summary. Read them yourself. The summary gives you the pattern; the transcript gives you the sentence that reorganizes your roadmap.

The 75 open leads are the loudest thing in that table. More than half of every opportunity we generated is still sitting there, unresolved. That, not any new capability, is what I’m building next: automated follow-up on stale leads, plus nudging owners to tell us what actually happened. The data told me where the value leaked. Judgment decides what to do about it.

Onboarding is where the story gets told

Elman calls onboarding the most important moment you have to tell your story, and he flags the problem I think about constantly: the blank prompt box. “Hi, ask me anything!” is a magic box with no story attached. Most people are not at their most articulate the first time they meet one.

We have two onboardings, and both start by refusing the blank box.

For the homeowner, the agent opens the conversation. It knows the business, the service area, and what this trade actually needs to know to quote a job, so it asks. The visitor answers questions instead of inventing prompts. Nobody has to learn what to type.

For the owner, the rule I set early was that setup could cost them exactly one thing: a single DNS record. Not a migration, not a data import, not a “white-glove onboarding call” that’s really a euphemism for a week of work. Everything else, we do. Every step I can remove from the first hour of an owner’s experience is a step that can’t be the reason they quit.

Elman’s distribution applies exactly: the eagers will get through anything, the fly-bys were never staying, and your real work is the curious middle. Local business owners are almost all curious middle. They’ve been burned by an agency, they’re skeptical, they’ll give you one evening. Tell the story in that evening or you don’t get another.

The artifact

If someone asked me today what artifact I produce, I wouldn’t say the spec. I’ve written those, and the best one I ever wrote described a product nobody wanted.

The artifact is the story. Nine words from a cleaning business owner: *I don’t want software to run, I want customers.*That sentence has survived a stack migration, a repositioning, a pivot, and every roadmap argument I’ve had with myself since. It decides what gets built and what gets deleted.

AI has made me faster at everything except this. It writes my code, runs my agents, summarizes my data, and drafts my documents. It cannot tell me which nine words matter, and it cannot tell me whether people are really using what I built.

That’s still the job. It might be the only part of the job that was ever really the job.


Thanks to Josh Elman for the piece that prompted this one.

I’m Arun Thachi. I build Supernila, an agentic AI platform for owner-led home service businesses, and I write about what building it teaches me about AI product decisions. If you’re working on agentic products, I’d like to compare notes on how you’re defining core actions when the agent takes most of them.