The brief was simple to say and hard to do: build and sell an autonomous agentic AI — an agent plus a dashboard — that can handle 80% of what a B2B renewals account manager does all day.

Not a copilot that nudges. Not a tracker that reminds. An agent that reads the whole account, decides what to do, and does it — while the human stays in control. Three weeks to go from a blank page to something a real revenue leader would put their name on as a design partner.

Here's how it actually went.

The Goal, Stated Plainly

Renewals are where SaaS revenue quietly lives or dies. Net Revenue Retention is the number boards scrutinize every quarter, and every deal counts toward it — not just the whales. But the economics of coverage break down fast: as you move from enterprise into mid-market and SMB, one rep is suddenly responsible for hundreds or thousands of accounts. Somewhere between 20% and 50% of revenue sits in that long tail with almost no visibility.

You can't fix that by adding headcount. You fix it with an agent that owns the renewal book end to end. That was the goal. The rest was execution.

Week 1 — Learn the Pain Before You Build the Fix

I spent week one not building. I ran customer discovery calls and did deep research on where renewals outreach actually breaks, then packaged it into a deck defining the problem and what the solution should look like.

The insight that shaped everything: outreach fails for structural reasons, not effort ones.

• Wrong timing. The average account gets its first renewal touch ~47 days out. The optimal window opens at 90–120 days. A huge share of deals start too late to influence the outcome at all.
• Wrong person. Roughly 28% of buyer-side champions leave every year, and CRM contact data decays about 30% annually. When the champion is gone and nobody noticed, the account quietly goes at-risk.
• Wrong framing. Most reps open with "are you renewing?" — an interrogative that invites a "let me think about it." An assumptive open performs dramatically better.
• Single-threading. A majority of reps contact exactly one person per account. Won renewals average three-plus contacts; lost ones average barely one.

From there I designed the outreach agent as a five-step loop: detect the signal, validate the champion, build a language profile from the account's past emails, generate an assumptive first-touch draft, and drop it in an approval queue (auto-send for the clean SMB happy path). The step I cared most about was the language profile — ingesting a customer's own email history so the draft sounds like a rep who's managed the account for two years, not a mail merge. That's what makes personalization survive scale.

Week one output: a shared, evidence-backed picture of the problem and a concrete spec for the fix. No code yet — and that turned out to matter.

Week 2 — Build Fast, Then Watch It Collapse Under Its Own Weight

Week two I built the agent, then combined it with what my co-interns had built into one full dashboard. We Frankensteined the pieces together until the thing actually worked end to end — which felt like a win.

It wasn't, quite. By the end of the week the dashboard was bloated. Every feature we'd bolted on had left a seam. It ran, but it was the kind of "runs" that gets slower and scarier to touch every day. Adding the next thing meant fighting the last three things.

The uncomfortable realization: the fastest way forward was backward. Instead of nursing the Frankenstein, we'd be better off rebuilding the dashboard from scratch — this time from a proper spec / PRD, designed as one coherent product instead of assembled from parts.

So we killed it. A new dashboard rising from the ashes of the old one — a phoenix from the pile we'd just torched.

Week 3 — The Phoenix, and Ten Demos a Week

Week three we rebuilt the dashboard clean, spec-first. Because the PRD did the arguing up front, the second version came together faster and held together better than the one it replaced. Building right turned out to be faster than building twice.

Then came the part that actually sharpens a product: selling it. I ran ten demos a week. Each demo was a live test of whether the story held up in front of someone who does this for a living, and each one sent me back with something to fix. Ten demos, ten iterations. The dashboard I ended the week with was shaped by real reactions, not my assumptions.

By the end, the product told a full six-stage renewals story — the agent working an account from Qualification → Outreach → Proposal → Buyer Evaluation → Negotiation → Close, reading the account 120 days out instead of the usual 30, with the human in the loop to approve every consequential action. It's positioned as a new teammate that owns the renewal book at roughly a tenth the cost of a loaded rep, and it was finally design-partner ready.

What's Next

Next week we start our first design partnership. After three weeks of research, one dashboard we had to bury, a rebuild, and ten demos' worth of feedback, someone's putting real renewals in front of it. That's the whole point — and it's genuinely exciting.

What I'm Taking Away

• Spend week one learning, not building. The research deck became the spec, and the spec became the guardrail for everything after.
• A working demo can still be the wrong foundation. "It runs" and "it's built right" are different claims. Bloat is a real cost, and sometimes a rewrite is the fast path.
• Write the PRD, then build. The rebuilt dashboard was faster to make because the thinking happened before the code.
• Demos are the best product feedback loop there is. Ten conversations taught me more than any amount of solo polishing. If you're building something to sell, start selling it before you think it's ready.

More soon — after the design partner kicks off.