Building a custom AI sales tool in 6 weeks: a Honka build log

Honka has been making Finnish log homes since 1958. Mark and Tom run the US arm. For decades they've hand-written the personalized spec sheets that go out to every serious prospect — the kind of high-touch follow-up that turns a $1M build into a closed deal. The problem: there are only two of them, and the spec sheets that don't go out the same week often don't go out at all.

Six weeks ago, the two of them decided to stop hand-writing those spec sheets and start having an AI do it instead. By the end of those six weeks, every prospect who finished their custom configurator was getting a branded, personalized PDF — written by Claude Haiku in under 10 seconds — that wove Mark and Tom's own founder-approved sales pitches into the prospect's exact build.

This post is the build log. The full system is documented in the Honka case study, with screenshots, founder testimonials, and the actual code that powers the AI. This is the behind-the-scenes: how we scoped it, what we built each week, the three architectural decisions that mattered most, and what we'd do differently if we ran it again.

If you've been told AI is years away from being useful for your specific business, or you're tired of generic SaaS tools that almost-but-not-quite fit, the math on a focused custom build might surprise you.

What we built (so you have the picture)

KotiConfig is an AI-native home configurator embedded in Honka US's existing HubSpot site. Prospects step through a branching pricing flow, hit a lead-capture wall part-way through, and on submit receive two things automatically:

1. A personalized PDF spec sheet — branded with Honka's serif typography and forest-green palette — that opens with the prospect's own vision in their own words and weaves in matching Honka pitches (fire safety, energy efficiency, legacy, healthy indoor air) based on what they answered "yes" to during the flow.
2. A lead tier — A, B, C, or D — assigned automatically based on land ownership, timeline, build size, and the depth of their open-ended answers.

Mark and Tom get an admin dashboard with the leads sorted by tier. They know who to call first without ever opening a spreadsheet.

You can see the actual UI, the generated PDF, and the founder testimonials in the case study. The rest of this post is how it got built.

The 6-week timeline (with what actually happened each week)

Six weeks isn't a marketing number. It's what's possible when the scope is right, the founder is in the room, and there's an existing site you can embed into instead of replacing.

Here's the rough week-by-week:

Week 1 — Discovery and scoping

We got Mark and Tom on a recorded call and mapped their actual sales playbook. Not their idealized playbook — the one they actually use on every prospect call.

The key insight came in this meeting: Mark and Tom didn't need a smarter generic CRM. They needed a tool that ran their specific sales motion automatically. We left week one with a one-page spec, a list of the six sales pitches they wanted in every spec sheet (the campfire-secret fire safety pitch, the energy-efficient logs pitch, the legacy-home pitch, and so on), and a decision: embed the tool inside Honka US's existing HubSpot site rather than replace it.

Week 2 — Architecture and the first cut

This is where we made the three decisions that shaped everything else (see next section). By the end of week two we had a working branching configurator with hard-coded options — no AI yet, no PDF yet, just the flow. Honka could click through it and confirm the structure was right.

Week 3 — The configurator engine and pricing

Each Honka home style (Ranch, Walkout, 2-Story) has a different downstream sequence — the windows that make sense for a Walkout aren't the windows that make sense for a Ranch. We built the branching logic this week so every choice cascaded the right way. Live pricing ran in parallel; the number on screen always matched what would land in the PDF.

Week 4 — The AI spec sheet system (the founder pitch weave)

This is the heart of the product. We took Mark and Tom's six sales pitches, codified them as a TypeScript array, and wired Claude Haiku to weave them into each prospect's PDF based on what they answered "yes" to in the configurator. The pitches are in the same words Mark and Tom would use on a call.

We named the pattern internally: the founder pitch weave. We've since used variants of it on other client builds.

Week 5 — Admin CRUD and lead tier scoring

By week five we had the customer-facing flow working end-to-end. This week we built what Mark and Tom would actually use day-to-day: a leads table with tier badges, full-text search, and a brand-guide editor so they could change their pitch language anytime without us being in the loop.

Week 6 — Polish, embed integration, launch

iframe-embed in the HubSpot site. postMessage SDK so lead events flow back to the parent page for analytics. Final QA. Launch.

The founder testimonials in the Honka case study were recorded a few weeks later, with the system already in production.

The three architectural decisions that mattered most

Every custom AI build worth doing comes down to three or four early decisions that determine the next six weeks. Naming and defending those decisions up front is more valuable than picking the "right" framework. Here's what we picked for Honka and why.

1. Iframe-embed first (don't replatform)

Honka US's traffic already lived on a HubSpot marketing site they'd built up over years. The configurator could have been its own standalone domain, but every prospect we'd have asked to leave the parent site was a prospect we might have lost.

So we built KotiConfig as an iframe-embed with a small postMessage SDK that passes lead events back to the parent. Honka kept their site, their analytics, their SEO equity. We added a new floor of capability without ripping out the foundation.

The lesson for your business: if you have a working website with existing traffic and brand equity, an embedded AI tool is almost always the right answer over a full rebuild. You preserve everything that's already working and add only what's missing.

2. Branching via explicit step IDs (admin-editable without code)

The configurator branches based on home style. We could have done that with tag-based logic — every step checks predicates against the prospect's answers and decides what to show next. That approach is flexible. It's also unreadable to anyone who isn't the original engineer, and unchangeable by Mark and Tom themselves.

We picked the opposite: every option on every step has an explicit "next-step ID" pointing to the next step. The flow is a graph anyone can audit by reading the data. More importantly, Mark and Tom can change the flow from the admin UI — add a new home style, rewire its sequence — without ever opening a code editor.

The lesson: when you build for a small team that wants to own their tool, optimize for founder readability, not engineering cleverness. The engineering team won't be there forever. The founders will.

3. Claude Haiku, not Sonnet, for the PDFs

For a personalized spec sheet that needs to ship in under 10 seconds, model latency matters more than model intelligence. Claude Haiku hits sub-2-second latency on Honka's prompt size with the same JSON-schema adherence we'd get from Sonnet at a fraction of the cost. The stylistic flair of a bigger model wasn't load-bearing — Mark and Tom's voice already carries the soul of the copy.

The lesson: don't default to the biggest model. Pick the smallest one that meets the task's actual constraints (here: latency, structured output, brand voice). You'll ship faster, pay less, and your prospects will get their answer before they tab away.

Three things that surprised us

Mark and Tom had a more codified playbook than they realized

When we asked them "what do you say when a prospect asks about fire safety?" we expected vague answers. We got a polished, three-sentence pitch with the exact analogy (campfires) and the exact regulatory citation (LA County codes). Same for every other pitch. Their sales playbook was already in their heads; we just had to write it down.

Most SMB founders are sitting on a 200-page sales playbook stored entirely between their ears. Pulling it out of their heads and into a code file is more valuable than any AI feature on its own. The AI is just the delivery mechanism.

The PDF was the load-bearing artifact, not the configurator

We thought the configurator was the product. It isn't — the PDF is. The configurator is just the questionnaire that generates the inputs. The prospect's actual takeaway, the thing they'll forward to their spouse, the thing that sits in their inbox for weeks while they decide — that's the PDF.

If we'd started with the PDF (designing the output first, then working backwards to the inputs), we'd have shipped the same product in five weeks instead of six.

The voice was the hardest part to get right

Code is easy. The configurator was built in days. The PDF layout was built in days. What took multiple rounds was getting Claude to write in Mark and Tom's voice — warm, premium, knowledgeable-friend, no marketing fluff.

The fix wasn't a better model. It was a better brand-guide section in the system prompt, written by the founders themselves and editable from the admin UI. For voice-led products, the model isn't your bottleneck. Your prompt's brand guide is.

What we'd do differently next time

• Start with the PDF. Design the output first, then design the questionnaire that feeds it. Working backwards from the artifact would have saved a week.


• Scope branching tighter. We supported three home styles from day one. Two would have been enough to ship a v1 and learn from real lead behavior before expanding.


• Build the admin UI in week one. We built it in week five. Earlier would have meant Mark and Tom could shape the copy as we went instead of waiting for the engineer to ship a brand-guide editor.

These aren't dealbreakers — six weeks was the right outcome. They're sharpenings for the next build.

Could you ship something like this?

Most SMBs probably can — if four conditions hold.

1. You have a subject matter expert whose voice IS the product. A founder, a senior salesperson, a craftsman, a clinician. Someone whose customers buy them, not just the company. If that person could be cloned and reach more prospects, AI is your cloning tool.
2. Your sale is high-touch. Long sales cycles, high ticket prices, lots of qualifying questions. The longer your sale, the more leverage an AI follow-up tool gives you per closed deal.
3. You have an existing site you can embed into. This is the iframe-first lesson. If you have working traffic and SEO equity, don't tear it down to add AI — add AI on top.
4. You're willing to spend ~6 weeks of focused scope. Not part-time exploration. Six weeks of clear, defended priorities with the founder in the room.

If three of those four hold, you have a real shortlist of candidate tools. The one we'd build first is almost always the one that automates the highest-leverage follow-up your team does manually today.

For more on the broader pattern of small businesses moving fast on AI, see our earlier post on why the AI adoption gap exists and how SMBs can close it.

Want to know what your version looks like?

This is the part we want to make easy. The three-step path:

1. Spec it in Mission Control — answer a few questions and our AI drafts a build spec for your specific business. Free, no signup, 10 minutes.
2. Paste the spec into any AI tool you trust — Claude, ChatGPT, Gemini, Grok — and ask "is this realistic in 6 weeks? what would it cost? what's the simplest version of this?" You'll get a second opinion in two minutes, for free, from a tool that has no stake in selling you anything.
3. If you want a partner who actually ships this stuff — get in touch. Same studio that built KotiConfig. Same approach. Same six-week scope.

Read the full Honka case study for the UI screenshots, the founder testimonials, the architecture diagrams, and a downloadable sample of the AI-generated spec sheet PDF.

FAQ

Does my business have to be tech-savvy to use AI like this?

No. Honka is a 68-year-old log home manufacturer whose product takes a year to build and is made of timber from Finnish forests. They didn't need to be a tech company to add an AI sales tool — they needed a partner who could build one for their specific sales motion. The "AI is for tech companies" framing is one of the most expensive misconceptions in small business today.

How long is a 6-week build really? Was it actually 6 weeks?

It was actually 6 weeks of calendar time, but that's because of three conditions that held: the scope was defended, the founders were in the room for every decision, and we embedded into their existing site rather than rebuilding it. Without any one of those three, the same project would have been 10-12 weeks. The lesson isn't "every custom AI build is 6 weeks." It's "every custom AI build is 6 weeks if you can hold those three conditions."

Could Honka have built this with off-the-shelf SaaS instead?

For the basic configurator step, yes — there are configurator SaaS products on the market. For the AI-generated personalized PDF that weaves the founders' own sales language into each prospect's spec sheet, in their own brand voice, with their own pitches — no. That's the part custom unlocks. The rule of thumb: if your differentiation is in the voice of the output, custom usually beats SaaS. If your differentiation is in the flow, SaaS is often fine.

What does the long-term maintenance look like?

Honka US owns the code, the cloud infrastructure, and the roadmap outright. Day-to-day, they can edit pitches, add steps, and tune the lead tier rules from the admin UI without needing an engineer. For larger changes (new integrations, new features), they're either able to extend it themselves with an AI coding assistant or they call us. The maintenance philosophy from day one was "the founders should be able to run this without us."

What's the "founder pitch weave" exactly?

The founder pitch weave is a pattern for codifying a small team's best sales pitches as data and having an AI deliver them inside customer-facing artifacts (PDFs, emails, chat replies) based on what the customer has said they care about. For Honka, the pitches live in a TypeScript array; the AI reads the prospect's "yes" answers and weaves the matching pitch into the spec sheet. We've used variants on other client builds since. The Honka case study has the actual code excerpt.

Where can I see the actual product?

The Honka case study has the actual UI screenshots, the founder testimonials, the architectural decisions in detail, and a downloadable sample of the AI-generated spec sheet PDF. It's the worked example for everything in this post.

---

Phil Thornton is a co-founder of Pilotship.io — an AI-native dev studio that helps small businesses turn ideas into shipped products with AI. We specialize in custom AI tools for established businesses in real estate, professional services, custom manufacturing, and high-ticket consumer brands. Clients own their code and their cloud from day one. Questions about anything in this post? Get in touch.