Maya Bashir

Founder · Wovenroot

Maya runs Wovenroot, a DTC furniture brand with a strong Instagram following. The voice that built the brand is fighting the support volume that came with growth.

The setup

Wovenroot is a DTC furniture brand Maya started four years ago, walnut credencas, linen sofas, brass-and-wood lamps, photographed in real homes by people who actually live with the pieces. The Instagram presence is the brand. 380K followers, every post written by Maya, every reply in her voice.

The voice is what built the brand. Customers buy a $1,800 walnut credenza partly because Maya’s product description felt like she wrote it personally. They follow because the captions read like a friend who happens to make furniture. They reorder because the brand feels like one person.

Then growth happened. Support volume is up 3x year-over-year. Maya can’t answer everything anymore, the queue would consume every hour of every day. The team has been answering, and the team is good, but the replies have started sliding toward templated. By her own audit, 35% of recent responses have read as Generic Helpful Brand: <em>We apologize for the inconvenience. Your satisfaction is our priority.</em> Every templated reply is a step toward Wovenroot becoming indistinguishable from the next DTC label.

She came to Zynkex with a knot. “Support volume is up 3x year-over-year. We’re falling back on templated replies and customers can feel it. The voice that built the brand is getting lost in tickets. I need a way to scale support without sounding like every other DTC brand.”

The session

The session opened with Maya framing this as a support-automation problem. Zynkex pushed back on the category. Support automation is the wrong frame because it treats the support reply as a logistics transaction, get the answer to the customer, fast. For Wovenroot, the support reply is a brand touchpoint. The category Wovenroot needs is voice-faithful automation, and the category is empty.

That reframe shifted the entire conversation. Maya had been thinking about response speed and ticket volume. The session pushed her toward thinking about voice fidelity, and the infrastructure that would make voice fidelity possible at scale.

The breakthrough was identifying that the voice itself was a defensible asset, but only if Wovenroot built the evaluation infrastructure to protect it. A bot trained on Maya’s past replies is easy. A bot with a continuous-evaluation pipeline that scores every response against a calibrated voice sample before it ships, that’s harder. That’s the moat.

The architecture fell into place from there. Live in-line scoring at a 0.85 voice-similarity threshold. Hard escalation rules on accuracy and refund caps. A draft-reply handoff so even escalated conversations are 80% done before the human picks up. And a weekly voice-drift report sent to Maya every Sunday, naming any response that scored below threshold and any phrase the bot started using that wasn’t in the calibration sample.

By the end of the session, the pitch had landed. Wovenroot isn’t adding a chatbot. Wovenroot is teaching the brand voice to scale, and building the infrastructure that proves the voice survived. That’s the story for the next investor conversation: not "we automated support," but "we proved the brand voice is an infrastructure asset, and we built the infrastructure first."

The Build Brief

For the engineering vendor building the bot across email, Instagram DMs, and live chat. Names the voice rules, the evaluation pipeline, the Slack handoff, the weekly drift report. The .docx the build runs on.

Build Brief for the Wovenroot voice-faithful support bot. Paste into your AI coding tool; the tool will determine the file structure for the build environment. The implementation produces a customer-support bot deployed to email, Instagram DMs, and live chat, a continuous-evaluation pipeline that flags voice drift before responses ship, and a Slack handoff to the human support team. Read end-to-end before generating code, the voice-as-product frame in Section 1 governs every reply, eval rule, and escalation decision downstream.

Wovenroot, Voice-faithful support bot

What this is. A support bot deployed across email, Instagram DMs, and live chat that handles ~80% of inbound support volume in the Maya voice. Plus a continuous-evaluation pipeline that flags voice drift before responses go out, so the brand voice survives scale.

The voice-as-product constraint. Every decision in this brief is gated on the fact that Wovenroot's support reply is a brand touchpoint, not a logistics transaction. Customers who reach support are at the exact moment the brand promise is tested. Templated replies, even fast, accurate ones, break the promise.

Wovenroot is not adding support automation. Wovenroot is teaching the brand voice to scale.

When uncertain whether to ship a feature, ask: would the response sound like Maya wrote it? If yes, ship it. If it sounds like Generic Helpful Brand, reject it, every Generic Helpful Brand reply is a step toward indistinguishable from the next DTC label.

Section 1. Channels

Three channels at launch. Each handled by the same bot with channel-specific formatting.

Email. hello@wovenroot.com. The highest-volume channel. Threading preserved; bot replies maintain the customer's existing thread.
Instagram DMs. Routed via the Instagram Graph API. Image attachments common (customers send photos of damaged pieces, fabric samples, room shots).
Live chat on wovenroot.com. Embedded widget. Real-time response expected.

The bot does NOT handle phone, Twitter/X DMs, or in-store inquiries. Out of scope for v1.

Section 2. The voice rules

The brand voice has been Maya's personal voice since launch. It is:

Warm but not effusive. That's frustrating, let me figure it out, not Oh no, I'm so sorry to hear that!
Specific, not categorical. The walnut on the credenza takes about two weeks to settle in dry air, not Hardwood furniture can be sensitive to humidity.
First-person from Maya, even when the bot is replying. The customer believes they're talking to Maya. Until they decide they need a person, this is intentional and the bot maintains it.
Uses contractions, em-dashes, and the occasional lowercase first word in informal replies. Reads like a friend who happens to make furniture.

Phrases to use: Let me figure this out. That's frustrating. I'll get back to you by tomorrow morning. The walnut, the linen, the brass, every material is named specifically.

Phrases to avoid: I apologize for the inconvenience. Your satisfaction is our priority. Per our policy. Unfortunately. Please understand.

Resist the urge to add a "support voice" mode that's distinct from the brand voice. The unified voice is the entire point. The moment customer support sounds like a different brand than the Instagram captions, customers feel the seam.

Section 3. Tool inventory

The bot has access to:

Order lookup. Shopify Admin API, read-only, scoped to the customer's email.
Shipping carrier status. EasyPost API for tracking events.
Product knowledge base. Maya's writing on each product, materials, care instructions, indexed in a vector store, retrieved per-conversation.
Refund / replacement initiation. Shopify API, mutating, capped at $200 per request without escalation. Above $200 triggers Slack handoff.
Slack handoff to the support team for anything outside the bot's scope or above the refund cap.

The bot does NOT have:

Authority on policy exceptions. Maya makes those calls.
Access to product roadmap or future-collection information. Frequently asked, must be escalated.
Send capability on email addresses other than the customer's. Hard privacy boundary.

Section 4. The evaluation pipeline

Every response is scored on three dimensions before it ships to the customer:

Voice score. LLM judge against a 50-sample calibration set of past Maya replies. Threshold: 0.85 cosine similarity on the voice-embedding vector. Below threshold, the response is rewritten and re-scored once; below threshold twice, the conversation is escalated to a human.
Accuracy score. Fact-check the response against the order details, the product KB, and the shipping status. Threshold: 100%, any factual mismatch routes to human.
Refund threshold. Hard rule: any reply that initiates a refund > $200 escalates regardless of voice + accuracy scores.

The eval pipeline runs in-line, adding ~600ms of latency per response. For live chat this is the load-bearing trade-off: slower but voice-faithful beats fast and generic.

Section 5. Slack handoff

When escalation triggers, the bot opens a thread in #support-overflow Slack channel with:

The customer's name + order number.
The full conversation history.
The reason for escalation (voice drift / fact mismatch / refund cap).
A draft reply the human can edit + send, or rewrite from scratch.

The draft is the load-bearing handoff feature. Even when the bot escalates, it provides 80% of the work the human would have done, the human's job is the last 20% (the voice fix, the policy call, the final send).

Section 6. Voice-drift report

A weekly report goes to Maya every Sunday at 8pm with:

Voice score distribution across the week's responses.
Any responses scored below 0.85 (the rewrites + final ships).
Any responses where customers explicitly responded "is this a bot?" (the test set for voice failures).
New phrases the bot started using that weren't in the calibration sample.

Purpose: Maya is the voice owner. The report is her tool for protecting the voice as the bot scales.

Section 7. Operational details to provide

Shopify admin token with scoped permissions.
EasyPost API key.
Instagram Graph API setup + the Meta business verification.
The 50-sample calibration set of past Maya replies, to be sourced from the support inbox archive.
The list of Maya-owned policy exceptions (return windows, replacement criteria) for the bot's knowledge base.

The Strategy Plan

For Maya to take into the next investor conversation. Names the voice-as-product thesis, the eval-pipeline-as-moat position, the year-one targets, and the explicit list of decisions the team protects. The .docx the Series A pitch is built around.

Wovenroot, Strategy Plan

Wovenroot is fighting the DTC voice-flattening cycle. Support volume is up 3x year-over-year; templated replies are eroding the brand voice that built the customer base. This plan ships a voice-faithful support bot, builds the evaluation infrastructure that protects the voice at scale, and turns the result into the case Wovenroot makes in the next investor conversation.

Opportunity

3x YoY support volume growth. Current handle time of 4 minutes per ticket at the human team; queue length growing by 30 tickets per week. Templated replies have crept into 35% of responses, measured by manual audit.

The recovery: ~$180K/year in support headcount avoided through v1 alone. The hidden upside: the only DTC furniture brand whose voice survived scale, the narrative the Series A pitch is built on.

The audience is the customer who bought a $1,800 walnut credenza partly because Maya's product description felt like she wrote it personally. They reach support because the credenza arrived with a scratch, or the linen color is darker than the website photo, or the delivery window slipped. The reply they get is the moment the brand promise either holds or breaks.

Every templated "We apologize for the inconvenience" reply is a step toward Wovenroot becoming indistinguishable from the next DTC label. Customers stop posting unboxing videos. Word-of-mouth flattens. The CAC has nowhere to go but up.

Our Position

Wovenroot is not adding support automation. Wovenroot is teaching the brand voice to scale.

The support-automation category is full of bots that promise speed and deliver generic. Their value proposition is throughput; their cost is brand. For Wovenroot, a brand whose differentiator is voice, that trade is upside-down.

The voice-faithful-bot category is empty in DTC furniture. Wovenroot is the first brand to ship support that sounds like the brand, at scale, with the eval infrastructure to keep it that way. That position is defensible by infrastructure: competitors can't copy the voice without copying Maya, and copying Maya requires the calibration set + the eval pipeline + the drift report that Wovenroot built first.

If a customer ever asks "is this a bot?" the strategy has failed for that interaction. The eval pipeline is the mechanism that makes "no, it's me" continue to be true even when it's the bot.

The Strategy

Voice as product. Every reply is a brand touchpoint, scored against the voice calibration before it ships. The 0.85 threshold is non-negotiable.

Eval pipeline as moat. The 50-sample calibration set, the voice-drift weekly report, the live in-line scoring, this is the infrastructure that lets Wovenroot ship the bot without losing the voice. Competitors can copy the bot architecture; they cannot copy the voice-evaluation discipline.

Human escalation is feature, not failure. When the bot can't maintain the voice or hits a policy ceiling, it escalates with a draft reply already 80% done. The human's job is the voice fix and the final send, not the whole response.

Channel scope: email, Instagram DMs, live chat. Phone, X DMs, in-store are out of scope for v1. Narrow channel set lets the voice eval be tight.

Year-one targets.

80% of inbound support volume handled by the bot without escalation.
< 0.5% of conversations end with "is this a bot?", the explicit voice-failure metric.
Weekly voice-drift report reviewed by Maya every Sunday; zero responses ship with voice score < 0.85.
Support headcount avoided: ~3 hires that would have been needed at current growth.

The decisions we made.

Voice-faithful, not voice-neutral. The bot speaks as Maya, not as a generic helpful brand.
In-line eval, not after-the-fact audit. The voice score gates every response.
Refund cap at $200, not "use your judgment." Hard rules protect against the LLM's worst mistakes.
Three channels at launch. Phone and X DMs deferred to v2.
Maya owns the voice. The weekly drift report goes to her, not to ops.

Risks

The single largest risk is calibration-set drift. The 50-sample set captures Maya's voice in a specific moment of the brand. As Maya's voice naturally evolves over the year, the eval pipeline scores the bot against a frozen Maya, and the bot becomes "the brand voice from January" rather than "the brand voice." Mitigation: quarterly re-calibration with 10 fresh samples added to the rolling window.

Second risk: customer-disclosed AI fatigue. If customers begin to clock the bot regardless of voice fidelity, through cadence, response time, or platform-specific tells, the brand promise still breaks even if the words are right. Mitigation: human-handled "is this a bot?" responses get tracked, surfaced, and reviewed monthly.

Open, whether to disclose the bot's existence publicly via a footer page about how Wovenroot handles support. Deferred until the first ninety days of customer-disclosed AI fatigue data tell us whether transparent disclosure builds trust faster than perceived authenticity.

What’s the asset your brand can’t afford to scale poorly?

Already have an account? Sign in