How I AI: 3 AI Workflows to Save $140K & Automate Marketing with ElevenLabs' Luke Harries

ElevenLabs Head of Growth Luke Harries builds custom AI systems for case studies, localization, and WhatsApp-based assistants. He calls the approach "vibe marketing": marketers use AI and code to build the operations they need instead of limiting themselves to off-the-shelf software.

Luke said the localization workflow replaced a $40,000 annual platform and more than $100,000 in agency work. Those are his reported avoided costs at the time of recording, not a guarantee that another team will achieve the same savings or quality.

In this episode of How I AI, Luke turns a recorded customer conversation into draft marketing assets, shows the translation system he built with Cursor, and connects Claude to WhatsApp through a Model Context Protocol server.

Turn a customer interview into draft marketing assets

The case-study automation workflow shows how Luke moves from a recorded customer conversation to reviewable social posts, a case study, and a sales brief without losing quote verification or editorial approval.

Luke's workflow moves from an interview to a case-study draft and social copy in under 15 minutes in the demonstration. Publication still depends on fact-checking, customer approval, brand review, and the rights to use names and quotes.

Capture the customer interview

Luke interviewed me about using ElevenLabs to prototype product keynotes at LaunchDarkly while Granola captured the conversation. Recording and transcription should be disclosed to participants, and confidential company or customer information should be excluded unless its use is explicitly approved.

Granola produced a structured summary and extracted details from the conversation. Its notes are a draft record, so names, roles, claims, and quotations need to be checked against the recording and confirmed with the customer.

Apply the ElevenLabs editorial brief

Luke built a custom GPT called "ElevenLabs Copy Editor." Its instructions define spelling, tone, formats, company messaging, and examples of approved work.

He supplies two versions of the same source:

1. The concise summary generated by Granola.
2. The full transcript for evidence, wording, and context.

The prompt begins with a direct editorial role and house style:

You are an expert editor and writing assistant specializing in the ElevenLabs communication style. You must enforce American English spelling. Your tone is serious and research-led, similar to Palantir or SpaceX.

The instructions also specify:

• Formats for assets such as case studies and social posts.
• Rules for skimmable headings and consistent structure.
• Approved examples that demonstrate the intended voice.

When the output repeatedly misses a requirement, Luke updates the shared instructions instead of correcting the same issue in every draft. Changes to the brief should be versioned and reviewed so a prompt edit does not silently alter every downstream asset.

Draft, verify, approve, and distribute

The GPT created a case-study draft about LaunchDarkly's use of ElevenLabs Studio, with headings and candidate pull quotes. The team still needs to compare every claim and quote with the transcript and obtain customer approval before publication.

Luke then asked for a social thread with placeholders for visuals. Repurposing speeds up distribution, but each format needs its own edit for context, substantiation, length, accessibility, and channel norms.

[Image placeholder: ElevenLabs Studio interface]

Luke also connected Salesforce and Zapier so eligible customers could receive an invitation to share their story. The automation should honor marketing consent, suppression lists, frequency limits, ownership, and a clear path for a person to decline.

Build a contextual localization pipeline

The custom translation-service workflow captures the language-specific prompts, changed-string trigger, engineering checks, and human language review behind the system that replaced a six-figure vendor.

Luke wanted to localize the ElevenLabs website without the platform and agency workflow the team had been using. He reported $40,000 in annual software cost and more than $100,000 in agency spend before building the internal system.

His team was already using ChatGPT to review agency translations, which prompted him to test whether a custom pipeline could handle the first translation pass directly.

Prototype the service in Cursor

Luke used Cursor to build the initial translation service over a weekend, then involved engineers to prepare it for company use. A working prototype is not ready for company use until owners review architecture, security, reliability, observability, and recovery behavior.

I think human-in-the-loop SaaS, where your job is about putting low-skilled workers in some sort of flow, which translation is, I think that's very risky.

Route changed strings through language-specific instructions

The website workflow has four parts:

1. Trigger: A change to a source string starts a GitHub Action.
2. Translation: The action sends the source string to the team's model service.
3. Language context: Each target language has its own glossary, brand rules, and stylistic guidance.
4. Reviewable change: The translation returns to the repository. A safer production setup opens a pull request, runs validation, and requires the appropriate language or content owner before merge.

The team connected CMS content to the same service through a Translate button. Luke said the system had not required maintenance at the time of recording. Ongoing cost still includes model usage, engineering ownership, monitoring, and qualified review for priority markets.

The workflow changes the build-versus-buy calculation, but it does not make every localization tool unnecessary. Product UI, legal text, safety information, regulated claims, and culturally sensitive copy may require professional localization, native-speaker review, and market-specific approval.

Query WhatsApp through a local MCP server

The WhatsApp MCP workflow keeps message retrieval local and bounded, then lets Claude query the approved history instead of granting a general agent unlimited access to private conversations.

Luke built an experimental Model Context Protocol server that lets Claude query WhatsApp data and invoke messaging tools. The capability is useful precisely because it is sensitive: private conversations, contacts, and the ability to send as the user sit behind the connection.

Model Context Protocol defines a way for an AI client to call tools and retrieve data from connected systems. The protocol does not make a tool safe by itself; permissions, authentication, data handling, and confirmation still belong to the implementation.

How Luke's WhatsApp MCP works

Luke published the project code on GitHub. The demonstration uses this local setup:

1. Local client: A script uses the whatsmeow library to connect as a WhatsApp Web client.
2. Sync: After QR authentication, the client copies message history into a local SQLite database and listens for new messages. Local storage is not automatically secure; the database, backups, device, logs, and model access all need protection and retention limits.
3. Tools: A local MCP server exposes functions such as reading recent messages and sending a message.
4. AI client: Claude can call those tools after the server is connected. Start with read-only, narrowly scoped access; restrict chats and time ranges; and require a preview and explicit confirmation before any send action.

Query and orchestration examples

Luke demonstrated a read request:

What are some recent messages I've received on WhatsApp?

He also described summarizing relevant discussions from busy groups and passing that summary to the editorial GPT. Participants may not expect their messages to become marketing research, so teams need a lawful basis, consent where required, redaction, access controls, and human review before reusing conversation content.

Luke chained the WhatsApp and ElevenLabs tools to turn a discussion summary into a voice note and route it back through WhatsApp. Before using a send workflow, confirm the recipient, source text, generated audio, account, and consent. Untrusted message content should never be allowed to override tool instructions.

Build around the bottleneck, then name the controls

Luke's systems join existing inputs, editorial rules, code, and distribution channels. The value comes from the handoffs: one interview can become several drafts, one changed string can start localization, and one query can retrieve a bounded set of messages.

Each handoff also creates an owner. Customer marketing needs approval and quote verification. Localization needs engineering and language review. Private-message access needs strict permissions, security, consent, and confirmation. Automation makes those controls more important, not less.

The $140,000 savings headline came from replacing a localization service, but Luke did not treat lower cost as proof of quality. The internal system still needed language-specific context, review from people who understood each market, and a safe way to retry or reject output when the model missed tone or product meaning.