We pulled £33,468 in revenue from 800 “forgotten” leads in 19 days – here’s how, and how you can do the same with your old CRM data. Most business owners I talk to are sitting on years of old leads that no one has the time (or energy) to call again. Missed timing. Missed context. And eventually… missed revenue. I set up an AI voice reactivation system for a UK housing disrepair claims company that had 35,000 dormant leads sitting in their CRM. We only touched 801 of them for the first test, and here’s what stood out: - The AI handled 98% of all calls end-to-end. - Only 18 calls needed a human, but those 18 represented ~£89k in realistic compensation… and about £33,468 in potential revenue for the company. Here’s the simple breakdown: - Let the AI agent call every old lead for youIt dials automatically, makes multiple attempts, speaks in a natural local accent, asks qualifying questions, and updates the CRM after every call. No manual dialing. No spreadsheets. No chasing. - Filter out all the noise and send only the good stuff to your teamEvery call ends with a clear outcome: not interested, do not call, no answer, follow-up, or transfer to human. Your staff only receives the “transfer” calls — the people who actually want to move forward. In our test, that was just 2% of leads… but that 2% carried nearly all the revenue. - Use data, not gut feeling, to decide if it’s worth scaling. From just 801 leads, we got 18 high-intent opportunities. Once we saw the numbers, it became obvious: rolling this out to the full 35,000-lead database is realistically a multi–six to seven-figure opportunity that was previously just… sitting there. This works for any business with a big pile of old leads: home services, claims, legal, medical, local services — anywhere people enquire, then go cold. If you’re sitting on a CRM full of “maybe one day” leads and want help thinking through whether an AI reactivation agent makes sense for your situation, just ask and I can walk you through the math for your numbers.

After scoping 15+ ai voice agent projects and shipping $30k+ in custom builds, I finally landed on a scoping workflow that turns messy client requests into clean, buildable technical plans. i keep seeing ai voice teams stuck in the same loop: clients speak in outcomes… we have to build in systems. that translation layer is where projects blow up. here’s the lightweight scoping “automation” I use to avoid that: what it does - takes the client’s non-technical wishlist - breaks it into: - required systems (crm, calendar, voice platform, phone, automation, sms/email, external apis, compliance) - core use case (e.g. booking agent, lead qual, receptionist) - caller intents (book, reschedule, support, out-of-scope) - for each intent, maps: - behavior → what the agent says - data → what must be collected - system actions → what tools/functions we need - turns everything into a simple flowchart that becomes the project blueprint everyone agrees on why it works - prevents scope creep (everyone sees all branches upfront) - catches missing dependencies early (apis, calendars, compliance rules) - makes tool design almost automatic - booking flows typically boil down to: - create_contact - retrieve_contact - check_availability - book_appointment - simplifies pricing ← complexity is literally drawn on the board if helpful, here’s the full step-by-step guide I use inside my agency: https://how-to-scope-ai-voice-ag-7i8lpn7.gamma.site/

After scoping 15+ custom voice agents, I realized most projects fail for the same reason: the scope is vague. Everyone jumps into “Let’s just build it,” but no one agrees on the systems, intents, or actions that actually define the project. Here’s the simple 3-pass checklist I now use to scope any AI voice agent cleanly: 1. Systems Write down every tool the agent will touch: CRM, calendar, voice platform, automation layer, SMS/email, external APIs, KB sources, phone routing, and any compliance checks. If it's not on this list, it’s not in the scope. 2. Intents Map all possible caller intentions from a single “Welcome” node: - New inquiry - Existing customer - Reschedule - Support request - Out-of-scope - Spam Each intent becomes its own branch you can design independently. 3. Actions For each intent, define two things: - Data: What fields must be collected? - System actions: What does the agent do in your stack? (create/update contact, check availability, book appointment, send SMS, tag lead, generate call report) Most scoping issues happen because one of these three passes was skipped. If you can’t visualize the full flow on one clean page, you're not ready to build yet. If you want the exact guide I use to scope any AI voice agent project, I’ve linked it here: https://how-to-scope-ai-voice-ag-7i8lpn7.gamma.site/

After building voice agents for 15+ clients, I turned my entire QA process into a quick LLM-powered workflow that cuts out 80–90% of the manual testing. Testing voice agents used to drain my week. Every small update meant another round of manual calls. Fix a node → retest. Update a fallback → retest. Change a price → retest everything again. It was brutal. So I built a tiny internal system to automate almost all of it. Here’s the exact flow: 1. Export the full agent as a JSON file. 2. Drop it into an LLM along with the client’s FAQ, KB pages, and policies. 3. Paste a system prompt that forces the model to understand every branch and condition. 4. Auto-generate 15–25 realistic test cases with: – emergencies – confused callers – angry callers – pricing checks – spam filters – function call tests – out-of-order info – multi-language attempts 5. Convert the whole output to Retell’s JSON format. 6. Import it into simulations. 7. Run everything at once and skim transcripts for anything weird. Total time: ~12 minutes. The real win isn’t just speed — it’s consistency. Every agent gets tested the same way. Every update gets retested with the same suite. No more “Oh I forgot to test billing flows” moments. If you want this level of sanity back, here's my exact SOP

I’ve built and delivered custom AI voice agents for 15+ clients, and this is the exact 7-step system I use to automate all testing so you don’t waste 5–10 hours manually calling your agent every time you update it. Hi all—testing voice agents takes way more time than most people expect. And I know a lot of you here are still testing everything by hand. I recently cleaned up my internal QA workflow and turned it into a simple, repeatable process. Here’s how you can set it up for your own builds: 1. Export your entire voice agent as a JSON file from Retell/Vapi. 2. Upload that JSON + any knowledge base files into your LLM workspace. 3. Paste in the system prompt (I shared it in the video + SOP). 4. Let the LLM analyze all branches, nodes, intents, and fallbacks. 5. Auto-generate 15–25 test cases covering emergency, spam, billing, scheduling, objections, retrieval accuracy, and more. 6. Download the output as a JSON file. 7. Import it straight into Retell Simulations and run everything in one click. A few tips that make this way more reliable: - Always review test cases manually. Passing doesn’t mean correct. - Test one thing per case. Don’t combine behaviors. - Include happy, unhappy, confused, and emotional callers. - Add tests for price retrieval, business hours, policies, and RAG accuracy. - Include speaking styles: short, long, fast, slow, out of order. - Use both simulations and manual listening to check tone and flow. - Refine and re-run until you’re happy with 80–90% pass rate. If anything here is unclear, let me know. Access my $30K SOP here: Voice Agent SOP