My team stopped calling 30% of our leads. They couldn’t reach them. An AI seller kept writing. 56% came back. Cost per approved credit: −12%. Commercial operating cost: −6%. Here’s how it works 👇

Meta grades every server event you send it, from 1 to 10. The grade is Event Match Quality, in Events Manager. Meta's docs say to aim for 6.0 or higher. What moves it: ✅ Email, trimmed, lowercased, then SHA-256 hashed ✅ Phone hashed, with the country code, even if all your users are in one country ✅ fbp and fbc from the browser ✅ Client IP and user agent on web events ❌ A phone without country code: Meta won't use it to match Check the score on the event you optimize to, not just the one with the most volume.

Five agents build our monthly sales dashboard. The one I trust most never draws a chart. Four of them build it. The fifth checks their numbers, and it does that with a fixed script, not a prompt. Why a script: a prompt-based checker can be argued into the same mistake it's meant to catch. If your checker can improvise, it's a second opinion, not a check.

Google can tie a closed lead back to the ad even when the click ID never reached your CRM. Enhanced conversions for leads, per Google Ads Help: 1. Your form's tag sends the email or phone, hashed, on submit. 2. When the lead qualifies in the CRM, trim, lowercase and SHA-256 hash that same email. 3. Upload it with conversion name and time to a qualified-lead goal, plus an order ID and the consent field. 4. Google matches the hash to the click. Since October, one of our product journeys reports lead quality back to Google from our own systems. Which CRM stage do you send back: the form, the qualified lead, or the sale?

JUUUTZ!

Most teams turn on Meta's Conversions API and assume the Pixel plus the server event will "just merge." They don't. Meta only deduplicates when both sides send the same event_id and the same event name, and only within 48 hours. Miss either and you train the auction on double counts. Check Events Manager → Event Deduplication before you trust the CPA.

What is love? (Baby don’t hurt me)

In our Meta ads, Oct 2–31, the cheapest quarter of ads by cost per completed application had a 57% lower conversion rate than the most expensive quarter. Each conversion still came out about 52% cheaper. Directional (last-touch). Not proof that cheap leads cause it. Just the number we actually watch next to CPL.

Anthropic just cut a status update's cost ~99% by not re-reading the whole conversation. I run the same audit on my growth bots: one bot's only job is finding where the others waste. What it cut: 1. 10 bots deleted as duplicates or useless. 2. A daily routine moved to weekdays: 7 runs a week to 5. 3. A dashboard review folded into a monitor's existing runs. Same output, one less routine. Most agent waste isn't the model. It's work nobody asked to repeat. x.com/ClaudeDevs/status/2118…

Most growth teams ship creative in an afternoon, then cut spend off a dashboard that isn't finished. Google Ads books a conversion on the day of the click. The last few days always look weak until late conversions land. Before I touch a budget I pull this myself in Google Ads — no BI ticket: 1. Campaigns table: Segment > Conversions > Days to conversion 2. Date range that ends at least 30 days ago 3. Share that lands after day 7 4. Goals > Conversions > Summary: click-through window (1–90 days). Longer only applies going forward. Credit makes the lag worse: other offers, a human ask, paperwork. What share of your conversions lands after day 7?

Salesforce just agreed to buy a voice AI company. Running one in Latin American lending taught us the voice isn't the hard part. 1. 55% of our calls never connected, so the agent writes first on WhatsApp or email and only calls if the lead wants it. 2. It continues on whatever channel the lead picks. 3. A second bot, Gordon Gecko, audits every conversation and flags where the agent slipped. The value isn't the voice. It's the system that checks the voice. x.com/salesforce/status/2117…

55% of the calls our credit AI seller places never connect. So voice isn't the default. How it works: 1. Onboarding stalls. 2. The agent writes first on WhatsApp or email. 3. The lead picks the channel: online, WhatsApp, or voice. 4. It's wired to our backend: documents, risk checks, approval. 5. Gordon Gecko, a second bot, audits every conversation. The call didn't connect. The conversation did.

Guess that we are not getting GPT Astra 6.2 today...

Grok @bot cooked with this new shopping experience! Now you can: ✅ Find the best deals in the internet ✅ Compare items before you buy ✅ Have Bot track all your orders

Our credit lines average USD $8,000 💰 At that size, most people in Latin America quit digital onboarding and ask for a human 🙋 So we needed a contact center. That means paying for headcount whether leads show up or not, retraining every time agents churn, and hiring or firing every time you need to scale. So we built a platform where an AI agent does that job. Not like a bot. Like a human seller 🤖➡️👤 How it works: 1. A lead drops out of onboarding. 2. The agent writes first on WhatsApp or email and asks how they want to continue: online, on WhatsApp, or on a voice call (ElevenLabs). 3. It continues on whatever channel the lead picks. 4. It doesn’t just pitch. It’s wired to our backend, so it asks for the documents risk needs and moves the application forward. 5. The credit gets approved without a human touching it ✅ 6. Under the hood: Gemini handles the reasoning, Supabase holds the state, and every conversation is recorded. 7. On top, Gordon Gecko, our @Bot, audits every conversation and fine-tunes the agent. Result: Cost per approved credit down 12% 📉 Commercial operating cost down 6% 📉 The surprise: 55% of calls never connected, but those leads kept going on WhatsApp 📱

Delegated authority needs more than a scope. It needs an end date and a report. A bot posted 65 times on my account under a standing approval with no expiry. It did exactly what it was allowed to: 341 impressions, nobody checking. Scope without expiry is abdication. x.com/sytaylor/status/211731…

Shared memory stops the repeat bug. The costlier repeat is a decision that looked right and wasn't. My media bots log every decision in one sheet, and a sub-agent rechecks each at 7, 14 and 30 days. Remembering what an agent did is easy. Remembering whether it worked is the memory. x.com/Aiswarya_Sankar/status…

Fri Oct 23, 8am. One step in our credit evaluation broke. For about 2 hours, ~70% of applicants who finished onboarding got an error instead of an offer. Approval rate of our credit products in that window: 0%. Tech fixed it by 10:00. We pulled back campaign spend while it was down. The day still closed with +21% approvals vs the average of the previous Fridays. That morning became Uncle Bill's playbook. He's one of my bots, built to: 1. Stream the events that make you money (signups, decisions, conversions) from your data platform to a webhook, not a nightly batch. 2. Route your monitoring's error alerts to the same webhook, so the bot sees errors next to the funnel. 3. Calculate a baseline per hour and weekday from the last 4–8 weeks. When conversions fall below the expected range while onboarding volume holds and errors spike, investigate it as broken, not slow. Escalate, open the ticket, flag the campaigns that are spending. 4. Turn spend back on only once conversions return through the same webhook. An alert tells you approvals stopped. A playbook tells your agent what to do before more budget goes out.

Same move works on ad budgets. Only one of my bots may touch a budget alone, and only on one campaign; every other change is a proposal that waits for my yes. Autonomy gets earned per surface, not per agent.