Jamie

Support lead at a B2B SaaS. I post about ticket deflection edge cases, macro fatigue, and where AI reply drafts miss what the customer actually meant.

[Jamie @jamieeem]

@jraphael10 we did this with a refund-flow macro. four months later no one knew who owned the carve-outs. shared skills age weirdly.

[Jamie @jamieeem]

the worse one had no dollar to flag. automation rewrote a customer field. found it from a csat dip three weeks later.

[Jamie @jamieeem]

first time an automation triggered another in our queue, refund came out 4x the ticket. finance caught it.

[Jamie @jamieeem]

same shape in support. bot handles the easy 80%. the team only sees the hard fifth and AHT keeps climbing.

[Tomasz Tunguz @ttunguz]

A laptop on my desk now handles 78% of my AI work. The cloud only gets the hard fifth. 🧵

[Jamie @jamieeem]

@PreslavMihaylov cs version of this hits the same wall. week 3 everyone trusts it, nobody re-reads the macros it's rewriting.

[Jamie @jamieeem]

in two of those meetings someone asked who handles the false positive. silence. that's where it stalls every time.

[Jamie @jamieeem]

who owns proactive AI alerts at your company, support or growth? at ours it's been three meetings and no answer.

[Jamie @jamieeem]

wait til product q&a quotes the wrong return policy and a million merchants ship the same wrong answer the same hour.

[Hedgie @HedgieMarkets]

🦔Meta just launched Business Agent, an AI that handles customer questions, product recommendations, appointment booking, lead qualification, and sales on WhatsApp, Messenger, and Instagram. Over a million businesses already use it. It's free to start, with paid subscriptions coming. The Information reports Meta plans to charge up to $200 a month for its planned "Hatch" AI agent. A basic website with a contact form has done most of this for a decade.

My Take
Meta cut 2,212 people from Menlo Park. Its AI support chatbot spent months giving away Instagram accounts to anyone who asked politely. And now the company wants small businesses to pay $200 a month to let a Meta AI agent talk to their customers unsupervised. The same week Microsoft's internal strategy doc said step one is "make people addicted," Meta is running the business version of the same play. Get a million businesses on the free tier, let them restructure around the tool, then flip on the subscription.

This reads like a 2015 Zendesk demo with "AI" stapled to it. Businesses have automated customer Q&A, appointment booking, and product recommendations for years with simple chatbots and web forms that cost a fraction of $200 a month and don't carry the risk of an LLM going off script with a paying customer. Meta needs this to work because the ad business that funds everything just lost the engineers who built it, and the AI spending has to produce revenue somewhere. Right now that somewhere appears to be small business owners who don't have the context to evaluate what they're actually getting.

Hedgie🤗

[Jamie @jamieeem]

@productwatch_io the drowning is real. but when the bot eats easy tickets, the team only sees the hard ones. handle time goes up, not down.

[Jamie @jamieeem]

@ZenoEleatis we shipped a bot that 'helped' callers change recovery email if a kb article said it could. blocked all auth intents by week 2.

[rohan @RohanVasishth]

"The dream is to be the system of record for customer support that really helps you identify how to make it really good."

This week on Skywatch, @getbluejay_ai's car podcast, I sat down with @johnjianwang, co-founder and CTO of @assembledhq, and we had a conversation about what it actually looks like when AI agents meet enterprise customer support at scale, and what most people are still getting wrong.

John did not start out building customer support software. After an MIT roommate steered him toward coding over economics, he finished school early for YC and joined Stripe. There, he and his brother developed an internal system saving millions, only for a PM to prioritize a Bitcoin project instead. That pivot point is what built Assembled.

A few highlights from this conversation I have not stopped thinking about:

👉Although a fully automated experience is ideal, any lapse in service causes customer satisfaction to fall off a cliff, highlighting that knowing exactly when to involve a human is the greatest challenge in customer support AI right now.

👉 While a top-tier human agent might still outperform AI today, the technology excels by delivering what John calls the Chick-fil-A effect, a level of absolute consistency that ensures every customer knows exactly what to expect, every single time.

👉John's analysis of his customer base revealed that 75 to 80 percent of companies implementing AI agents increased their headcount, due to automation driving higher volume, creating more complex support tickets, and requiring long-term change management.

If you are building in customer support, voice, or enterprise AI, this one is worth your time.

Full episode on YouTube and Spotify in the comments!

[Jamie @jamieeem]

hit 42% deflection last year. next 5 points took 4 months of edge cases. anyone past 50 without hardcoding answers?

[Jamie @jamieeem]

the cleanup queue rarely has an owner. our bot double-emailed last quarter, surfaced in csat two weeks later, not on any dashboard.

[Haktan Suren, PhD @HaktanSuren]

AI agent governance is not a policy PDF.

It is knowing who gets paged when the agent is confidently wrong.

Lead routed.
Discount approved.
Record overwritten.
Customer emailed twice.
Workflow retried 14 times.

Autonomy is exciting until the cleanup is manual.

[Jamie @jamieeem]

@HuseyinTheBrain the catch: when AI auto-summarizes and routes, the engineer-needing tickets end up tagged 'medium' and buried. you're triaging the triage now.

[Jamie @jamieeem]

@luisgomes21 we shipped a kb-backed bot. 6 weeks in half its answers were from a sunset feature. nobody flagged it til CSAT tanked.

[Jamie @jamieeem]

the boomerang isn't even tracked as a reopen, it's a new ticket. dashboard says 4 deflected. nobody's linking them back.

[Jamie @jamieeem]

audited 30 AI tickets last week. 4 closed clean, customer thanked the bot, then reopened two days later. same issue. deflection up. retention not.

[Jamie @jamieeem]

true on volume. but the weird requests we now handle are also the audit pile. half are calls the agent shouldn't have made.

[MorningAI @MorningaiBrief]

Operators who understand AI have an unfair advantage right now.

Not because they can build models — they can't. But because they understand what AI is actually good at.

The gap is huge. 🧵

[Jamie @jamieeem]

@elsontec tried to ship priority auto-tag last quarter. three CSMs labeled the same ticket type three different ways. the model learned the inconsistency.

[Jamie @jamieeem]

@luisgomes21 for us the truer metric is escalation accuracy. when the agent says human needed, was that call right? we audit it weekly.