🎧 Prefer to listen?

Klaviyo just bought itself an AI brain. The e-commerce marketing platform acquired Agency, a customer-success AI startup founded by serial entrepreneur Elias Torres, who now joins as chief product officer. The plan: fuse Agency’s tech with Klaviyo’s existing agents — Composer, which builds marketing campaigns, and Customer Agent, which handles post-sale questions like returns and order tracking — and roll them out to 200,000 businesses. Torres spent $32 million of investor money (Sequoia, Menlo, Felicis) learning how to build this. You get to skip that bill.

If you caught my coverage of Asana acquiring StackAI, you know the pattern: the platforms you already pay for are absorbing agent capabilities and shipping them to millions of users by default. This one matters if you sell anything online. Customer Agent answering “where’s my order?” and “can I return this?” for you is genuinely useful — until it answers wrong, promises a refund, or mishandles an angry customer at 2am with no human watching.

Here’s what the acquisition really signals, and the prep work that decides whether AI support makes your store better or just cheaper-feeling.

What I actually did (mapped the rollout against a real store)

I took a mid-size Shopify-style store setup and mapped what Klaviyo’s announced agents would actually touch, then tested how each would behave with the store’s current state of organization. The picture:

  • Order tracking is ready today. This is structured data — order numbers, statuses, dates — and agents handle structured data well. If your store’s order system is even half organized, Customer Agent can absorb the “where is my stuff?” tickets immediately. That category is typically 30-50% of support volume for physical-product stores.
  • Returns are where it gets spicy. A returns policy is rules plus judgment: “unopened, 30 days, except final sale, but we bend for good customers.” Agents follow rules fine; the bending is what your data has to teach. A store with clean, written, consistently-applied policies gets a consistent agent. A store with verbal exceptions gets an agent that makes promises nobody authorized.
  • The data moat is yours, not theirs. Klaviyo’s pitch is that years of customer data give its agents an edge over AI-support startups like Decagon and Sierra. But notice whose data: yours. The agent is only as good as your CRM history, your past support threads, your email engagement patterns. Which means the prep work isn’t learning new tools — it’s cleaning what you already have.

The reunion angle is a great story (Torres hired Klaviyo’s CEO back in 2010; the CEO later made him an angel investor), but for a solo builder the headline is simpler: agent support is becoming a default feature, like SSL. You don’t decide whether to have it. You decide whether it’s any good.

Do this yourself (the 5-step agent-readiness sprint)

Do this before the rollout reaches you — it takes one focused afternoon:

  1. Write your policies like a robot has to follow them. Literally. Open your returns, shipping, and warranty policies and check every sentence against: can this be executed exactly as written? Kill phrases like “case by case” unless you also define the cases. This is the same discipline as the instructions in my Grok Bot delegation test — ambiguity is where agents fail publicly.
  2. Feed the agent your real history. Past support emails, product FAQ, sizing guides, common confusions (the ones customers actually ask, not the ones you wish they asked). Klaviyo’s agent advantage comes from this data, so make sure yours is in it and accurate.
  3. Set escalation rules before launch, not after. Define what the agent must always hand to a human: refunds above a threshold, legal language, anything with “chargeback,” anything angry. Write these as hard rules, not vibes. This is the non-negotiable layer from the agent security gap post.
  4. Run a two-week shadow mode. Let the agent draft responses to real tickets but have a human approve them, and log the disagreements. The disagreements are your policy gaps. This mirrors the eval habit from my AI eval harnesses post — measure before you trust.
  5. Keep the tone yours. Agents learn voice from your data. If your brand is warm and a little blunt, make sure past emails reflect that, or the agent will default to polite-corporate and your customers will notice.

If you’ve never set up automation at all, do build your first automation in 15 minutes first — agents are easier to trust when you’ve already built the boring version by hand.

What failed

  • “AI handles everything” is still fiction. In my mapping, roughly two-thirds of tickets were agent-ready; the rest — disputes, edge cases, emotional conversations — needed a human from the first reply. Budget for that, don’t pretend otherwise.
  • Data preparation is the actual work. The tools are one-click. Cleaning three years of inconsistent order notes and thread-hijacked email chains is not. If you skip it, the agent inherits your mess and repeats it confidently.
  • Rollouts arrive whether you’re ready or not. Platform-shipped features don’t wait for your consent the way tools you install do. Your prep window is now, between the acquisition and the default-on moment — same lesson as every platform AI push I’ve covered in ChatGPT’s scheduled tasks.

Who should skip this

Skip the prep sprint if you don’t sell anything and never will — this is an e-commerce story. Skip the panic if you already run a disciplined support operation with written policies and tracked tickets; you’re most of the way there, and the agents will mostly amplify your system. And if your business is high-touch consulting where every customer conversation is the product, automated support is the wrong instinct entirely — consider the angle in my AI for small business post instead: agents for prep and follow-up, humans for the room.

The bottom line

Klaviyo buying Agency means agent-based customer support is about to become a checkbox in platforms you already use — for 200,000 businesses, then millions more. The winners won’t be the stores that bolt on the feature; they’ll be the stores whose policies, data, and escalation rules were ready before the switch flipped. Spend the afternoon on the five steps. Your future agent will only ever be as good as what you feed it — and right now, that part is still your job.

New to AI tools entirely? Start here.