twimco.Let’s talk ↗︎
Menu
Field notes / Reviews

CALL-E review: an AI that makes your phone calls once you approve the plan. Read the call report before you trust it

Ask your AI assistant to cancel a subscription or call three suppliers for a quote, and CALL-E will dial for it, but only after you've approved what it plans to say. The part worth your time is what comes back when the call ends.

Editorial engraving of an owner reading a one-page printed report at a sunny desk with her phone beside a short checklist while an assistant waits in the doorway.

CALL-E, launched by AI Rudder on November 14, 2026, lets an AI assistant plan a phone call, show you the plan, dial after you approve it, and return a summary, transcript and structured result with a confidence label. It fits low-stakes errands in the US and Canada. Keep irreversible calls, like cancellations and payments, out of it until you've read a few full transcripts against what the summary claimed.

What CALL-E does

AI Rudder, a Singapore company that has deployed voice agents for financial-services work in Southeast Asia, launched CALL-E on November 14. The idea is simple. You're already talking to an AI assistant, such as ChatGPT, Claude Code, Codex or Cursor. You tell it what you need done by phone. CALL-E plans the call, you confirm, it dials, and you get back what was said and what happened. Developers can reach it through an API, SDKs and MCP. AI Rudder's announcement says it works with those tools, though it doesn't claim a partnership with their makers.

The typical errands in CALL-E's own description are the ones that eat an afternoon: getting through a phone menu to a person, changing or cancelling a subscription or order, confirming or moving an appointment, checking opening hours or stock, calling around for quotes, leaving a voicemail. Before it dials, it asks for anything missing, including who to call, when, in what language and what counts as done. During the call it handles pickup, voicemail, hold, transfers, silence and interruptions, and it can press keys to move through a phone menu.

Calls can be scheduled or sent to several numbers as a batch. That's convenient, and it's also where a mistake multiplies: a vague plan sent to forty numbers is forty vague calls. Run a batch of two before you run a batch of forty. Coverage is local numbers in the US and Canada. Twenty more countries can be reached on international numbers, which CALL-E says are mainly for development and testing.

The plan screen, and the report

The best design decision here is the approval step. The assistant shows the call plan before anything is dialed, and AI Rudder says that is its safeguard against an agent making a call you never authorized. Read that screen like a short contract. Does it name the right company and number? Does it say what the agent may agree to, and what it must not? Does "done" mean "got a quote" or "accepted a quote"?

Then comes the report, which is where we'd spend the time. It's also the part you can build a routine around, because it arrives in the same shape every time. Each call returns a summary, the full transcript, and a result in a shape you define, with quoted lines from the call as evidence and a confidence label about whether the job was completed. That's a good set of ingredients. A summary tells you what the agent thinks happened. The transcript lets you check. TechNode's report adds a useful caution: a request being accepted by the service only means it entered the workflow. It doesn't confirm the person answered or that the task worked.

So the habit to build is boring. For your first twenty calls, read the transcript next to the summary. Notice where the summary rounds a "maybe" up to a "yes," or where a quoted line doesn't say what the result field claims. A call summary is a claim. The transcript is the evidence.

How the work moves

One call, start to finish, as CALL-E describes it

  1. You ask your assistant

    From ChatGPT, Claude Code, Codex, Cursor or the API: "Cancel this subscription" or "Get three quotes."

  2. CALL-E fills the gaps

    It asks who to call, when, in which language and what counts as done.

  3. You approve the plan

    It shows the call plan before dialing.

  4. It makes the call

    Handles pickup, voicemail, hold, transfers, silence and phone menus.

  5. You get the report

    A summary, the transcript and a structured result with quoted lines and a confidence label.

Steps as described in CALL-E's own documentation (source: heycall-e.com llms.txt and docs.heycall-e.com); not a measured result.
A call summary is a claim. The transcript is the evidence.

What it costs, and what's unproven

Pricing is published and usage-based. A one-shot call costs a model fee of $0.0296 per ten seconds plus a carrier fee of $0.04 a minute for US and Canadian numbers, billed in whole minutes. CALL-E's own examples are $0.1288 for a 30-second call and $0.2176 for a minute. Reusable "Goal" runs, which the docs recommend for batches, cost half as much per call. Model fees can apply even if the call doesn't connect, and new accounts get 100 free credits. At those rates, five supplier calls are cheap. The real cost is whoever reads the results.

What isn't established is how well it handles the awkward calls. AI Rudder cites a leaderboard result for how well its model handles interruptions, but TechNode notes that the claim came from the company and wasn't independently reproduced. CALL-E introduces itself as an AI voice agent by default. Users can change the introduction, subject to a check meant to block pretending to be emergency services or the police. And AI Rudder itself says consent, disclosure and recording rules differ by place and use. That last point is yours to settle. Know what your state requires before you record or place calls for your business.

One more practical detail: CALL-E's docs say version 1 of its calls interface retires on December 31, 2026. If a developer builds on it, ask which version.

How it compares

The obvious alternative is making the call yourself, or handing it to someone on your team. That's slower and costs wages, but a person hears the hesitation in a voice and knows when to stop. A human assistant or answering service sits between the two, with judgment and a price to match.

CALL-E sits closer to a very patient intern who always writes things down. It's good for repeatable, checkable calls, and weaker as the stakes and the ambiguity rise. One thing it can't replace is the relationship. If a supplier knows your voice and bends a rule for you, a first call from an agent may not get the same answer, and you won't learn that from a confidence label.

Decision guide

Which calls to hand over first

ConsiderationFine to try earlyWait until you trust the reports
InformationOpening hours, whether an item is in stockAnything a regulator or contract treats as a statement
QuotesCall five suppliers and compare the answersAccepting a quote or agreeing to terms
AppointmentsConfirm a time you can easily moveCancelling something that can't be rebooked
MoneyAsking what a late fee would bePaying, refunding or changing a billing detail
If the report is wrongYou lose a few minutesYou lose money or a customer
Decision framework: sort your own calls by what happens if the report is wrong; examples are illustrative.

Plain verdict

Worth a trial, on low-stakes errands, with someone reading the transcripts. Start with information calls and quotes, which cost you little if one goes wrong. Keep cancellations, payments and anything contractual out of it until you've seen enough reports to know how often the summary and the transcript disagree. Read the plan before every approval. Check your recording and disclosure rules first. And before the first call, write down who reads the reports and how often a person spot-checks one against its transcript, because a tool that saves time only helps if somebody owns the checking.

If you're building something similar into your own product or operations, where an agent makes calls and a person signs off, the design work is in the approval step and the record it leaves. That's the part we do in our AI applications and agents work. Talk to us.

Related work: Spin:Market

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎