← All case files
On the recordArlo trial · Jul 2026

Draft, then delegate

Octopus Energy lets AI write the email — and keeps the hard cases human

Octopus Energy · Energy retail · United Kingdom

Octopus spent years using generative AI to draft emails that staff reviewed before sending. Only then did it let an assistant reply on its own — to about 4% of emails, never to vulnerable customers, always labelled. In a three-month trial it out-scored comparable human replies.

35%

of emails AI-drafted for staff review, 2024

4%

of UK emails answered by Arlo on its own

76% v 72%

satisfaction: Arlo v comparable human replies

Supervised drafting first; narrow, labelled autonomy second.

Autonomy granted narrowly, after years of supervised drafting.

The record

  • By mid-2024 about 35% of Octopus customer emails were written with help from Magic Ink, a generative-AI drafting tool built on its Kraken platform; staff are trained to review everything it writes before sending. (techUK, Aug 2024)
  • In a three-month trial, Arlo, an assistant answering routine emails autonomously, handled around 8,000 emails a week — about 4% of UK customer email — and scored 76% customer satisfaction against 72% for comparable human responses. (Octopus Energy, Jul 2026)
  • Ofgem’s consumer survey of early 2026 put Octopus’s customer-service satisfaction at 86% against a market average of 77%, and overall satisfaction at 90% against 81%. (Octopus Energy, citing Ofgem, Jun 2026)

Years of drafts before a single autonomous reply

Octopus began building generative-AI tools for customer email in 2023, and chose to put them behind its staff rather than in front of its customers. Magic Ink, built on the group’s Kraken platform, summarises calls and drafts replies; an energy specialist edits and sends. By August 2024 it was involved in about 35% of customer emails, around a third of drafts needed little or no change, and staff were trained to check every word (techUK, Aug 2024).

That sequence matters. Supervised drafting produced a large record of what a good Octopus reply looks like, measured against what staff actually sent, before any message left the building unreviewed. It kept accountability where customers and the regulator expect it — with a named person — while the company learned where the model was reliable and where it was not. Autonomy, when it came, was based on evidence rather than on a vendor’s demonstration.

Autonomy with a fence around it

Arlo, the autonomous assistant trialled in 2026, answers only straightforward emails: tariff renewals, payment dates, account details. It does not handle vulnerable customers, sensitive cases or complex complaints, which go straight to experienced staff; every AI-written email is labelled, and customers can ask for a human at any stage (Octopus Energy, Jul 2026). In the trial it answered around 8,000 emails a week, roughly 4% of UK customer email.

Within that fence it performed well: 76% satisfaction against 72% for comparable human replies (Octopus Energy, Jul 2026). The comparison holds only for routine queries, and that is the point. The company’s deputy chief technology officer, Ashley Firth, framed the design against the industry norm: “Too many chatbots have been built to keep customers away from people.” Octopus has said it plans to extend Arlo to more customers, not to more kinds of case.

Service as the metric, not headcount

Octopus has consistently presented the tools as working alongside its service team, “without replacing the humans at the heart of it” (Marketing Week, Aug 2026), and its public metrics are service metrics, not cost per contact. Ofgem’s survey of 3,911 households in early 2026 put Octopus’s customer-service satisfaction nine points above the market average; it was the only supplier significantly above average on overall satisfaction (Octopus Energy, citing Ofgem, Jun 2026).

This is the inverse of the over-automation pattern at the overheated end, where a deflection rate becomes the goal and the hardest cases get trapped in a loop. Here the hard cases are excluded by design, and the AI earns autonomy one narrow category at a time. The cost case is real, but it is treated as a by-product of letting staff spend their time where judgement is needed.

The lesson

It worked because autonomy was earned in stages: years of human-reviewed drafting before any unsupervised reply, then autonomy only for routine, low-stakes queries, with vulnerable and complex cases excluded and a human always one request away. Judge the AI on customer satisfaction, not on deflection.

How we’re reading this

Every AI-specific figure here is company-reported: the 35% drafting share, the trial volumes and the 76%-versus-72% satisfaction comparison, which covers routine queries over three months only. Ofgem’s survey is independent but measures Octopus as a whole; it cannot show what part of the lead, if any, is due to AI. Octopus has not published staffing numbers over the period, so we do not claim the tools either saved or cost jobs.

Sources

  1. 01
  2. 02
  3. 03
  4. 04
The pattern, anonymizedGood automation, bad automation

Compiled from public filings, court records, company statements and reputable press. Figures are attributed to their sources; allegations are labeled as such. Not legal or investment advice.