The problem
Energy suppliers receive a large volume of repetitive written queries: tariff renewals, payment dates, account details. Each one takes an advisor’s time, and that time is not spent on customers with complex or sensitive problems. Octopus serves 11 million households globally and calls its customer service team award-winning (Octopus Energy).
The company’s stated aim is to clear the routine work, not to cut people. “Too many chatbots have been built to keep customers away from people. We wanted to build one that does the opposite,” said Ashley Firth, Octopus’s deputy chief technology officer.
What they did
Octopus has experimented with generative AI in customer email since early 2023. In May 2023, chief executive Greg Jackson wrote in The Times that AI was doing the customer support work of 250 people, according to TheWrap’s report of his article (TheWrap). At that stage the AI drafted personalized replies that a team member could review and then send.
Arlo is the next step. It was built in partnership by Octopus’s own technology team and Kraken, the operating system Arlo is built on, which has since spun out of Octopus Energy Group as an independent company (Octopus Energy). It answers straightforward emails about tariff renewals, payment dates and account details.
The design choices are the core of the case:
- Narrow scope. Arlo deals only with routine enquiries and works within strict guardrails.
- Hard exclusions. It never handles vulnerable customers, sensitive cases or complex complaints, which go straight to experienced members of the customer service team.
- Disclosure. Every AI-written email is clearly labelled.
- An exit to a person. Customers can ask to speak to a human advisor at any stage.
The company describes a careful, step-by-step introduction, and positions Arlo as support for its team rather than a replacement.
Results
During the three-month trial, Arlo handled around 8,000 emails a week, which Octopus says was 4% of all customer emails it received in the UK (Octopus Energy). The company reports that Arlo’s replies achieved a 76% customer satisfaction score, against 72% for comparable responses from human advisors. These figures are self-reported: Octopus has not published how it chose the comparable human responses, the survey response rate or the trial dates.
The 2026 numbers are lower than an earlier claim. In 2023, Jackson said AI-drafted replies scored 80% customer satisfaction against 65% for trained staff, as reported by TheWrap (TheWrap). That figure covered a different setup, in which staff could review drafts, so the two are not directly comparable. Both point the same way: on routine queries, customers rated AI-written answers at least as well as human ones.
Octopus has not published cost savings, response times or the number of advisor hours released by Arlo.
What didn’t go to plan
Public sources report no setbacks from the trial. The open questions are about what the numbers leave out. The 4-point satisfaction gap compares Arlo’s answers with human answers to “comparable” emails, but the human advisors also handle the harder cases Arlo is barred from, so the comparison may flatter the AI. A 4% share of email shows the scale is still small. Expanding beyond routine queries would test the guardrails that made the trial work, and a misrouted vulnerable customer would carry regulatory and reputational risk for a supplier. The quoted customer praise comes from the company’s own release.
Lessons
- Fence the scope before you launch. Arlo answers only routine enquiries and never handles vulnerable customers or complex complaints.
- Label AI replies and keep a route to a person. Octopus labels every AI-written email and lets customers ask for a human at any stage.
- Compare against a human baseline. Octopus reports Arlo’s 76% satisfaction next to 72% for comparable human responses, not on its own.
- Grow in steps. The company moved from staff-reviewed drafts in 2023 to a three-month trial covering 4% of email before planning wider use.
Adapt it in your organization
Is this for you?
- A large share of your inbound email is routine: dates, balances, renewals, account details.
- You can identify vulnerable customers and complex cases reliably before routing.
- You already measure satisfaction on individual replies, so a baseline exists.
- Your customer records sit in a system the assistant can read accurately.
- You operate under consumer-protection rules that require a clear route to a person.
A 90-day version
Pilot: let an AI assistant answer one category of routine email, such as payment-date questions, labelled as AI-written, with an opt-out to a human and automatic exclusion of flagged vulnerable customers. Owner: head of customer operations. Success measure: satisfaction on AI replies at or above the human score for the same category, with no rise in repeat contacts.
Metric to track
Customer satisfaction on AI-written replies against human replies to the same query type, plus the share of email handled. Octopus reports 76% against 72% on 4% of its UK email. For a contrast where scale came first and quality concerns followed, see the Klarna case.
Failure modes to pre-empt
- Comparing AI replies on easy emails with human replies on all emails.
- Letting a vulnerable customer reach the assistant because the flag was missing.
- Hiding that a reply was AI-written, which undermines trust when customers find out.
- Expanding scope faster than the guardrails are tested.
Questions to take to your leadership team
- Which email categories are routine enough to automate, and how do we know?
- How would we detect a vulnerable customer the assistant should never have answered?
- What would we do with the advisor time released?
Executive FAQ
Did Octopus Energy’s AI beat human agents on satisfaction?
On routine emails in its trial, Octopus says Arlo scored 76% satisfaction against 72% for comparable human responses.
Does Arlo replace Octopus’s customer service team?
No: Octopus says Arlo handles only routine enquiries, with complex and sensitive cases going straight to its team.
About the curator & AI disclosure
Curated by Luciano Ayres
Luciano Ayres is Director of Artificial Intelligence at ArcTouch, with more than 20 years of experience in software and technology leadership and specialized studies in deep learning. Connect with him on LinkedIn.
AI-generated briefing. This case was produced through an automated AI research and writing workflow curated by Luciano Ayres. Sources are linked for independent verification; summaries and analysis may contain errors and do not replace professional due diligence.