05On the record

Named companies · measured outcomes

The MiddleWhere it works

The case files and The Freeze document the two ways to fail. This is the band between them: named companies where AI produced measurable results — scoped, baselined, with humans kept where judgment and liability sit. Same bar as the failures: attributed figures, and an honest note on what is self-reported.

6 on file · Q2 ’26

FrozenBalancedOverheated

What the cases here share

  1. 01

    One workflow, owned

    A named owner and a single, bounded job — not a portfolio of pilots looking for a sponsor.

  2. 02

    A baseline first

    Resolution rate, handling time, error rate: measured before the switch, so the after means something.

  3. 03

    Humans where it counts

    Escalation paths and sign-off kept wherever trust, liability or judgment live.

  4. 04

    Scale what survives

    Expansion follows the numbers, not the announcement. What doesn’t hold up is cut quietly.

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.
On the recordJul 7, 2026
Works

Draft, then delegate

Octopus Energy

Energy retail · United KingdomArlo trial · Jul 2026

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.

Nov 2022

−13%

global staff cut; headcount had outgrown revenue

Jun 2026

$3.6bn

Salesforce deal for the company renamed Fin

From headcount outrunning revenue to a company named after its agent.
On the recordJun 15, 2026
Works

Product reinvention

Fin (formerly Intercom)

Customer-service software · IrelandAcquired · Sep 2026

In late 2022 Intercom was a stalling software unicorn cutting 13% of its staff. It rebuilt itself around Fin, an AI agent billed per resolved conversation, renamed the company after it and sold to Salesforce for about $3.6bn. The headline metric behind that price is one the company defines itself.

≈S$1bn

economic value attributed to AI, 2025

2,000+

models across more than 430 use cases

~4,000

contract/temp roles not renewed over 3 years

Many scoped models, each tested against controls, valued in aggregate.

A portfolio of models valued against controls — and a disclosed cost.
On the recordFeb 9, 2026
Works

Measured against controls

DBS Bank

Banking · Singapore≈S$1bn value · FY2025

DBS attributes about S$1bn of economic value in 2025 to more than 2,000 analytics and AI models, measured, it says, against control groups. Unusually, it also said in advance what AI would cost its workforce: about 4,000 contract and temporary roles over three years.

+29%

cancers detected v standard double reading

−44%

radiologists’ screen-reading workload

−12%

interval cancers (82 v 93) at two years

AI routes and flags; a radiologist still reads every scan.

More cancers found, fewer missed, with less reading.
On the recordJan 30, 2026
Works

AI triage, human read

MASAI screening trial

Public healthcare · SwedenLancet RCT · Jan 2026

In a randomised trial of about 106,000 women, AI decided which mammograms needed one radiologist and which needed two, and flagged suspicious areas. Cancer detection rose 29%, screen-reading workload fell 44% and — the result that mattered most — 12% fewer cancers surfaced between screenings.

50,000+

responses themed in ~2 hours, for £240

0.79–0.82

F1 score against two expert groups

0.74

F1 score between the two expert groups

AI proposes the themes; experts check, refine and decide.

Benchmarked against how much human experts disagree.
On the recordOct 16, 2025
Works

Benchmarked against humans

UK Government · Consult

Public sector · United KingdomBeta · rollout planned 2027

The UK government’s in-house AI tool sorted more than 50,000 consultation responses into themes in about two hours for £240, then had experts check the result. Its agreement with those experts beat the experts’ agreement with each other. It is still in beta, and the headline savings remain projections.

2.58M

patient encounters, Oct 2023–Dec 2024

~16,000

hours of documentation saved (estimated)

7,260

physicians used it; 3,447 in 100+ visits

Ambient drafting, with physician review before sign-off.

Large aggregate savings, built from small ones per visit.
On the recordApr 1, 2025
Works

Clinician-reviewed scribe

Kaiser Permanente

Healthcare · United StatesSystem-wide · since Aug 2024

Kaiser’s Northern California doctors used an ambient AI scribe in more than 2.5 million visits and saved an estimated 16,000 hours of documentation, according to a peer-reviewed evaluation. The tool drafts; the doctor edits and signs. The gains are real but modest per visit — and an independent trial elsewhere found them smaller still.

Inclusion here is not an endorsement, and a case can move along the scale. Several of these outcomes are company-reported; where they are, we say so. The middle is a practice, not a status.