BRENNEN

Industries / Finance and business services / Call centers and BPO

Back-office processing

Back-office processing is the work that outsourcing firms and in-house shared service centers do away from the phones: typing data in, checking forms and claims, matching invoices to orders, checking new customers' identity documents (KYC, "know your customer"), reviewing posts against a platform's rules, and labeling data for AI.

7

Milestones achieved, from 1964 to 2024.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1964

    Keying moves offshore

    Time described Caribbean Data Services, an American Airlines subsidiary in Barbados, where about 500 keypunch operators typed up passenger ticket coupons flown in each morning, about 1,100 pounds a day, and sent the data by satellite to American's computer in Tulsa, for $2.20 an hour against the $9 an hour American had paid US operators. The same article dates earlier moves: ILM Corp. founded a Jamaican subsidiary in 1964, and in 1972 John Hancock Insurance hired Key Universal to have workers in Grenada computerize its records.

    LessonOnce paper could be shipped out and data sent back by wire, wage rates decided where the work was done.

    time.com

    Credit

    American Airlines (Caribbean Data Services); ILM Corp

  2. 1997

    Captive back offices in India

    Pramod Bhasin, then head of GE Capital in India, built a small division to do back-office work for all of GE Capital, which Harvard Business Review's editors call "the beginning of Genpact—and of an entire industry"; Genpact dates its start inside General Electric to 1997, its spin-out to 2005 and its listing to 2007. WNS was founded in 1996 as an in-house unit of British Airways, and in October 2025 Capgemini completed its purchase of WNS for $3.3 billion.

    LessonA captive (an in-house center serving one parent company) can be sold as an independent vendor once its processes are proven.

    hbr.orggenpact.comclearygottlieb.comcapgemini.com

    Credit

    Pramod Bhasin and General Electric; British Airways

  3. 2009

    An outsourcer's books collapse

    In January 2009 the founder of Satyam Computer Services, B. Ramalinga Raju, said its revenue and profit figures had been inflated for several years; India's government took charge, ordered the accounts restated and picked Tech Mahindra to invest. In April 2011 the US Securities and Exchange Commission settled charges that Satyam had overstated revenue, income and cash by more than $1 billion from at least 2003 through September 2008; Satyam paid a $10 million penalty and agreed to an independent review of its controls, and its former auditors were sanctioned separately.

    LessonA vendor's audited accounts do not prove the vendor is sound, so clients need their own checks and an exit plan.

    research.seed.law.nyu.edupcworld.com

    Credit

    US Securities and Exchange Commission

  4. 2015

    Software robots in back offices

    Telefónica O2 in the UK had moved back-office work to a BPO provider in Mumbai from 2004 (375 full-time staff by 2009), until, as volume passed a million transactions a month, its back-office head said "low cost wasn't so low anymore". After a 2010 trial it adopted Blue Prism's robotic process automation (RPA, software that logs into the same systems a clerk uses and repeats rule-based steps), and by April 2015 ran over 160 "robots" handling 400,000 to 500,000 transactions a month, about 35% of back-office transactions, with a three-year return on investment of 650–800% by the researchers' count.

    LessonEY wrote in 2016 that it had seen "as many as 30 to 50% of initial RPA projects fail", mostly from planning and governance mistakes, not the technology.

    researchonline.lse.ac.ukeyfinancialservicesthoughtgallery.ie

    Credit

    Wayne Butterfield's back-office team at Telefónica O2, with Blue Prism; case study by Mary Lacity, Leslie Willcocks and Andrew Craig (LSE Outsourcing Unit, with Blue Prism as research sponsor)

  5. 2020

    Moderators' harm reaches the client

    Facebook used at least five outsourcing vendors in at least eight countries to review content, by a Reuters count, and in October 2019 one of them, Cognizant, said it would leave parts of that work, cutting about 6,000 jobs, after The Verge's reports of conditions at its sites. In May 2020 Facebook agreed to a $52 million settlement with 11,250 current and former US moderators (at least $1,000 each, up to $50,000 for diagnosed conditions) and to tool changes such as muted audio and black-and-white video by default; in June 2023 Justice Byram Ongaya of Kenya's employment court held that Meta was the "primary and principal" employer of 184 moderators hired through Sama, which he called "merely an agent".

    LessonOutsourcing a harmful task does not outsource liability for it.

    m.thewire.inengadget.comtechcrunch.com

    Credit

    Selena Scola (lead plaintiff); The Verge; Kenya's Employment and Labour Relations Court

  6. 2023

    Data labeling becomes offshore piecework

    Data labeling for AI means tagging images, telling pedestrians from palm trees in driving videos and editing chatbot text. In August 2023 The Washington Post reported that at least 10,000 people in the Philippines did this work on Remotasks, a platform owned by Scale AI; of 36 workers interviewed, all but two said payments had been delayed, reduced or canceled, and the Oxford Internet Institute's Fairwork assessment scored Remotasks 1 out of 10. Scale AI said payment delays were "exceedingly rare".

    LessonGig platforms can move back-office work offshore without the contracts and protections of a BPO center.

    stripes.com

    Credit

    Rebecca Tan and Regine Cabato (The Washington Post); Oxford Internet Institute (Fairwork)

  7. 2024

    A test for software-driving agents

    OSWorld, published in April 2024, set AI agents 369 tasks in real desktop and web applications; humans completed 72.36% and the best model 12.24%. In October 2024 Anthropic released "computer use" in public beta, in which the model reads screenshots, moves the cursor and types; it scored 14.9% on OSWorld from screenshots alone. By July 2025 the top agent verified by the OSWorld maintainers on their own machines scored 60.76%, but on the harder OSWorld 2.0 (June 2026), with 108 long workflows, the best agent completed 20.6%.

    LessonScores on short tasks rose fast, but long multi-step work, the shape of most back-office processes, is still far from reliable.

    arxiv.organthropic.comxlang.aiarxiv.org

    Credit

    Tianbao Xie, Tao Yu and co-authors (XLANG Lab); Anthropic

Watch for

  1. Agents settle small insurance claims

    Allianz's Project Nemo, launched in Australia in July 2025, uses seven AI agents (planner, cyber, coverage, weather, fraud, payout and audit) to handle food-spoilage claims under A$500; Allianz reports "an 80% reduction in claim processing and settlement time", from several days to a day or a few hours, with a claims professional still making every payout decision and no baseline data given.

    Achieved whenan auditor, regulator or published study reports processing times, error rates and fraud missed on a stated number of claims, compared with the human process.

    allianz.com

    Claimed by

    Allianz SE, November 2025

  2. Agent-run accounts payable

    Genpact made its AP Suite generally available, four modules for capturing, processing, tracking and answering questions on supplier invoices, each run by "a network of AI agents", and claims more touchless processing (invoices no person handles) and up to 90% capture of early-payment discounts.

    Achieved whena client or independent auditor publishes touchless rates, error rates and the share of invoices passed to people, over a stated period and volume.

    seekingalpha.com

    Claimed by

    Genpact, June 2025

  3. Human-level score on OSWorld

    Simular Research reports that its Agent S3 with a "Behavior Judge" reached 72.6% on OSWorld, just above the 72.36% human figure, by keeping the best of 10 attempts from two models; a single run averaged 62.6%.

    Achieved whenthe OSWorld maintainers list a verified single-attempt score above 72.36%.

    arxiv.orggithub.com

    Claimed by

    Simular Research, February 2026

Open

  1. 01

    A green-screen keying test

    A mock claims-entry application on an emulated IBM mainframe, 50 synthetic scanned claim forms (some faulty and meant to be rejected), and a harness in which an AI agent reads each form and keys it through 3270 terminal screens, the text-only "green screens" of mainframe software. A checker reads the application's data file and scores every field.

    The brief
    Credit · unclaimed
  2. 02

    One recording, a working automation

    A pipeline that watches one screen recording of a person doing a back-office web task, such as updating an order in a shop's admin panel, writes a step-by-step procedure (an SOP, standard operating procedure) and a script, then runs the script on 10 new cases with different inputs, scored by the share that end in the right state.

    The brief
    Credit · unclaimed

Open 01 · the brief

A green-screen keying test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A mock claims-entry application on an emulated IBM mainframe, 50 synthetic scanned claim forms (some faulty and meant to be rejected), and a harness in which an AI agent reads each form and keys it through 3270 terminal screens, the text-only "green screens" of mainframe software. A checker reads the application's data file and scores every field.

Why it is still open

IBM's 3270 protocol is "still commonly used" through emulators for "green screen applications", by Wikipedia's account, and its field-based screens were built for form-style data entry. Public agent benchmarks test other software: OSWorld covers desktop and web applications. Servers that let AI agents drive 3270 and 5250 sessions appeared on GitHub in 2026, such as claude-tn3270 and tn5250-mcp, but they ship tools, not task sets or published tallies.

en.wikipedia.orgarxiv.orggithub.comgithub.com

Built with what exists

The SDL Hercules 4.x Hyperion mainframe emulator; MVS 3.8j built with MVS-sysgen; KICKS, a free "CICS replacement for MVS 3.8"; x3270/s3270 driven from py3270; any agent model that can call tools or use a computer; synthetic forms generated in Python; a small extraction model such as Meta Muse Ember for the form-reading step, for comparison (see the Note "The jobs agents repeat").

  1. First step

    Bring up Hercules with MVS 3.8j and KICKS, write one KICKS transaction with one entry screen backed by one file, and write a script that dumps the file to compare with a form's answer key.

  2. Done looks like

    50 forms, 20 runs per form per agent, and a published tally of fields keyed correctly, bad forms rejected and wrong records committed, plus how many forms each agent got right on all 20 runs (see the Note "All twenty times").

Open 02 · the brief

One recording, a working automation

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A pipeline that watches one screen recording of a person doing a back-office web task, such as updating an order in a shop's admin panel, writes a step-by-step procedure (an SOP, standard operating procedure) and a script, then runs the script on 10 new cases with different inputs, scored by the share that end in the right state.

Why it is still open

WONDERBREAD (NeurIPS 2024) published 2,928 recorded demonstrations, each with a screen recording, an action trace and a written SOP, but scores generated SOPs only against reference SOPs and does not run them on new cases. ECLAIR reported 40% end-to-end completion and names validation as an open challenge, and Sharingan turns desktop recordings into action sequences that RPA can replay, which repeats the recorded case rather than handling a new one.

arxiv.orgarxiv.orgarxiv.org

Built with what exists

WONDERBREAD's recordings, traces and SOPs (Apache 2.0); the WebArena sites they come from (a Magento shop admin, GitLab and a forum); a vision-language model to read the video; Browser Use 1.0, which writes Python over the Chrome DevTools Protocol, to produce scripts that can be rerun (see the Note "Deleting 80% to start over").

  1. First step

    Download WONDERBREAD, pick 10 workflows from the shop-admin site, and write a generator that makes 10 input variants per workflow with a checker for the end state.

  2. Done looks like

    30 workflows with one recording each, each SOP and script run on 10 unseen input variants, with the published pass rate per workflow, the number of scripts that pass all 10, and a comparison with scripts written by hand from the same recording.