Market Signals

Imagine a large global enterprise with dozens of call-centres handling thousands of chats and calls daily. Their biggest bottleneck: junior agents who are slower, more error-prone, and need intensive ramping. To close the gap, the company deployed an AI "agent assist" tool: as agents chat or talk with customers, the system listens, surfaces relevant policies and past cases, drafts response suggestions, and recommends next steps.

The result? Agents, especially newer ones, start working closer to "expert mode." In the field, that translated to ~14% higher issue resolution per hour, with >30% gains for novices. (The study is documented in an NBER working paper.)

Why it worked: The AI doesn't replace the agent—it amplifies them by reducing lookup time, easing context switching, and embedding best practices in real time.

Klarna's vision was bold: could their AI take over the bulk of support inquiries across markets? They launched a generative assistant that understood order systems, refunds, return logic, SKU mapping, policy rules—and could reply autonomously for many cases.

Within its first month, it handled two out of every three chats. Behind the scenes, each chat required orchestration: LLM layers for intent, retrieval of order/support history, and fallback to human agents for edge cases.

Why it worked: Klarna had access to rich structured and unstructured data (orders, customer history, policy docs), and they carefully scoped to only automate the frequent, well-understood cases first.

Dev teams deployed GitHub Copilot (or equivalent) to help with boilerplate, scaffolding, test generation, doc comments, and patch refactors. Because devs often lose flow when switching between code, tests and docs, the headroom is big. User studies and telemetry show up to 55% faster performance on repeated tasks; in many teams, 30–50% of code is being touched or suggested by Copilot.

Core insight: much of software work is repetitive pattern recognition. By training on open repos + private internal repo fine-tuning, Copilot learns what you expect.

In a UK government pilot involving ~20,000 civil servants, Microsoft inserted Copilot into Word, Excel, PowerPoint, and Outlook. Tasks ranged from summarising meeting notes, drafting responses, generating slide decks, to email triage. According to the report, users saved ~26 minutes per day — equivalent to 2 weeks of work per year per person. Over time, the knock-on effect is less overtime, fewer handoffs, better consistency.

"Liberating knowledge workers from administrative drag using tools they already love."

Amazon keeps pushing the envelope: combining computer vision, path optimization, AI slotting, robots, and worker interfaces to streamline fulfillment center throughput. Cameras identify items, robots shuttle goods, and the system continuously adjusts slot assignments based on demand patterns.

They report up to 25% reductions in processing time in upgraded sites, higher accuracy, and faster shipping windows.

John Lewis deployed goods-to-person systems, automated vertical storage, and robotics in its distribution center. The story: ahead of Christmas, they needed to scale without simply hiring a ton of seasonal staff. The automation enabled 75% more storage density and smoothed throughput peaks.

This kind of case is powerful to show: retailers can modernise not just for cost, but for capacity and resilience in peak seasons.

Lemonade built "Jim," an AI claims bot that ingests a customer's submission (photos, policy data, text), runs instant fraud/eligibility checks, and — in many cases — approves and pays the claim in seconds. While the headline "3-second claim" is attention-grabbing, the deeper value is in cost of service reduction and building trust via transparent AI.

They report ~30–40% of claims as instantly approved: every one you don't have to staff reduces cost and friction.

Zurich built a system to ingest catastrophe (CAT) claims (surge events) and automatically classify them in minutes. Using NLP over customer descriptions, photos, geospatial data, and policy metadata, CATIA flags which claims are "high severity," routes them to the right adjusters, and accelerates reserve decisions and reinsurance processes.

Using tools like Celonis or Process Intelligence (PI), organizations feed execution logs (ERP, CRM, workflow engines) into a miner. The miner reveals rework loops, bottlenecks, manual deviations, and exception paths. The team then prioritises high-leverage nodes (e.g. invoice exceptions, credit holds, multi-touch orders) and automates or semi-automates them with bots, microservices, or workflow logic.

In studies, such firms saw ~383% 3-year ROI and typical payback in ~6 months. Many cite millions in annual savings and working capital release via lower processing times. (These are in vendor TEI studies and independent clients.)

Lessons: start with "where you leak value," not shiny tools.