Operations AI has to prove itself where the work happens
Insurers use AI widely and prove it rarely, Waydev sells tools to watch AI agents at work, and one food maker reports a 12% yield gain. The common thread: measurement, not rollout.
- Industry data cited in Five Sigma's guide, says 82% of insurance carriers use AI for routine claims tasks, but only 7% have achieved AI success at scale in claims. Self-reported m:…7151
- Five Sigma's guide describes a 'claims AI productivity paradox': AI speeds up individual tasks but fails to reduce overall claim cycle times. Self-reported m:…7151
- Waydev says it launched three engineering-intelligence tools: the SDLC 30B model, the OLA Agent for delivery target tracking, and Turing for AI-human commit attribution. Self-reported m:…2947
- Maple Leaf Foods achieved a 12% yield improvement across multiple lines using Braincube's Real-Time Process Optimization, Braincube says, and reached return on investment in three months. Self-reported m:…7929
The paradox: widely deployed, rarely proven
Most insurers now use AI on claims, and almost none of them can show it working at scale. Industry data cited in Five Sigma's executive guide, says 82% of insurance carriers use AI for routine claims tasks, while only 7% have achieved AI success at scale in claims. Five Sigma calls the gap the claims AI productivity paradox: AI speeds up individual tasks, the guide says, but fails to reduce overall claim cycle times. Files move faster through each step and then sit idle waiting for an adjuster to intervene, because the decision-making between tasks is still queued behind people. m:…7151
The guide says most claims AI today focuses on reading, summarising and drafting tasks faster than adjusters, and generative AI in claims is limited to reading and writing, requiring adjusters to manually copy output into files. The work gets quicker in pieces and no faster as a whole. That is the pattern the rest of this story tests against: new tooling that measures the work, and one manufacturer that reports a measured result. m:…7151
Orchestration inside set limits
Five Sigma's answer is orchestration inside set limits. The guide advocates AI orchestration that connects different AI agents across the life of a claim, so work keeps moving instead of stopping between tasks. Agentic AI, the guide says, can perform specific actions such as scheduling statements or updating reserves within set permissions. Governance, it argues, should mirror human settlement authority, granted by action, line and role. The point, the guide argues, is that adjusters should spend their time on complex judgment calls while routine work moves on its own. m:…7151
Five Sigma developed Clive as a multi-agent AI claims expert for orchestration, the guide says, and the company says its automation can be integrated on top of existing claims management systems. Clive is Five Sigma's own product, presented in Five Sigma's own guide, so the orchestration argument and the product pitch arrive together. The test the guide sets for itself is whether files actually move faster end to end, not whether individual tasks get quicker. m:…7151
New tooling to watch the agents
The same measurement instinct is showing up in software engineering. Waydev says it has launched three engineering-intelligence tools: the Waydev SDLC 30B model, the OLA Agent for delivery target tracking, and Turing for AI-human commit attribution. The SDLC 30B model is a multilingual model with 30B parameters, Waydev says, and it supports on-premise deployment. The OLA Agent sets and tracks delivery targets based on user data, according to Waydev. m:…2947
Turing is an open-source tool that attributes commits to a human or an AI agent, Waydev says, and it supports Claude Code, OpenCode, Codex CLI, Gemini CLI, Factory Droid, Copilot CLI and Cursor. Turing stores commit lines as hashes to minimise data collection, the company says. The pitch is visibility: a team should be able to see what its AI agents actually did, rather than assume the work happened. Waydev also says it provides a SOC 3 report for compliance. None of these tools is independently verified in the sources used here; the claims are Waydev's. m:…2947
The one measured win
The clearest measured result in this set comes from manufacturing, not software. Braincube argues that fixed operating targets can lead to lost throughput, yield or margin when production conditions fluctuate, for example when variations in incoming material moisture lead to wasted energy and over-dried or under-dried product. Its answer is dynamic setpoints: operating targets that respond to actual production conditions, recalculated by its Real-Time Process Optimization platform, which Braincube says continuously evaluates live production data against historical performance to calculate optimal setpoints. m:…7929
Maple Leaf Foods, Braincube says, achieved a 12% yield improvement across multiple lines using the platform, and reached return on investment in three months. Those are the vendor's numbers, reported by the vendor, and they cover one customer. But they are the kind of numbers the insurance guide is asking for: a measured outcome at the point where the work happens, not a count of tools deployed. The contrast is the story. Insurers have the tools and not the results; engineering teams are buying tools that watch the tools; one manufacturer can point to a yield number and a payback period. m:…7929 m:…7151
What to ask before the next rollout
Taken together, the three sources describe the same turn. The question is no longer whether to deploy AI in operations, but whether the deployment can be measured where the work happens: claim cycle times rather than tasks completed, delivery targets rather than tools launched, yield rather than setpoints set. The open question is whether the orchestration and measurement layers now being sold will produce results like Maple Leaf Foods' in insurance and software, or whether they become another layer of deployment to count. Nothing in these sources answers that yet, and executives should treat the vendor numbers here as claims to test in their own pilots, not as results already banked. m:…7151 m:…2947 m:…7929
What is still unresolved
- The 82% and 7% figures come from industry data cited by Five Sigma; the underlying survey is not in the sources used here.
- The Maple Leaf Foods result is a single customer case reported by Braincube, with no independent verification in these sources.
- Waydev's tools are described only in the company's own launch post; no independent testing is cited.
- Whether AI orchestration closes the claims productivity paradox in insurance is unproven in these sources.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.