The Week AI Stopped Suggesting and Started Doing

If you are scoping agentic AI for a service operation, start with our 5 Pillars of AI for customer service and How to choose AI tools.

Two weeks ago I wrote that distribution is sprinting on AI and claims is walking. The pattern held. What wasn't yet visible is that the whole conversation was about to move.

In five days in mid-May, Anthropic put Claude inside QuickBooks, PayPal, HubSpot, Canva, DocuSign and the Microsoft 365 stack, with fifteen ready-made workflows. OpenAI put Codex on the ChatGPT mobile app, with live terminal output and approvals on a phone. GoCompare launched inside ChatGPT as a conversational price comparison front door. CFC kicked off Lane Assist, an agentic AI pilot inside Lloyd's underwriting submissions. The LMA published an AI adoption toolkit. ArXiv said it will ban AI-authored academic work without proper attribution.

That's a different week to what we're used to. Less news about model benchmarks. More news about deployment, governance and accountability. The conversation has gone from "what can AI do" to "what's it allowed to do, where, and under what controls".

The harness, not the model

For two years the AI conversation has been dominated by which model is best. GPT versus Claude versus Gemini versus Llama. League tables of benchmark scores. Bigger context windows. More parameters.

That conversation isn't over, but it's stopped being the one that matters most.

What matters now is what the industry is starting to call the harness around the model. Custom instructions, skills, plugins, the Model Context Protocol (MCP), APIs, command-line interfaces. The stuff that lets a model actually do work inside your tools, with your data, against your rules.

The framing of MCP as the USB-C of AI integrations is roughly right. One tool server that any compatible AI can speak to, instead of rebuilding the same integration in every app. That's why the news this week wasn't really about Claude or about Codex. It was about both of them stepping into harnesses that connect to where small and mid-sized businesses actually work.

The strategic implication is direct. The team with the cleanest harness will out-execute the team with the newest model. The new advantage is owning the workflow map, not engineering the next breakthrough.

What this means for SOS Operations

We've used the 3Rs lens (Repetitive, Rules-based, Resource-intensive) for years as a planning tool. It tells you which tasks are AI candidates. Until recently, knowing the answer wasn't enough. The connectors weren't there, the deployment surface wasn't ready, and "AI candidate" meant "thing you might pilot in nine months".

That's changed. With Claude inside the back-office stack and Codex on a phone, the 3Rs filter has moved from a planning tool to a deployment tool. If you can name the three most repetitive, rules-based, resource-intensive tasks in your operation this morning, you can almost certainly start automating one of them this month.

That's not hyperbole. It's a statement about deployment surface, not about model capability. The model capability has been there for a year. The surface to act on it is what landed in May.

A concrete example, agentic AI on the underwriting desk

CFC's Lane Assist pilot is the cleanest illustration of what this looks like in practice. It takes broker submissions out of email and turns them into structured quote recommendations on specialty lines. It's agentic in the sense that the model is doing multi-step work, calling tools, holding context across the work, and producing a draft outcome the underwriter then signs off.

What's interesting isn't that AI can do this. We've known that for a year. What's interesting is who's deploying it and where. A Lloyd's specialty underwriter, in production, with named scope and named controls. The LMA's adoption toolkit landed in the same week for a reason. The market is starting to standardise how to govern these deployments, not just whether to attempt them.

For credit hire and motor claims, the implication is direct. If underwriters can have agentic AI on broker submissions, claims teams can have it on TPI correspondence, BHR challenges, period disputes and settlement nudges. The model capability is there. The deployment surface is here. The harness is what you build now.

What to do this week

Open a blank page. List the three most repetitive, rules-based, resource-intensive tasks in your back office. Pick the one that wastes the most hours. Map it on a single page in five steps. Mark every step a human must approve.

That map is your first automation candidate this quarter. It is the harness, before you've plugged anything in. Without it, you'll buy the wrong tool. With it, almost any compatible AI will do useful work.

The model conversation is settling into a stable comparison. The harness conversation is just starting. The teams that win the next twelve months are the ones who can answer "where are we going to let AI do the work, and under what controls" with a specific operational answer, not a deck.