McKinsey's latest review confirms what we've been saying to mid-market clients for months: banks are burning money on AI agents. The headline is stark—token costs could hit 20% of total labor spend—but the real story is what this reveals about implementation failure. Banks are now creating dedicated AI FinOps teams to route queries to cheaper models based on complexity. This is triage, not strategy. For UK regulated firms subject to PRA SS1/23 governance requirements and FCA Consumer Duty obligations, this matters because it shows that rushing to deploy AI agents without proper architectural thinking doesn't just waste money—it creates compliance debt. You can't explain to the FCA why you're rationing AI access to control costs if your business case was built on unlimited productivity gains.

What's happening at the banking sector level is a predictable cycle: enthusiasm drives deployment, deployment drives cost shock, cost shock triggers governance. The pattern applies equally to law firms, insurers, and accountancy practices. Everyone deployed generalist agents (think broad Microsoft Copilot integrations or off-the-shelf Harvey implementations) to everything, discovered they were haemorrhaging tokens, and now they're retrofitting cost controls. This is backward. The firms winning on AI cost and compliance are the ones that started with use-case specificity—not agents that answer anything, but systems trained to handle narrow, high-value tasks precisely. They're also the ones that treat AI governance as an operational requirement from day one, not a damage-control measure after deployment.

Trovix's approach differs because we don't start with agent deployment; we start with outcome definition. Trovix Brief handles intake and matter automation by design—it knows exactly what it's solving for and operates at query cost that's a fraction of general-purpose agents, because precision doesn't require token sprawl. More importantly, firms using Trovix systems build compliance documentation contemporaneously. When the ICO audits your AI use under UK GDPR, or the SRA questions your AI governance framework, you have a clear record of what your AI does, why it does it, and what it costs—not a spreadsheet of retroactive cost controls. The McKinsey finding shouldn't surprise anyone. It should alarm you if you've already deployed without this thinking.

If you're a mid-market firm considering AI agents or reviewing existing deployments, do this now: audit what your agents actually do versus what you expected them to do. Most firms discover their agents are burning tokens on tasks that don't matter or could be handled by simpler automation. Second, ask whether your deployment is defensible under your regulatory framework—not just FCA rules or SRA Code requirements, but ISO 42001 principles for AI governance. Third, look at Trovix Audit, which gives you an operational dashboard for AI spend and compliance that doesn't require you to hire a new team. The banks in McKinsey's report are solving the cost problem; they're not solving the governance problem. You can do both if you change the question from 'How do we ration AI?' to 'What should AI actually do here?'

Source: Reuters