Newsletter

How to test-drive the self-driving company

Subscribe on Substack

You enter your email address on Substack. Unsubscribe at any time.

What a self-driving company looks like today

It sounds like the future. But the future is already here. Replit calls it the self-driving company. People set the goals, use their judgment and stay accountable, and agents do the work through the company's own systems. The company sets the limits on when the agents may act, and when they must ask. At BNY, nearly 140 agents operate as digital employees, where each has a login and a human supervisor who reviews their work. Mercury's rule, as Lex Sokolin reports, is that an agent never gets more access than the person it works for, and it can prepare a payment but never send one.

It doesn't always go to plan. Ethan Mollick left agent teams polishing a project and found that "the AI had built a bureaucracy", a rule book of 204 sections after 12 days. Only 5% of the US leaders Deloitte surveyed between April and June 2026 say their processes are highly prepared for agents, and the UK's CMA, FCA, ICO and Ofcom have said that an agent acting alone doesn't reduce our responsibility for what it does.

So the first job is ours as leaders: deciding which calls an agent may make alone, which need a yes from a person, and who owns that rule book and the log.

How to run a 30-day test in your company

Britain's National Cyber Security Centre advises starting small, with one low-risk, tightly bounded task. Here's how we'd run it:

  1. Pick one repetitive task with a clear right answer, name the person who owns the agent, and record how long the task takes today.
  2. Write the agent's rule book on a page: what it may do alone, what needs a yes, and what it must never touch.
  3. Build it where your data already lives: Copilot Studio for Microsoft 365, Gemini Enterprise for Google Workspace, Claude Cowork on your own desk, or n8n if you want every action approved in Slack or Teams first. Engineers can use the OpenAI Agents SDK.
  4. Give it its own service account or login with read-only access, and test it on a batch of past cases before it sees a live one.
  5. Run it supervised in office hours, keep the log, and check that someone can switch it off at once.
  6. At day 30, compare it with the baseline and decide whether to widen it, fix it or stop.

When Asurion did this with invoice exceptions, the share handled without a person rose from under 20% to 55% in the first week.

Also worth your time

A&O Shearman accelerates AI strategy with a USD1.1 billion lawyer-led commitment, A&O Shearman. The law firm is putting $1.1bn into AI over five years, most of it as time rather than cash: every practice gives at least 3% of its time to building AI tools for its own area of law, and a qualified lawyer stays responsible for every piece of advice. In September, 96% of its partners used the firm's AI tools. This example shows us funding AI from the business rather than the IT budget.

2026 Core Accounting Software Market Map, Mostly Metrics. CJ Gustafson asks finance-system vendors how they let agents near the general ledger, and the careful ones keep the accounting itself rule-based and treat the AI as just another user, with people approving payments and anything material. For our CFOs, it's the rule book above: an agent uses all the access it's given, so give it less.

Is your leadership team ready to light the fire?

Upskill with AI Firestarter bootcamps and workshops, built for leaders who are too busy leading to be experimenting with AI.

As a leader, you set the tone for how AI is used in your organisation.

Until next time,
Ryan Snyder
ryan@aifirestarter.com

Read past issues