The step I always keep human

Thanks for coming back for issue three and Happy Friyay. This fortnight has been a mix of one-on-one coaching, a discovery process with a logistics team, and working sessions with a client whose accreditation process we’re automating end to end. A couple of things kept showing up in more than one room.

The first is a decision I keep making the same way, in completely different projects. Automating the accreditation process, we got to the step where a reviewer gets assigned to a case. My instinct was to automate it: match workload, assign, done. But workload isn’t really data, it’s a judgment call about who’s stretched thin right now. We left that one step human and automated everything either side of it. Almost the same conversation came up with a logistics team I’ve been interviewing about how they plan deliveries. Their planners have twenty years of knowing a vehicle’s capacity by sight and which driver can handle which route. On one call, a planner heard tiredness in a driver’s voice and pulled them off the road there and then. No system flags that. The pattern across both: figure out which single step in a process is actually a judgment call, keep a human on that one, and automate the rest properly rather than half-heartedly. (Forrester’s research on human-in-the-loop AI points to the same shift: automation can accelerate the work, but it shouldn’t own the judgment.)

The second is a client I coach one-on-one who built a personal tool in Claude Code for their own use. Two people who saw it asked if they could buy it. They asked Claude to turn it into a proper product, with login and payment, and it was live overnight. From a tool built for one person to something two strangers wanted to pay for, in the time it takes to sleep.

Things to try

• If you’re automating a process, name the one step that’s a judgment call before you touch anything else. Automate everything around it and leave that one alone.

• Built something for yourself in Claude Code that solves a real problem? Ask Claude directly whether it could become a standalone product, and what it would take.

• Before a high-stakes conversation, brief Claude on the material and ask it to predict the objections you’ll get. Rehearse against those, not a generic run-through.

• If you’re burning through your monthly token budget on routine, repetitive tasks, ask about running an open-source model locally instead. It’s a weekend setup, not a rebuild.

What’s new in AI (worth knowing)

Claude Opus 5 is out, and it’s not just for the hardest problems anymore

Anthropic released Claude Opus 5 on 24 July: the current flagship Opus-tier model, priced the same as its predecessor (Opus 4.8) but noticeably stronger on agentic and coding work, with a 1-million-token context window. Practically: the model you’d usually reserve for your gnarliest problem now handles more of your everyday work without a price jump. If you’re on a paid Claude plan, it’s already there. Open the model picker and select Opus 5.

OpenAI’s “Presence” is handling real customer calls, not just chat

OpenAI has quietly turned its voice-and-chat agent technology into a deployed enterprise product called Presence, pairing the model with policies, guardrails and escalation rules. OpenAI is using it for its own phone support and says it resolves 75% of inbound issues without a human. Banking and telco names are already live on it. Worth knowing because it’s a live number, not a promise: if a board conversation this quarter touches customer service automation, “resolves 75% without escalation” is the benchmark to ask your own vendor about.

Boards are starting to ask where the AI budget actually went

A cost governance report released this fortnight found AI has eroded gross margins at four in five enterprises for a second year running, with most unable to fully account for what they’re spending. Separately, Gartner found only about a quarter of companies say their governance frameworks are keeping pace with AI adoption. Two different sources, same signal: AI spend is moving from “innovation budget” to something the board wants a line item and an owner for. If nobody in your organisation currently owns that number, that’s the gap to close before it becomes a bigger conversation.

If anything from the above resonates with where you are, reply to this email.

Gareth

P.S. The insights in this issue came from workshops and one-on-one coaching sessions I run on AI. If any of it sounds useful for your team, or you’d like to work through your own AI adoption, reply to this email and we’ll work out what makes sense.