Who gets your AI working after the demo?

A demo can turn an email into a tidy job summary in seconds. Putting that same tool into a working business raises less glamorous questions. Which inbox does it read? What happens when a customer changes the address? Who notices if it stops running on a Friday afternoon?

Those questions are where much of the work sits. A new announcement from Anthropic is a useful reason to look at what you should actually be buying when you pay for AI help.

What Anthropic has announced

On 2 October, Anthropic announced Claude Frontier Academy, backed by a $100 million commitment and a target of training 10,000 Frontier Deployed Engineers by the end of 2027.

The programme combines practical training with a 12-week residency on a real project. Its first cohorts include engineers from large consulting firms and enterprises. Participation is through nomination, so this is not a public training offer for every local business owner.

Our interest is in the approach: people who can work through the messy details of putting AI into use. This article is our assessment of the announcement, not a review of the course or a claim of accreditation.

Start with someone watching the work

For a small business, hands-on support might begin with an hour beside the person who answers enquiries. Ask them to show you a straightforward job, an awkward one and something that fell through the cracks.

You might discover that the biggest delay is finding a previous quote. Or that two people are replying to the same customer. Neither problem necessarily needs a new AI model. An agreed owner, a searchable record or a better connection between existing tools could do more good.

Before approving a build, ask for a short description of the job it will do. “Prepare an enquiry summary and flag missing details for the office to check” gives everyone something concrete to test. “Transform customer communications” does not.

Pay for the awkward cases as well

Imagine a heating firm wants help sorting incoming enquiries. This is an illustrative scenario, not a Kerbside client story.

A useful trial would include a clear service request, a message with no address, a forwarded email containing two customers’ details and a follow-up that changes the original request. An urgent message should still reach a person through an agreed route. The model should not be deciding how long someone can safely wait.

Ask whoever builds it to demonstrate what happens when information is missing, contradictory or unavailable. Look at the actual output together. Agree when it should produce a draft, when it should ask for help and which actions always need approval.

Our guide to choosing between custom AI and a better process covers that decision before you commit to development.

Make the handover part of the job

A working tool needs a named owner. Before the project finishes, you should be able to answer these questions:

• Which accounts and information can it access, and how do we remove that access?
• Where do failed runs appear, and who checks them?
• Can staff pause it and carry on with the old process?
• What are the software and usage costs, separate from support?
• Who updates it when an email provider or connected system changes?
• What do we receive if we move to another supplier?

A short walkthrough with the people using the tool is often more useful than a lengthy document nobody opens. Keep the account details, recovery steps and ownership arrangements somewhere the business can access.

Keep the first piece small enough to judge

You do not need a resident engineer on the payroll to apply this approach. You can commission one contained piece of work, try it with staff and decide whether ongoing support is worthwhile.

For that heating firm, the first release could simply prepare summaries for review. Booking appointments and sending replies can remain with the team. After a trial, compare the time spent checking summaries with the time previously spent preparing them. Include corrections and failed runs in the comparison.

That gives you a sensible basis for the next decision. Expand it, change it or leave it alone. At Kerbside, AI adoption and managed support can cover different stages of that work. The scope should follow the job you need doing, with clear responsibilities after launch.