Your AI Can Talk. Can It Get the Job Done?

Imagine hiring a bright assistant and asking them to handle customer bookings. They can write clearly and understand what people want. But they still need access to the calendar, the booking rules, a record of earlier conversations and someone to ask about unusual requests. Even a capable person would struggle without that setup. AI needs something similar. Its ability to understand a message is only one part of doing the job.

The AI model is the part that interprets information and generates a response or proposes a next action. The harness is the surrounding software that puts those suggestions to work: it can fetch a record, run an allowed action, bring the result back to the model and repeat until the task is finished or needs help. Cloudflare’s documentation describes this ongoing process as the core of a harness. You may also hear ‘AI agent’: that is the working system, using a model and a harness to pursue a task. The harness is usually built into the product you use, so you may never see the name on a button.

The useful question is what the whole setup can complete in your business.

Here is an illustrative example, not a report of a customer deployment. Suppose a service business allows customers to move appointments to an available slot at the same price. With the right connections and rules, an AI system could handle the request like this. If Friday has several possible times, it would first ask the customer which one they want.

• Find: Locate the correct customer and existing appointment. If the match is unclear, ask for help.

• Check: Read the live calendar and the business’s rules for moving bookings.

• Act: Change the booking once the time is agreed and the action is allowed. A change outside the rules goes to a person.

• Confirm: Check that the booking was saved, then send the confirmation and record what happened.

A polished reply is useful. A correctly updated booking is the result the customer needs.

You do not need to understand the code to judge whether the setup makes sense. Ask how it gets the facts, what it can do, what it remembers and how it knows the job is done. Anthropic’s work on longer tasks shows why recorded progress and checks matter: an agent can lose track or declare success too early without suitable support.

• The right information: Current records and clear rules, with a way to spot missing or conflicting details.

• The right tools: Connections to the systems where work happens, with access limited to the job.

• A record of progress: A saved account of what has already happened, so an interruption does not mean starting blindly again.

• A check on the result: Evidence that the change succeeded, plus a clear point at which to stop or ask a person.

This is a revealing demo question. If there is no suitable slot, the system should offer valid alternatives or pass the request to someone who can help. If the calendar goes offline, it should report the problem and leave a clear record of unfinished work. If an update gives no clear result, it should check what happened before trying again. Otherwise it could create a duplicate. These behaviours have to be designed and tested; the word ‘harness’ does not guarantee them.

Writing ‘always ask before making an exception’ in an instruction is not the same as enforcing that rule in the software.

AI can misunderstand a request. Your records can be wrong. A check can miss a mistake. A harness helps you manage the work, but it does not make the system infallible or decide your business policy for you. Someone still has to define which changes are allowed, who handles exceptions and what counts as completion. Give that person a way to review actions and pause the system. Start with a task where errors are easy to spot and correct.

Many AI products already include some of this setup. Check what the product actually supports and whether it connects to your systems. A custom solution may make sense when your workflow has needs that existing tools cannot meet. For a task that always follows fixed steps, ordinary automation may be enough. Anthropic’s guidance on building agents recommends starting with the simplest suitable solution. The aim is to remove work from your team’s day, and you can assess that without buying into every new term.

Choose one everyday request and follow it all the way through. Watch where the information comes from, which records change and what happens when something is missing. Then try a small pilot and compare completed jobs, staff review time, mistakes and total cost with the current process. Include awkward cases such as an unavailable slot or a repeated request. That gives you something concrete to judge before handing over more work.

Ask: ‘Can you show me the finished result, and what happens when the job cannot be completed?’