Last year an AI tool could write you a reply. This year an agent can answer the inbox.
Draft the replies to every enquiry. Chase the late updates. Build the Monday report from the numbers. Those used to be someone's job. Now a computer worker can carry them. That is a whole responsibility, not one task.
It is easy to read that as "the person is replaced". That is not what I have seen in the businesses I have worked with.
What I have seen looks more like a car.
The car does the miles. You do the driving.
A car does the hard physical work. It turns fuel into motion and gets you across the country before lunch.
But you still sit at the wheel. You choose where to go. You watch the road. You brake when a child steps out. Leave the keys in it and walk away, and it goes nowhere.
Today's AI agents are cars with a person at the wheel. The agent does the work: reading, drafting, sorting, calculating, moving information between systems. The person decides where it goes and checks the driving.
The responsibility has not disappeared. It has split in two. The doing moves to the agent. The steering stays with a person.
Five things decide how well it runs
Two businesses can buy the same AI product and get different results. The difference is not the product. It is what each business gives the agent to work with.
There are five. I call them the five Cs.
Microsoft names three of them when it describes its own agents. An agent "follows the instructions you give it, draws on the knowledge sources you connect, and uses tools to take action" (Microsoft Copilot Studio documentation). Instructions, knowledge, tools. Add the model underneath and the skills you teach it, and you have all five.
Capability is the engine. The model underneath the agent sets the ceiling. A stronger model reads a messy email thread more accurately, follows a longer brief, and notices when something does not add up. You do not always need the biggest engine. You do need to match the engine to the journey. Sorting enquiries into three piles is a short trip. Drafting a quote from a rambling voice note is a long one.
Context is the map. An agent that knows nothing about your business can only give generic answers. Give it your price list, your service procedures, your last twenty replies to the same question, and it starts to sound like you. This is also where access matters most. It should only see what the person it works for could see. I wrote about that in AI answers from your documents, with the right access.
Clarity is the destination. Most people skip the brief. "Reply to customers" is not a brief. "Draft a reply to any enquiry about repeat orders. Confirm the quantity and the delivery date. Never quote a price. Leave the draft for Sam to send." That is a brief. Write it down. Change it when the work changes. The agent follows the brief you wrote, not the one in your head.
Craft is the driving lesson. Every business has jobs it does a particular way. The way you lay out a quote. The order you check a new supplier. The tone you use with a customer who is annoyed. Anthropic calls this a skill. It is a folder of instructions, code and reference material, "organized like an onboarding guide you'd create for a new team member". The agent loads it when that job comes up (Anthropic's Agent Skills documentation). That means you teach the job once, in writing, and the agent uses the same method every time. A good skill is the difference between a new starter and someone who has done the job for a year.
Connectors are the roads. An agent that can only talk to you in a chat window is a car in a garage. Connectors let it read your inbox, look up an order, check the calendar, put a row in the spreadsheet. There is an open standard for this, the Model Context Protocol. It describes apps like Claude or ChatGPT connecting to data sources, tools and workflows (modelcontextprotocol.io). OpenAI's API connectors list Gmail, Google Calendar, Outlook and SharePoint among the services an agent can reach (OpenAI developer documentation). The more of your systems it can reach, with the right permissions, the more of the responsibility it can carry on its own.
Improve one of the five and the car runs better. Improve all five and you can hand it a bigger job.
Course correction is the job that stays
Even a good driver on a good road checks the mirrors.
Models make mistakes. The vendors say so themselves. Microsoft's guidance for its own Copilot says users "should still use their judgment when reviewing the output before sending them to others" (Microsoft 365 Copilot privacy page). OpenAI's developer guidance recommends "having a human review outputs before they are used in practice" (OpenAI safety best practices). So decide, in advance, where the person looks.
Three review points cover most jobs.
Before anything leaves the building. A draft reply, a quote, a report: a person reads it first. Keep it until the agent has earned your trust on that job.
When the agent is unsure. Tell it to say so. "I could not find the delivery date" is a useful sentence. A confident guess is not.
Every week, on the pattern. Look at ten pieces of work together. If the same correction comes up three times, the fix is not another correction. It is a line in the brief or a change to the skill. That is the moment the agent gets better for good.
The UK government's AI regulation principles ask for the same habit: "effective oversight of the supply and use of AI systems, with clear lines of accountability established across the AI life cycle" (AI regulation white paper, 2023). They are not law. They describe what a well-run business does anyway. Someone is named, and that person looks.
Where responsibilities move first
The jobs that move first share a shape. They repeat. The inputs already live in your systems. A person can check the output in a minute.
Take enquiries. Say a hundred a week arrive by email, and forty of them ask about the same three things. An agent with your order records and a clear brief can draft all forty. Someone reads them over coffee and sends. That person's responsibility used to be "answer the inbox". Now it is "check the inbox is answered well".
Take reporting. Say the numbers live in four places. Someone spends Monday morning copying them into one sheet. An agent with connectors to those four places builds the sheet overnight and flags the number that moved most. Monday morning becomes a conversation about that number.
Take knowledge. Say new starters ask the same questions for six months. An agent with your procedures and the right permissions answers them, with a link to the source. The person who used to answer becomes the person who keeps the source current.
In each case a person is still at the wheel. They just do less of the work.
Start with one road
You do not need to hand over the whole business. You need one responsibility that repeats, one agent with the five things it needs, and one person who checks the work and adjusts the brief.
Pick the job. Write the brief. Give the agent the map and the roads. Then drive it for a month and see where you keep correcting.
That is where the next skill goes. You already know which job to start with. Pick it.
Vendor documentation checked on 8 September 2026. Features and plan requirements change; check the current page before you rely on one.