An agent that triages your inbox and one that turns rough notes into a finished document are both worth an afternoon. Both need a decision from you before they touch company mail.
Microsoft's research on the infinite workday found people interrupted every two minutes during core hours, 275 times a day. The average inbox takes 117 emails, most of them skimmed in under a minute, and six in ten meetings never reach a calendar in advance.
An agent wired into your mailbox and calendar can have the overnight pile sorted before you open it. Another turns a folder of rough notes into a document you can send. Neither takes more than an afternoon or a line of code, and both have to reach into company mail to be useful, which means reading personal data about your clients and your staff.
Triage that survives contact with your inbox
Connect an agent to your mailbox and calendar, then give it categories that match how your business runs. Client, supplier, internal and noise covers most people. It reads what came in overnight, sorts it, drafts replies for anything a client is waiting on, then checks the pile against your diary so tomorrow's conflicts surface before tomorrow does.
A prompt that holds up fits in a sentence. Review my unread mail from the last 24 hours, sort it into client, supplier, internal and noise, draft a reply for anything a client is waiting on, and send nothing without my approval. That sentence names the job, the data it may touch, the output you want back, and the limit it can't cross.
On the calendar side, ask it for what needs preparing before a meeting and where two commitments collide. That beats a tidy summary of your day, because prep slips first when the morning gets away from you.
Keep that limit tight while you learn what the output is worth. Let it draft rather than send, and read the calendar rather than move it. After a week, count how often you still open every message to check the triage yourself. If the answer is most of them, your categories need the work rather than the software.
From rough notes to something you can send
The second agent produces a real file. Point it at a folder of notes, say who's reading and how long you've got, and a PowerPoint or Word document comes back that you can open and edit. A board paper or a deck for the finance director are the obvious first jobs.
Then turn your own brand template into a reusable instruction set, and every deck from then on arrives on brand without anyone reformatting it.
These agents also fill gaps. Hand over thin notes and you'll get confident sentences nobody sourced. Tell it to ask you questions where your notes run out, and read the draft for claims you never supplied before a client or a lender sees it. Vague notes produce vague documents, and no amount of prompting repairs that.
The data protection side
The ICO is direct about where this lands. In the majority of cases, using AI involves processing likely to result in a high risk to people's rights and freedoms, and that triggers the legal requirement to carry out a data protection impact assessment. The same guidance expects you to be clear, open and honest about how and why you use personal data in an AI system.
An agent that drafts a reply for you to read and send keeps you as the author. An agent that replies on its own puts software into correspondence with your client, and the transparency duty above applies to whatever it sends.
Ask where the drafts and any cached copies of your mail are held, and for how long. That answer belongs in the DPIA, and it differs between vendors, so ask before you connect anything.
Your team has probably started without waiting for a policy. Microsoft found 78% of people who use AI at work bring their own tools, and at small and medium companies it's 80%. More than half won't admit to using AI on their most important tasks, and only 39% have had any training from their employer. That's shadow IT, and it's running in your business now. Ask people what they've already connected before you write any policy, because the policy has to describe what's happening.
Setting boundaries and permissions
Put the approval boundary in the prompt itself, then widen it slowly. Make the work visible first, then make it faster, then let it run to a schedule, and only hand over judgement once you've watched it behave for a few weeks. Treat it like a new starter, because nobody hands a new hire the client inbox on their first morning. Schedule an unattended run across live mail on day one and the first person to catch a mistake will be a client.
Connect agents through named company accounts whose permissions someone has reviewed, rather than personal logins nobody can audit. Training matters as well, because that 39% is why so much of this happens in the dark.
A connector somebody set up on a personal account keeps its access when they leave, and nobody revokes what nobody approved. Add AI connectors to the leaver checklist you already run, next to the mailbox and the laptop. If you're weighing up which tools to sanction, we've compared the main AI plans for UK businesses and looked at the agents most UK SMEs are missing in Microsoft Copilot.
A first week that costs nothing
Give it one person and one agent for a week. Choose the mailbox carefully. Somebody in operations with high volume and low sensitivity will teach you more in a week than the managing director's inbox, and it costs far less if the categories come out wrong. Drafts only, with nothing sent or scheduled unless somebody reads it first. Write down what data the agent will touch before it touches any of it, and run the DPIA if that includes personal data, which for a mailbox it always does.
We do this as part of AI consulting, which covers the guardrails and the training that stops people improvising around you. Often the bigger win sits upstream, in the process feeding the inbox: that's business process automation work, and it usually beats putting an agent on the symptom. We're ISO/IEC 27001:2022 certified and Cyber Essentials Plus, so the governance side is a conversation we have most weeks. Our wider view on where AI genuinely helps a smaller business covers the rest of the picture.
If you'd rather talk it through before anyone connects anything, book a discovery call and we'll look at your setup with you.
