Guide
The AI agents guide: how AI runs a back office
What an AI agent is, what it should and shouldn't do on its own, the guardrails that make it safe, and how to put your first one to work.
Key takeaways
- An AI agent does more than answer: it does the work, in your systems, against your rules.
- OFFHOURS packages agents into roles — an Accountant, an HR Admin and an Office Manager — each with its own access and rules.
- The rule that matters most is when to stop: new counterparties, amounts over a threshold, anything personal.
- Safe agents have need-to-know access, approvals, a record of every action and people supervising them.
- Start with one role. It is working within a week, on probation at first.
What an AI agent is
A chatbot answers questions. An AI agent gets work done: it reads what came in, decides the next step against a set of rules, and acts in the systems where the work lives — the ledger, the HR system, the inbox, the spreadsheet.
That makes it different from classic automation too. A script follows one fixed path and breaks when an invoice looks different. An agent reads the invoice the way a person would, works out which supplier, which account and which approval it needs, and knows when something doesn't fit.
From agents to roles
An agent on its own is a capability. What a business needs is someone who owns a piece of work. OFFHOURS packages agents into roles — the Accountant, the HR Admin and the Office Manager — each with a job description, the systems it works in, the rules it keeps and the people it reports to.
Thinking in roles rather than tools answers the questions that matter: who does this, with what access, who checks it, and what happens when it's unsure.
What agents do well
The best back-office work for an agent is recurring, rule-based, high in volume and spread across more than one system. It's the work that piles up and lands on someone's desk.
- Booking supplier bills the day they arrive and reconciling the bank to the cent
- Updating invoice statuses from what has actually been paid
- Answering routine leave and policy requests against your own rules
- Preparing joiner and leaver checklists and gathering payroll inputs before cutoff
- Placing orders inside the spending policy and flagging renewals before they fall due
Where an agent should stop
The most important rule an agent keeps is the one that tells it to stop. A good agent doesn't guess: it brings a person in, with the evidence attached, and waits.
- A supplier, employee or counterparty it hasn't seen before
- An amount above the approval threshold you set
- Anything personal or sensitive — salaries, health, disciplinary matters
- A document that doesn't match what the rules expect
- Any decision that needs judgement rather than a rule
The guardrails that make it safe
Need-to-know access. Each role only gets the systems and information its job needs. The HR Admin sees leave and payroll inputs, not the bank.
Approvals where they belong. Payments above your threshold wait for a person. The agent prepares the work; your team signs it off.
A record of everything. Every action carries the rule it followed, who approved it and when, so nothing happens in the dark.
People supervising. OFFHOURS roles are supervised by our engineers, who review the work and tighten the rules as the role learns your business.
How to start
Start with one role, not a platform. Pick the part of the back office with the most repetitive work and the clearest rules — for most companies, that's the books.
- Show us your processes: the recurring work, the systems and the approvals
- We add your rules: the tools are connected and the permissions, boundaries and logic are set
- Within the week it starts work on probation, with every output checked
- Then the role works around the clock, supervised, and reports back every day
How to know it's working
Measure the work, not the technology. Count the hours your team no longer spends on it, how many tasks it clears each day, how many come back as exceptions and how many needed a correction. In our own back office, where we're client zero, the Accountant registers dozens of supplier bills a day and keeps the bank reconciled in full.
Questions to ask any AI agent provider
- Does it work inside our systems, or do we move our data to yours?
- Who can see what — and can we limit access role by role?
- What does it do when it's unsure?
- Where is the record of what it did and why?
- Who supervises it, and who fixes it when it's wrong?
