Automation
Which tasks should you hand to AI agents, and where does a person still sign off?
Hand AI agents the repeatable, rules-based work that is quick to check: data entry, chasing and follow-ups, first-draft quotes, triage and routing, and routine reporting.
Keep a named person signing off anything that moves money, sets a price, goes out to a client, carries legal weight or decides something about a person, such as hiring. If a mistake would be hard to undo, or would land on someone outside the business, a person approves it before it happens.
What to hand over, and what to keep.
Most processes mix both kinds of work. The agent prepares; a person makes the call.
Hand to AI agents
Runs on its own, checked by sampling
- Data entry Details from emails, forms and PDFs keyed into your systems.
- Chasing and follow-ups Reminders for overdue paperwork, invoices and replies, in wording you've approved.
- First-draft quotes Built from your price list and past jobs, ready for a person to check.
- Triage and routing What comes in is read, tagged and sent to the right person or queue.
- Reporting Figures pulled together into the weekly or monthly pack.
Keep a human sign-off
The agent prepares; a named person decides
- Payments Releasing, refunding or changing payment details, above the limits you set.
- Pricing Final prices, discounts and terms.
- Anything client-facing Quotes, complaint replies and bad news, checked before they're sent.
- Legal and contractual Contracts, filings and anything signed in your name.
- Hiring and people decisions Who is shortlisted, hired, rated or let go.
Where the two overlap, set a limit. A routine chaser can go out on its own within approved wording; anything outside the script, or any reply to an unhappy client, comes to a person first.
Three questions that settle most cases.
- Can it be undone? A draft can be deleted. A payment, a sent email or a deleted record can't. Hard-to-reverse actions need approval.
- Who finds out if it's wrong? If it's a colleague at the next step, let the agent run. If it's a client, a regulator or a candidate, a person signs off first.
- Is checking quicker than doing? If someone can check the output in a fraction of the time it takes to produce, hand it over. If not, the agent isn't saving anything yet.
Keep approvals for the moments that matter. Ask someone to approve 99 routine actions and they will miss the one that counts.
How we build sign-off into bespoke builds.
Off-the-shelf tools come with someone else's idea of control. When we build an agent for you, the sign-off points follow your approval rules, sit inside your own systems, and you own all of it.
- Approval steps. The agent does the preparation, then pauses at the points you choose. The right person sees what it wants to do and why, and approves, edits or rejects it in one step. You set the thresholds: refunds under an agreed amount go through, for example, and anything above goes to finance.
- Role-based permissions. Each agent gets its own identity and only the access its job needs. It can read before it can change anything, and who approves what follows the roles you already have.
- An audit log of every action. Everything the agent reads, drafts, changes or sends is recorded, with who approved it and when. If a question comes up in six months, the answer is in the log, in your own systems.
We start with tight limits and widen them only once the agent has proved itself on your own work. See how we approach bespoke AI development.
Find out where the line sits in your business.
The AI Opportunity Audit maps your processes as they run, scores where AI pays off, and marks the points where a person should stay in charge. Fixed price, and credited against your build.
FAQs
Frequently asked questions.
What tasks can AI agents do for a business?
Agents handle repeatable work that follows clear rules or patterns: keying data from emails and documents into your systems, chasing overdue paperwork and payments, drafting quotes and replies, sorting and routing enquiries, and pulling figures into reports. They work best on high-volume tasks where the output is quick for a person to check.
Which decisions should AI not make on its own?
Anything that moves money, sets a price, goes out to a client, carries legal or contractual weight, or decides something about a person, such as hiring or performance. An agent can gather the information and draft a recommendation, but a named person should approve it before it takes effect.
What does human in the loop mean?
It means a person reviews and approves an AI agent's work at set points before it takes effect. In a well-designed system those points are kept for actions that are hard to reverse or that affect someone outside the business, so people aren't asked to approve routine work.
Will my team have to approve everything the agent does?
No. Routine, reversible work runs on its own and is checked by sampling, with every action recorded in an audit log. Approval is kept for the decisions that matter, so the people approving have the time and context to make a real judgement rather than clicking through.
Can AI make decisions about people in the UK without a human involved?
Only with safeguards. Since 5 February 2026, UK GDPR, as amended by the Data (Use and Access) Act 2025, allows significant decisions based solely on automated processing only if the person affected is told, can make representations, can ask for human intervention and can contest the decision. Tighter limits apply to special category data such as health. For hiring and similar decisions, most firms keep a person making the final call. Take legal advice on your own case.