# AI Agents That Take the Manual Work Out An agent reads your emails, documents or reports, pulls out what matters and files it into the systems you already use. The AI does the work: you review it, override it, and stay in control. ## Human in the Loop, Always How I build every agent, and the reason they're safe to use. ### The AI does the first pass The reading, extracting and filing that would otherwise eat your day, without getting bored or making typos. ### You review and override Everything lands in a review view where you approve, correct or reject it. A person always has the final say. ### Narrow access, full log Each agent gets only the access its job needs, and every read, change and approval is logged. ## Where This Has Already Worked Two examples, both built by me inside an enterprise compliance platform, both doing real work. ### Control mapping, in a single click Mapping the controls in a SOC2 report against the controls a business actually has took a full day. The agent does it in a single click, with a person reviewing the output and able to override it. ### Due diligence reviews, three hours faster A 150-question questionnaire checked against cyber security frameworks and best practice, returned as a structured report for the compliance team. Each run saves around three hours. ## Jobs an Agent Is Good At Where the input is messy and the answer has to be worked out. If the answer is the same every time, [automation](https://lukereynolds.co.uk/automations/) is cheaper to build and costs nothing to run. ### 💷 Invoices into your accounts Supplier, date, total and VAT pulled from each invoice and posted to your accounts package. Anything that doesn't add up goes to a person. ### 📥 An inbox that sorts itself Enquiries, quotes and complaints read, routed to the right person, with a draft reply ready for you to approve. ### 📝 Quotes drafted The enquiry and your price list go in, a draft quote comes out in your format. You still set the price. ### 🔍 Records checked against the rules A form, questionnaire or report checked line by line against a framework or your own policy, with the exceptions flagged for a person. ## Agent Pricing Priced by how complex the workflow is, not how clever the AI sounds. Fixed prices agreed before I start, with no VAT to add. ## Start with the £750 pilot One job, two to four weeks, a fixed price. You keep the finished agent either way, with human review built in, the process documented, and your team trained to run it. In return I ask for honest feedback and permission to write it up as a case study, anonymised if you prefer. [How the pilot works](https://lukereynolds.co.uk/automations/#pilot). ### Data Extractor From Typical projects £950–£1,500 - Reads structured or semi-structured email - Extracts 3–5 fields you specify - Appends them to a spreadsheet or database - Simple review step before anything is saved ### Operations Agent From Typical projects £2,450–£3,500 - Reads messy, conversational email - Structures the data reliably - Checks for duplicates before it acts - Pushes it into your tools: FreeAgent, Stripe, your CRM - Review dashboard for one-click approval ### Multi-Step Agent From Typical projects £4,500–£7,000 - Keeps multiple systems in sync - Drafts replies and actions for your approval - Handles different cases differently - Reads scanned PDFs and invoices - Slack or WhatsApp alerts when a person is needed ### Running costs - Automation Retainer: monitoring, break-fixes, model and prompt upkeep, two hours of changes a month, no lock-in: £300/mo - Model usage, billed at cost and never marked up: typically £15–£55/mo - Additional work outside the agreed scope: £450/day What keeps your price low: aim the agent at one inbox or label rather than everything, decide the exact fields up front, and keep a person approving the output. Review-and-approve takes roughly half the build time of full autonomy, and it's safer to run. Bigger workflows are quoted individually. ## Frequently Asked Questions ### What does it cost to run an agent? Model usage and hosting are billed at cost, and for a small business handling a few hundred to a couple of thousand items a month that's typically £15–£55 per month. The Automation Retainer is £300/month on top, covering two hours of development or tuning a month plus monitoring, break-fixes and a quarterly review. ### What happens when the AI gets something wrong? That's what the review step is for: anything the agent isn't confident about goes to a person rather than being guessed at. When a connected system changes and something breaks, fixing it is covered by the retainer, so nothing quietly stops working. ### How long does an agent take to build? A data extractor is usually a couple of weeks. More involved agents take 3–6 weeks, depending on how many systems they touch and how much the source data varies. ### Which provider do you use, and does my data train it? I pick the smallest model that reliably does your job, tell you which provider it is, and can change it later. The paid API tiers don't train on your data, while free tiers often do, so I build on paid tiers only. Most model calls are processed in the US; EU processing is available at a premium, and no mainstream provider offers a UK region. If your data can't leave the country, say so at the start and I'll tell you what that rules in and out before you spend anything. ### Can the agent send emails or post things on its own? Only if you ask for it, and I'd talk you out of it for anything customer-facing. The agent does the reading, finding and drafting, and a person presses send. Where you do want something automatic, it's a fixed rule rather than a judgement call, which is automation rather than an agent. ### Why not just use a £20 a month AI bot? You can, if you're happy to set it up, give it your logins, teach it the job and check its work yourself. Those tools sign in as you and often act unsupervised. What I sell is the finished process: built around your systems, with the narrowest access it needs, a person approving every handoff, and someone to fix it when something changes. If a subscription is the better fit, I'll tell you. ### Why not build it fully autonomous? It costs more to build, because autonomy needs far more testing against edge cases, and it's riskier to run. A human review step roughly halves the build time and means you can safely automate processes you'd never hand over unsupervised. ## What's eating your day? Tell me the process you'd love to never do again, and I'll tell you whether an agent can take it off your plate.