Document agent
Reads incoming invoices and certificates, extracts issuer, amounts and expiry dates, names the file and files it in the right place.
AI agents · Automation · n8n
AI agents for business
An agent isn't a chat window. It's a digital worker that takes a job from start to finish: reads, decides, acts, and comes back to you when something is unclear.
✓ Ten agents running in our own product every day
What an AI agent really is
The word gets used for everything. Here's the useful definition.
A language model answers a question. An AI agent is given a goal, has tools available — database, email, file storage, APIs — and works through several steps towards that goal without someone triggering each one.
The practical difference: you ask a model "how should I write this product description?" and you get a paragraph. You give an agent your catalogue and it writes 20,000 descriptions, checks them against your rules, files them in the store and flags the twenty where the source data was wrong.
What matters isn't how clever the agent is, it's where you put the approval. Anything that leaves the building or moves money goes through a human. That isn't distrust, it's operational safety.
Agents we've built
Concrete examples from live projects.
Portal agent
Uploads paperwork to compliance portals, checks whether it was accepted and requests renewals before anything lapses.
Quoting agent
Turns site measurements or bills of materials into a finished quote with your prices and your layout, ready for review.
Catalogue agent
Writes and optimises product content at volume, in your tone of voice and with your mandatory fields.
Inbox agent
Sorts messages, spots urgency, prepares draft replies and escalates what a human needs to see.
Reporting agent
Pulls numbers from several systems, checks them for inconsistencies and writes the report in plain language.
What goes wrong when it's done badly
We've built enough agents to know the usual failures.
Too much autonomy too early. An agent that emails customers without approval is a reputation risk. We always start in draft mode and only loosen it once the hit rate holds up over weeks.
No audit trail. If you can't see why the agent decided something, the first mistake is also the last — nobody will turn it on again. Every decision gets logged.
One agent for everything. Several small agents with a clear job work better and can be fixed one at a time.
How an agent is put together
No jargon, but precise enough for you to ask the right questions.
An agent has four parts. The brief describes in plain text what it should do, what it must never do and when to stop. The tools are what it's allowed to use: a database query, a file operation, an API call, an email draft. The memory keeps what it has already done in this job and what it needs to carry over from previous ones. The controls define where a human has to approve.
The most common mistake is in the tools: giving the agent full access because it's easier. We do the opposite — it gets exactly the permissions its job requires and nothing more. An agent that reads invoices needs no write access to the customer directory.
The same applies to memory and data: anything that doesn't need storing is discarded after the job. That isn't only about privacy — it keeps the agent reliable, because it isn't working from stale context.
From prototype to production
An agent that works in a demo is about halfway there.
Stage 1 — draft mode. The agent does the whole job, but everything it produces waits for approval. This runs for at least two weeks on real work. We count how often a human had to correct it, and what they corrected.
Stage 2 — partial autonomy. The case types the agent has been reliable on for weeks run without approval. Everything else stays in draft. Where that line sits is a business decision, not a technical one — it's yours, not ours.
Stage 3 — production with sampling. The agent runs largely on its own, but a share of jobs still gets checked, so that a slow drift in quality shows up before it gets expensive.
We've never let an agent run fully autonomous on day one and we wouldn't start now. The two weeks in draft mode aren't caution for its own sake: that's when you find the edge cases that appear in no process document.
Frequently asked questions
Does an agent replace a role?
No, it replaces tasks. In our projects no role has disappeared; what disappeared were the hours spent copying, renaming and uploading.
What if the agent gets it wrong?
That's what the approvals and the audit trail are for. Anything sensitive passes a human first, and every decision is traceable. When it's unsure, the agent escalates instead of guessing.
Which model do you use?
It depends on the task and your data requirements. We tell you upfront which provider sees which data, and there are tasks that can run entirely on your own infrastructure.
Which job would you hand over?
Tell us which recurring task eats the most of your team's time. We'll tell you whether an agent can take it.
Book a free call