Back to home

Service

AI agents

Q: Is it going to replace someone on the team?

That is not how we frame it. An agent absorbs the repetitive part of a role, not the role. And the work that appears afterwards, reviewing what it does and tuning its limits, is done best by whoever already knew the process.

AI systems that do more than answer: they query your data, decide within set limits and carry out tasks with a log of everything they do. And before that, the consulting that decides whether you need one at all.

Tell us your case See what we implement

We work with

WordPress
Shopify
WooCommerce
n8n
Airtable
Supabase
Hermes Agent
Artificial Intelligence

What an agent is

The word gets used for everything. Here is what it means here.

An automation walks the steps someone drew: if this comes in, do that. It is predictable and cheap, and for most repetitive work it is exactly what you need.

An agent works the other way round. You give it a goal, a set of tools it may use and limits it may not cross, and it picks the route in each case. That helps when the task changes depending on what it finds, and it is money wasted when the task is always the same.

The useful question is not whether you want an agent, but whether the work varies enough to need one. That is why nearly every project starts with the consulting.

What we implement

Six starting points. The first one often saves you the other five.

Process consulting

Before implementing anything, look.

We walk through how the team works, where work stalls and what data actually exists, rather than what should exist. Out of that comes an ordered list: what to automate, what deserves an agent and what has to be fixed by hand first.

It is the service that most often ends up recommending you do not buy AI, because a good share of what people want an agent for is fixed by tidying up a database.

Who it is for: Teams that do not know where to start
Works on: Your current processes, data and tools

Sales and support agent

It goes further than answering. It checks the catalogue, confirms availability, prepares a quote using your discount rules and leaves it ready for a person to sign off.

When conversation is the main channel, the entry point is the AI chatbot. The agent goes a step further, because it also acts.

Who it is for: High-volume sales and support
Works on: CRM · catalogue · stock · price lists

Document agent

Invoices, delivery notes, contracts, policies and case files. The agent reads them, pulls out the fields that matter, checks them against what you already hold and flags anything that does not line up for a person to review. What it really adds is knowing when it is unsure and saying so, rather than filling a field blind.

Who it is for: Back offices with a lot of paper coming in
Works on: Email · folders · ERP · document manager

Analysis agent

It answers questions about your own data in plain language. "Which customers ordered less this quarter and what did they have in common?" without anyone writing the query by hand.

It always returns the figure with its source and the query it ran, so anyone can reproduce it.

Who it is for: Management waiting on others to see a number
Works on: Database · ERP · CRM · spreadsheets

Internal knowledge agent

One place where the team asks and gets the answer from your own documentation: procedures, per-customer terms, incident history, the manual nobody can find.

It respects each person's permissions: nobody sees through the agent what they could not see by opening the folder.

Who it is for: Companies where the knowledge lives in three heads
Works on: Drive · Notion · SharePoint · email · tickets

Custom implementation

When the process is yours and resembles nothing else. The agent is designed with its tools, its limits and its approval points, and put to work first on a small slice of real work. We never start with the critical process, but with one where a mistake is fixed by a phone call.

Who it is for: In-house processes with rules that shift
Works on: Whatever it takes, with human oversight

Putting it on a leash

What an agent is handed before it is allowed to touch anything.

A goal "Leave every incoming invoice matched against its order." Specific and checkable. what it has to achieve
A set of tools Read the mailbox, query the ERP, write to a sheet. Nothing else. Anything off the list does not exist for it. what it may use
Limits A spending cap, a volume cap, and the list of actions that need a person to say yes. how far it goes
A log Every lookup, every decision and every action, timestamped. It is what makes correction possible. what gets written down

If the agent reaches past its tool list, the call is refused and the attempt is logged. The model does not get to set its own permissions.

Agent or automation

Most of the time the right answer is the automation. It is worth knowing when it is not.

Criterion	Automation	AI agent
The task	Always the same steps	Changes with what turns up
Result	Predictable, every time	Right almost always, not always
Running cost	Very low	Per model usage
Who decides	The rules you wrote	The agent, inside your limits
Human review	Optional	Required wherever there are consequences
Fits when	The process is settled	Something has to be interpreted, compared or written

The usual answer is both: the agent decides and an automation executes.

How the process works

Five phases. It always starts on a small slice of real work.

Diagnosis

Current processes, data and tools. The output is a list of candidates with what each one costs and saves.

Picking the case

We pick one where a failure does no damage, with success criteria written down before we start.

Sign-off

Build

Tools it may use, limits it may not cross, human approval points and a log of everything.

Supervised pilot

It runs on real work with a person reviewing every output, measured against the criteria from phase two.

Sign-off

Rollout

If the pilot holds up, volume grows and supervision eases off gradually. If it does not, we stop and say so.

What gets written down

Half an hour of a wholesaler's invoice agent, as its log reads.

09:14:02readInvoice F-2291 received from Arce Supplies
09:14:05extractNet 1,840.00 € · VAT 21% · due 30 Nov
09:14:07queryOrder PD-8812 in the ERP
09:14:08mismatchOrder says 1,640.00 €. Difference of 200.00 €
09:14:08stopOutside tolerance. Not posted
09:14:09alertCase opened for accounts, both figures attached
09:41:30resumeMaria approves: there was unbudgeted freight
09:41:31writePosted against PD-8812

The two orange lines are the work you were actually buying. An agent that had posted that invoice without blinking would have been faster and far worse.

Control, data and limits

The part that decides whether this can go into a company at all.

A closed list of actions. The agent can only use the tools you give it. Anything not on the list does not exist for it.
Human approval where there are consequences. Emailing a customer, moving money or changing a record goes past a person, unless you ask otherwise.
A full log. What it looked up, what it decided and what it did, with timestamps. Without a log there is no way to correct anything.
Your data does not train models. Business accounts with training switched off, and EU-hosted or self-hosted models where the data demands it.
Inherited permissions. The agent sees what the person asking sees, not one document more.
Spending and volume caps. They are set before it is switched on, because the model is paid per use and nobody wants to find out the bill at month end.

Frequently asked questions

What is the difference between an AI agent and an automation?

An automation follows the steps you drew, always the same ones. An agent picks the route: you give it a goal, a set of tools and some limits, and it decides what to do in each case. That helps when the task varies; when it does not, a plain automation is cheaper and more predictable.

Where do we start?

With the consulting, nearly always. A working session on your processes usually makes clear that half of what you wanted AI to solve is fixed by tidying data or building an ordinary workflow. What is left after that does justify an agent.

Does our data end up training a model?

Not with the setup we build. We use business accounts with training switched off, and where the data is sensitive we work with models hosted in the EU or on your own server. That is decided at the start of the project, not at the end.

How much control is there over what the agent does?

As much as you decide. Every agent gets a closed list of allowed actions, spending and volume limits, and a log of everything it has done. Actions with consequences, such as emailing a customer or moving money, stay behind human approval unless you ask otherwise.

Is it going to replace someone on the team?

An agent absorbs the repetitive part of a role, not the role. And the work that appears afterwards, reviewing what it does and tuning its limits, is done best by whoever already knew the process.

What if the agent gets it wrong?

We assume it will. That is why there are limits, human approval where it matters and a full log. In a well-built agent a failure shows up immediately and could not do any damage.

Let us start by looking

Tell us which process worries you. If you do not need an agent, we will say so, which is usually the answer.

Talk to us

We also build automations, AI chatbots and websites.