AI assistants & agents
Building an AI assistant or AI agent starts with the task, not the model: well-scoped, measurable, and with a human in the loop where it matters. That’s how we develop assistants and agents that carry the work - on your own knowledge and data.
Open about our stack
Model-independent: Claude · GPT · open source - on your own infrastructure or in the EU cloud
Your team answers the same questions every day, digs through the same systems for the same information and copies data from one screen to another by hand. Public chatbots don’t help there: they know nothing about your customers, your agreements or your systems.
An assistant or agent built on your own knowledge and data can - but only when the task is properly scoped. That’s why we don’t start with the technology, but with the work itself. It’s the common thread in all our AI work.
- Internal assistants on your own knowledge base, documents and systems
- Agents that carry out well-scoped tasks autonomously, within boundaries you define
- Approval steps for actions with impact: a human signs off before anything happens
- Connections to your existing systems: CRM, ERP, email, planning
- Measurable quality criteria and logging, so you can see what happened and why
What an AI agent does - and doesn’t do - in practice
The difference is autonomy. An AI assistant helps people do their work: it finds the right answer in your knowledge base, summarises a file or drafts a first version - the employee decides. An AI agent goes a step further and carries out a task in steps, autonomously: reading an incoming request, pulling the right data from your systems, preparing a proposal and only sending it after human sign-off.
An honest caveat belongs here: an agent that “handles everything” doesn’t exist. Agents work when the task is well-scoped, the outcome is measurable and there is a clear moment where a human reviews or approves. Vague goals, missing data, or actions where a single mistake causes immediate damage with no checkpoint - we don’t take those on, and we say so up front.
Besides, the biggest gains are often not in the conversation but in the steps around it: reading documents, transferring data, checking results. We cover that ground with AI integration & document processing - that’s how Letselinzicht extracts specialist calculations from loose documents, transparently and reproducibly. And sometimes the honest answer is that you don’t need AI at all, but an internal tool that automates your process.
Our approach: start small, control from day one
week 1
Scope the task
Together we pick one task with a clear start, end and measurable outcome. Not “AI for everything”, but one workflow that matters right now.
week 2-4
Working version on real data
The assistant or agent runs on your real documents and systems. Up front, we define measurably when an outcome is good enough.
before go-live
Build in control
Approval steps, permissions and logging: the agent only does what was agreed, and a human approves sensitive actions.
ongoing
Measure and adjust
We monitor quality and only expand once it works. If it doesn’t work well enough, we scale back or stop - which is possible, because the steps are small.
Start here - AI QuickScan
Know first whether an assistant or agent fits your work.
A focused scan of your processes. You leave with a prioritised, realistic AI roadmap - including an honest answer to where an assistant or agent does and doesn’t make sense right now.
1 session
with senior engineers
5+ use cases
scored on value & effort
1 roadmap
ready to execute
Frequently asked questions
What you want to know before you start.
What’s the difference between a chatbot, an AI assistant and an AI agent?
A chatbot answers questions. An AI assistant helps a person do the work: looking up information, summarising, preparing drafts - the person decides. An AI agent carries out a task in steps, autonomously, towards an agreed result. The more autonomous the system, the more scoping and control matter. That’s why we decide per task which form fits - not the other way around.
Which tasks are a good fit for an AI assistant or agent?
Tasks that recur often, have clear steps and a verifiable outcome: sorting and preparing incoming email or requests, bringing information together from multiple systems, summarising documents, transferring data. Tasks with vague goals, or actions where a single mistake causes immediate damage with no checkpoint, are not a good fit - and we’ll simply tell you so.
How do we stay in control of what an agent does?
By building control in, not hoping for it afterwards. An agent only gets access to what it needs, sensitive steps - sending, changing, paying - require human approval, and every action is logged. So you can always see what happened and why. Where the line sits between “the agent does it” and “a human approves” is up to you, per task.
What does building an AI assistant or agent cost?
That depends on the task, the systems it connects to and the control requirements. That’s why we work scope-first: we start with an AI QuickScan with a fixed, short lead time, and then build in small, budget-controlled steps. After the first conversation you get a concrete estimate - you’re never locked into a long project.
What if it doesn’t work well enough in practice?
Then we scale back to an assistant role with more human control, or we stop. Because we define measurably up front when an outcome is good enough and build in small steps, you find out early - not after months. Nobody benefits from an agent that almost works.
Do we need AI expertise in-house?
No. We take care of the technology: models, integrations, evaluation and maintenance. What we do need from you is knowledge of the work itself - the people who do the task today know where it pinches and when an outcome is right. They help define what the assistant or agent should do and review the first results.
Not sure whether an agent fits your work? Jasper Kums, partner at eenvoud, will tell you honestly where an assistant or agent does and doesn’t work - and where you’d be better off starting. Or start with an AI QuickScan.