Many companies already have “an agent.” Often, it is a chat window beside the actual workflow. Someone pastes text in, reviews the response, and manually moves the result into another system.
The expensive part rarely ends with writing. It is finding information, checking it, transferring it, and assembling a usable result.
Documents, email, SharePoint, shared drives, business applications, and individual employees all hold pieces of the answer. People keep looking up the same information, often in competing versions with no clear indication of which one is current. Data moves between systems by copy and paste. Project knowledge leaves when a colleague does.
You already own the knowledge. You keep paying for the work of finding it.
What a custom AI agent takes on
A custom agent works inside a process. Three types of work come up repeatedly:
Finding knowledge with sources. The agent searches approved documents, compares versions, and provides an answer with its source and date. The answer should be traceable to the material you have made available.
Turning raw material into a draft. Notes, voice messages, photos, and unstructured input become a report draft that the team can complete and approve. People retain the final decision; the agent handles collection and organization.
Carrying out recurring workflows. Checking systems, transferring data, and preparing emails or invoices can form part of a defined task. The agent requests approval at the points you have specified.
These are possible applications. Scope, access, and operating boundaries are agreed for each project.
Custom means you define the boundaries
“Custom” has a practical meaning. You decide:
- Which documents and systems the agent can access.
- Which task it can finish independently.
- Where it must stop and ask.
- What makes the result ready for someone else to use.
The assignment is specific: carry an agreed piece of work through to a verifiable result. Permissions and approval points turn that assignment into an operating process.
Boundaries make responsibility clear. Sources let people check the result. Approval points keep consequential decisions under control.
Your archive does not have to be spotless
“We need to get our data ready first” is a common concern.
A cleaned-up data environment is not a prerequisite. Old versions, duplicates, and expired material belong in the assessment, not in the agent's answer.
We start by reviewing which existing documents are useful for a narrow task. Access is limited to the selected material. Conflicting information should be flagged so that the team can resolve it.
Working through real cases reveals which gaps actually interfere with the job. That gives data cleanup a concrete purpose and a manageable scope.
Some workflows need a different tool
We start with the operational problem. Sometimes an agent fits. Sometimes a small custom application or a standard tool you already pay for is the better choice. We say so.
An agent is worth considering when a task involves several steps, depends on sources, and produces something another person or system will use. A chat window can be enough for occasional writing or summaries. An existing business application may already handle a clearly defined workflow.
The useful question is: which part of the work consumes time because someone has to search, transfer, or assemble information, and can that task be defined well enough for an agent to take it on?
What this looks like in our work
The Sentinel web portal has handled video data for years, from capture through processing to delivery. The customer also works directly with our agent, Satoru. Satoru checks running systems, resolves issues within agreed boundaries, and asks for approval where required.
We use the same agent internally for development, testing, and recurring office work such as incoming invoices. We operate the kind of software we propose to customers. Everyday use is the test that matters.
Our own products illustrate a related focus on usable output. Zedl turns photographed receipts into structured accounting data for DATEV, BMD, CSV, or Excel. Nodl turns spoken notes into structured documents. Both are our products, and their value comes from what people can do with the result.
Measure the work, including corrections
We assess usefulness through three measures:
- Processing time, including rework.
- Accuracy of results.
- Day-to-day effort.
Prompt counts and a polished tone do not tell you whether the workflow improved.
A draft that needs a complete rewrite has saved little work. An answer without a source makes the team repeat the investigation. A data transfer that needs manual correction may simply move effort elsewhere.
If real cases do not show an improvement, the task may have been scoped poorly, or an agent may be the wrong tool. Either finding is useful before committing to a larger project.
Start with a small, usable piece
The first project covers a limited part of the work so you can judge it in practice.
- A free, nonbinding 30-minute conversation about your situation and whether we are a good fit.
- A workshop covering workflows, data, access, boundaries, and goals. Scope and a fixed price follow.
- A proof of concept (PoC) using real cases. The target is an initial deployment after about a month, with scope and schedule agreed together.
- Further development of what proves useful, with operational support to the agreed extent.
The indicative PoC range is EUR 2,500 to EUR 15,000, with a fixed price after the workshop. Further work is arranged separately as a block of hours or a fixed number of hours per month.
We assess integrations with Microsoft 365, SharePoint, email, shared drives, databases, and business applications individually. The aim is to work with the IT you already operate.
Frequently asked questions
What is a custom AI agent? It carries out a defined task across several steps using approved sources and systems. The result is intended for use in the next part of the workflow.
Which back-office tasks are suitable? Examples include searching project records, drafting reports from notes, and transferring information between systems. Sources, expected output, and approval points need to be specific.
Do we need to train our own language model? No. Customization concerns the assignment, sources, system access, and approval points; it does not require a company-specific language model.
Can ex-nihilo connect an agent to our existing systems? Microsoft 365, SharePoint, email, shared drives, databases, and business applications are possible integration targets. Each connection is assessed individually, rather than assumed to have a ready-made connector.
How much does a proof of concept cost, and how long does it take? The indicative range is EUR 2,500 to EUR 15,000, with a fixed price after the workshop. An initial deployment after about a month is the target; scope and schedule are agreed for the project.
Who controls what the agent can do on its own? You define the assignment and required approvals with us. Permissions restrict access to the documents and systems selected for that work.