A good start: AI agents with Klardaten, Claude, and Langdock in the daily routine of a tax advisory & accounting firm

Anyone who has followed the posts of tax advisory & accounting firms on LinkedIn in recent months will recognize a pattern. More and more tax advisors are building their own agents using Klardaten and Langdock: for pre-capturing documents, for provision calculations, for specialized research on tax-related questions, or for draft responses in the firm's style. Some go further and have the agent create posting lines directly in DATEV accounting. Klardaten provides the connection to DATEV, Langdock the AI platform with the models. This creates a tool that makes a specific task in everyday life significantly faster.

This development is good. It shows a profession that is actively engaging with technology. And it creates real value: anyone who uses a cleanly built agent for a recurring process today feels the relief immediately. This is a good start. The decisive factor is where the path leads next.

What already works well today

An AI agent with DATEV access can do more than many realize. It reads and writes documents, researches facts, creates posting suggestions, or drafts client communication. For clearly outlined, recurring tasks, this is a very good solution. An agent that categorizes incoming client emails and prepares a draft response in the firm's style saves time every day.

The charm lies in the immediacy: you identify a problem, build an agent for it, and have a working tool the next day. Such tools make the individual processor faster. For many tax advisory & accounting firms, this is exactly the right first step.

The difference lies in the entire process

A single agent handles a self-contained task: posting a document, performing a calculation, drafting an email. The actual work in a tax advisory & accounting firm, however, is rarely a single task. It is a process of many steps: gathering documents, posting, checking against rules, clarifying open points, approving.

In software development, dealing with this has long been a part of everyday life. There, an agent takes on a larger task, plans the steps, works through them, and checks its result against defined criteria. The developer does not follow every step individually, but steps in where their judgment is needed. In tax advisory, such a process looks different, but the principle can be applied: the agent runs through the entire process, the specialist checks where it counts professionally.

This is precisely where the limit of the single agent lies. It solves one step, but it does not carry the process. Three points show what matters here:

First: from copilot to autopilot. A self-built agent is a copilot. It sits next to the specialist, makes a suggestion, and the specialist accompanies every step. This helps, but it does not take away all the work, because in the end, every process still depends on a human who picks it up and continues it. The actual leap is the one to autopilot: the agent runs through the entire process independently, from processing to checking and presentation, and the specialist only intervenes at the points where their judgment counts.

Second: the depth of accounting. Accounting is not a text problem that a language model solves on the side. It is a system with hard, built-in truths: the balance sheet must balance, balances must be reconciled, deadlines interlock. An agent can post every single document plausibly and still deliver accounting in the end where the balance sheet does not balance. German tax law consists largely of exceptions, and checking these exceptions is precisely the more time-consuming part of the work. A generic tool processes accounting on the surface. A platform built solely for tax advisory knows the rules, checks against them, and catches special cases before they become errors that only the annual financial statement brings to light.

Third: secure write-back. Reading data from DATEV is something a connection like Klardaten can do. Writing back is the other half, and that is the delicate one. If an agent runs twice because a connection briefly drops, the booking is in the system twice without protection. The fact that a process ends up exactly once and correctly in DATEV is decided not by the model, but by the infrastructure behind it. A platform built for everyday tax advisory & accounting firm life has this infrastructure.

What an agentic platform does differently

Atlas, the platform from Limetax, builds exactly this level. Atlas builds on DATEV and uses the same models that are also behind a self-built agent. The difference lies in what is created around the models: the continuous process up to approval, the link with the posting proof, the automatic check against the rules of accounting, the secure write-back to DATEV, and the control of similar processes across many mandates. Because a chat is only as fast as a human types and reads. Querying and approving missing documents across a hundred clients in a single run is something else entirely, and that is exactly where the gain in capacity lies that a tax advisory & accounting firm needs.

This includes that the check does not have to take place in the chat. If an agent presents a result for control, the specialist opens a specialized application for the respective subject area instead of going through each point individually in the chat window. A dedicated accounting tool is integrated directly into Atlas for this purpose, and every check that happens there flows back into the platform.

An agentic platform does not just solve individual tasks faster; it regulates the interaction between human and agent across the entire workday. This is exactly where the sensible use of AI in tax advisory is decided. And this level will not disappear with the next model leap: if the models improve, Atlas will also improve, while the process, check, write access, and proof remain unchanged in value.

Why a group has an advantage

Every correction a specialist makes to a result is valuable information. It shows where the agent was not yet precise enough. With a single built agent, this correction remains local and, at best, improves one's own tax advisory & accounting firm. In a group with a shared platform, it flows back into the entire system: what a specialist corrects in one tax advisory & accounting firm improves processing in all others.

The same applies to new skills. If a tax advisory & accounting firm builds a skill for a specific process, it is available to the entire group. What was solved once in one place does not need to be built again in any other.

The right path for your own tax advisory & accounting firm

For tax advisory firms building their first agents with Klardaten and Langdock today, this is explicitly not a warning, but an encouragement. Anyone who starts automating processes gathers experience that will be valuable later. The first agent for a specific bottleneck is a sensible step.

The key is to keep the next stage in mind. A single agent makes the processor faster. An agentic platform changes how many clients a tax advisory & accounting firm can manage because it controls, checks, presents for approval, and learns from every correction the entire process. The difference lies not in the model, but in everything needed for a tax advisor to sign their name to a completed piece of work.

This is exactly the level we are building at Limetax, together with our tax advisory firms and deeply integrated into their processes. Not as a replacement for the specialist, but as infrastructure that takes care of the routine and leaves the responsibility where it belongs.