Challenge us
AI agents in the process

AI that asks instead of guessing.

Put AI agents straight into your workflows: an agent checks applications including their attachments, looks things up in your data and knowledge sources, asks questions and delivers a validated result. Where the case goes next is set by your process model – not by the AI.

An agent as a process step

Many AI features end with a single call: text in, text out. An agent can do more. It works through a task in several steps, uses approved tools to do so and decides for itself when it is finished or needs a person.

In the workflow, the agent is a node of its own. You describe the task in plain language, choose the fields and attachments it may see, and define which tools it may use. Its result lands in fields of your workflow form and is validated like a submitted form.

From there, the process continues through fixed exits: result, question, handover to a person, limit reached or error. Each exit is a visible path in the process model. That keeps what the AI did traceable – and what happens next under control.

Illustration: a robot checks an application with a magnifying glass at a desk; five fixed tracks lead from its desk to five exits, one of them back into the process.

What the agent looks like in the process

Application
AI agent
Five fixed exits
  • Result validated result, on to approval agent_done
  • Question question to a person, after which the agent continues agent_question
  • Handover to a person handed over for manual processing agent_handover
  • Limit reached cost, time or step limit reached agent_limit
  • Error error, with its own error path agent_error

How a run works

  1. The process reaches the agent node. The case waits; a background service picks up the run immediately.
  2. The agent reads the approved fields and attachments. PDFs are converted to text beforehand.
  3. It uses its tools: reading tables, reading long documents page by page, searching knowledge sources.
  4. If something is missing that only a person knows, it asks. Once the question is answered, it carries on from where it left off.
  5. It delivers its result into form fields. If the result passes validation and the agent is confident enough, the case moves on. Otherwise a person takes over.

What the agent node brings

The process sets the paths

The agent can only trigger one of the fixed exits. It cannot choose a path that nobody has modelled.

Result validated like a form

Required fields, allowed values and rules come from the workflow form. If the result fails twice, a person takes over and sees the draft.

Fixed limits

Model calls, tool calls, run time and cost are capped per run. The installation sets upper limits that no workflow can exceed. Once a limit is reached, the case continues through an exit of its own.

Repeatable testing

Record runs and replay them at no cost. Fixed test responses check paths such as question or limit without calling a model.

Traceable

Every step in the log, with tokens and cost, without keys or secrets. In the history, every AI output is marked as such.

Your models

Any model with an OpenAI-compatible interface: Azure OpenAI, an in-house model farm, locally operated models. No automatic fallback to a public provider.

Protection against injected instructions

Form texts, attachments and search results reach the model as data with their origin, not as instructions. Whoever writes an application cannot give the agent orders.

People in the process

Questions and handovers are modelled tasks with fixed assignees. Below an adjustable confidence threshold, the agent hands over instead of guessing.

Tools only with approval

Without approval, the agent reads only what the node gives it. Tables are approved one by one, down to field level if required.

External knowledge sources

Read-only tools from external MCP servers (e.g. documentation, DMS, search). A person approves each tool individually. If the provider changes a tool’s description, it has to be approved again.

Long research

Intermediate results appear in the case history. On long runs, older search results are shortened once the model has been able to note what it still needs.

Configured in the editor

Task, inputs with their origin, result section, model, tools and limits in the workflow editor. An AI assistant can create agent workflows too; validation then reports missing exits and approvals.

Orchestration

Several AI agents, one process – and people keep the decision

A single AI model always sounds convinced, even when it is wrong. That is why, at important steps, the workflow application lets several agents work independently of one another, if you wish with models from different providers or with your own model environment. The process then evaluates the results by a fixed rule, such as unanimity or majority.

If the agents agree, the case continues automatically. If they do not, it goes to a person. A further agent then prepares the decision: it shows where the assessments differ and proposes a way forward. If an agent does not respond in time, the process evaluates what is available once a deadline has passed.

Every step stays traceable: which model decided what, and why, is recorded in the case history. You can use the same building blocks without AI, for instance when several departments are to comment in parallel.

Case
  • Agent 1 model from provider A
  • Agent 2 model from provider B
  • Agent 3 your own model environment
Evaluation unanimous · majority · after deadline: what is available
  • Agreement The case continues automatically.
  • Disagreement A person decides. A further agent shows the differences and proposes a way forward.
More on orchestration on our blog
Acting agents

Agents that may act – but only through your process

An agent can do more than read and summarise: it creates customers, drafts quotes and triggers approvals. To do so, you hand it process steps such as "Create customer" or "Send quote". You model each step in the workflow as usual: whether data is saved, who has to agree beforehand and who gets notified.

The agent chooses for itself which step it needs and when, and fills in the required details. Then the process takes over, and the agent learns the result. If sales management rejects a quote, it revises the quote and resubmits it.

The AI decides what needs doing. Your process carries it out.

Three examples: master data, a quote with approval, two quotes at once
  • Control What the AI may do is in the process diagram. Delicate steps such as deleting or sending you tie to an approval.
  • Traceability The history shows what the agent wanted, who approved or rejected it and what was saved.
  • Flexibility The agent orders the steps itself and responds to rejections. Not every special case has to be modelled in advance.
  • Clear limits Cost, run time and number of steps capped per agent. A running agent can be stopped.

Further reading on our blog

Five ways to use AI agents in business processes

From the assistant in a business application to the AI that builds the process: who triggers it, what the agent may do, who has the last word.

Several AI agents, one process

Why one model alone is not enough, and how a person decides when the agents disagree.

AI agents that may act

How an agent creates customers and drafts quotes without working around your data. With three examples.

Challenge us.

Send us a complex process. It does not have to be your internal one – an invented one, a publicly documented one, one from an old manual. Anything with roles, deadlines and approvals. The trickier, the better.

You get it back: built, played through by every role, with the test report and with test accounts. That includes a process with an agent step: the AI builds it and plays it through as every role. Paths such as question or limit are checked with fixed test responses, without calling a model.

What you get and what you do not: Frequently asked questions

You have files – Visio, PDF, BPMN? Send them to info@flyingdog.de. We need no particular format.

Frequently asked questions

Who is this for?

For medium-sized and large companies and organisations, for public administration and universities – and for BPM consultants and integrators who build processes for their clients. What they have in common: many workflows with roles, deadlines and approvals, often across departments. For a single small workflow in a small business we are not the right tool.

What does the AI check – and what does it not?

It logs in as every role, works through the process, lets deadlines actually run out and then reads the stored state and the send log. It checks the flow, not whether a decision was professionally right. It checks what your material said – what was not in there, it cannot find. Whatever the material does not settle is marked in the process as an assumption, visibly. A person decides on every assumption before anything goes live.

What do I get when I challenge you – and what not?

A running process with forms, roles, group inboxes, deadlines, escalations and mail templates – plus the data model, which the AI generates along with it, including individual user interfaces inside the process forms or outside them. Also test accounts for every role and the test report from the play-through. Interfaces to your systems are possible only to a limited extent. It is a test environment, not production; deadlines run on test values – minutes instead of days. One process per challenge.

What does the test process cost?

Nothing. There is no charge for the test process – one process, in our test environment, with test accounts for every role. We only talk about licence and installation once you have played it through.

Do I have to send an internal process?

No. An invented one, a publicly documented one or one from an old manual will do – as long as it has roles, deadlines and approvals. Send material to info@flyingdog.de; the form deliberately has no upload.

Where does it run?

The test process runs in our test environment. With a licence, the platform runs on your servers – AI operation included – with the source code shipped.

What happens after the challenge?

You play the process through with your people. Then we talk – about licence and installation. You can still buy afterwards.