Process assessment with a clear decision
A prioritised decision on where AI creates measurable relief, where classical automation is more reliable, and which risks, data and approvals must be considered.
01Custom software & automation
From need to reliable delivery
Software for people. Ready for machines.
Fewer manual handoffs. Reliable workflows. Systems that work together securely.
I connect existing systems through clearly defined interfaces and integrate AI only where it provides measurable relief. Employees, third-party systems and AI use the same validated business logic — traceable, limited and without unnecessary data access.
At a glance
- Your processes remain the starting point — not the model
- Existing systems are connected rather than replaced prematurely
- Every access path receives only the data and functions it needs
- Every action remains traceable, limited and reversible
The underlying problem
When systems slow workflows down
Many applications work well on their own. But when data needs to reach another system, dashboard or automated workflow, people become the interface: exporting, copying, reconciling and correcting. That costs time, creates errors and makes the real status difficult to understand.
An integration-ready application exposes selected business functions through stable APIs for third-party systems and controlled tools for AI. Whether checking an order, classifying a document or retrieving a status, business rules are implemented once and apply equally to employees, connected systems and AI.
No uncontrolled side door
Third-party systems and AI bypass neither business logic nor permissions. They receive clearly limited access to the same validated functions.
Not every process should be operated by AI. I first determine where classical automation is more reliable, which actions require human approval and how large the permitted impact radius may be.
Architecture principle
One application. Three controlled access paths.
Employees use familiar interfaces, third-party systems use standard APIs and AI models use a dedicated, secured interface. All rely on the same rules, processes and data.
Employees
work as usual in day-to-day operations
Third-party systems
ERP, SaaS or partner solutions
AI systems
internal, external or replaced later
through clearly defined access
User interface
web, mobile or business application
Standard APIs
process-specific data instead of full access
limited
Secured AI access
only clearly approved actions
into the same business logic
Your business logic
one reliable source of truth for every access path
ProcessesRulesData
Dedicated identity and minimal permissions
Approvals, limits and shutdown controls
Complete traceability
No access path bypasses permissions or business rules.
Concrete delivery
What you actually get
Integration-ready custom software
New or existing applications gain stable APIA clearly defined interface through which systems exchange data and securely call functions. for third-party systems and controlled AI access. Processes and business logic remain the shared foundation without a premature full replacement.
Process-specific interfaces
APIA clearly defined interface through which systems exchange data and securely call functions. and AI tools expose only the functions and data required by the process. Depending on the need, GraphQL, REST, events or MCP provide clearly permissioned and traceable access.
Automation with a clear handover
AI can prepare information, pre-qualify cases or perform clearly bounded tasks. Uncertain and consequential cases are deliberately handed over to employees.
Transparent operations and measurable value
Technical quality, cost and business workflows remain traceable through appropriate dashboards. Depending on the need, Grafana & Elastic StackProven platforms that make technical metrics, logs and operational flows visible, searchable and traceable through dashboards. provides the foundation. For critical business processes, I build custom dashboards that make states, bottlenecks and outcomes easy to understand.
Production experience
What this looks like in practice
Hoffmann Group — opening business functions safely to AI
Data Hubs support business processes for orders, products and identities. A controlled interface opens selected functions to AI without duplicating data or bypassing permissions. Curated AI guidance helps consuming teams connect existing applications to the APIA clearly defined interface through which systems exchange data and securely call functions. and adapt integration code within defined interfaces, rules and security boundaries. Business logic and data ownership remain in the existing system.
Controlled AI support in engineering
Custom AI tools work on clearly bounded development tasks through the ticketing system. Roles, automated quality checks and human reviews constrain the assignment — AI supports the workflow without taking uncontrolled responsibility.
Business controlling — clear decisions instead of spreadsheet workarounds
A custom controlling dashboard for a multi-site healthcare practice consolidates planning and actual data in one reliable view. Utilisation, revenue, capacity, site comparisons and data quality become transparent — deliberately without AI where transparent calculations lead to better decisions.
Herr Daniel zeichnet sich dadurch aus aktiv auf Problemstellungen oder Lücken in Konzepten hinzuweisen und im Team eine optimale Lösung für solche Fälle zu erarbeiten. Ich bin Herr Daniel persönlich sehr dankbar für die stets offene, zielführende und angenehme Zusammenarbeit. Er ist kommunikationsstark, verlässlich, hat hohe Ansprüche an guten Code und die Zusammenarbeit mit ihm hat wirklich Spaß gemacht.
Dorothea Schwarz
Digital Product Owner Data Hubs · Hoffmann Group
My approach
Control is part of the architecture
AI that can act does not need special privileges. Four principles keep integration, operations and future model changes under control:
- 1
Dedicated identity instead of shared authority. Every AI integration receives exactly the permissions, tools and data access required for its specific task.
- 2
One business logic for every access path. User interfaces, APIA clearly defined interface through which systems exchange data and securely call functions. and AI tools use the same rules, checks and roles. Each process receives only the data and functions it actually needs, avoiding unnecessarily broad access and parallel shadow systems.
- 3
Approvals before consequential actions. Limits, confirmations and defined handovers to employees constrain the impact radius and keep accountability visible.
- 4
Models remain replaceable. The control layer separates your processes from the provider, so outages, cost changes or model swaps do not require rebuilding the application.
Clear answers
Questions you might be asking
When is AI not the answer?
When rules can be expressed completely and unambiguously, classical software is usually cheaper, faster and more reliable. AI is most useful for variable language, unstructured information and tasks with judgement. Irreversible or safety-critical decisions still require human approval.
How can an external AI operate our software securely?
Not through direct database access or shared user accounts. AI receives its own identity and only approved tools. Permissions, data scope, limits and required confirmations are defined per task; all actions remain logged and access can be disabled at any time.
Does this lock us into one model provider?
No, when the integration is separated cleanly. Business logic and tool interfaces remain independent of the model. Cloud models, local models or multiple providers can therefore be used and replaced later without rebuilding the application.
How does quality become measurable?
Verifiable examples and edge cases are defined for the actual use case. Measurements can include domain correctness, unsupported answers, cost, runtime and safe handovers. The right metrics depend on the real task rather than a generic AI score.
What does AI cost in production?
That depends on the model, data volume, request volume, response length and operating model. Costs per transaction are measured early and optimised through limits, caching or smaller models. A reliable cost range follows the use-case assessment, not a generic request count.
Do existing systems need to be replaced?
Usually not. I first identify which functions and data already work reliably. APIs and controlled AI access then extend those systems step by step without putting daily operations at risk. Replacement only makes sense when technical limits or ongoing costs justify it economically.
Related areas of expertise
Get an initial assessment
Could a custom solution be the right fit for your project?
Briefly describe where time, quality or room to act is being lost today. I will personally assess whether custom software is the right path and what a sensible first step could be.