AI integration services connect model capabilities to an existing application or workflow. Obdurate Systems designs implementations for classifying requests, extracting document data and finding approved information, with validation and review around important actions. Start with a useful task and a way to judge whether AI helps.
An illustrative workflow
- 01
Receive
Accept a defined input from a source the workflow is authorised to use.
- 02
Interpret
Ask the model to classify, extract, retrieve or draft within a bounded task.
- 03
Check
Validate the response against expected fields, source data and business rules.
- 04
Review
Let a person approve material changes or resolve an uncertain result.
- 05
Continue
Pass the approved output to the existing workflow, with a manual fallback.
What arrives, what needs interpreting, and what happens next? A useful scope names the input, the desired output, the person who owns the result and the repetitive work involved.
Exact calculations, status synchronisation and explicit approval rules usually belong in ordinary code. AI is a candidate where the input varies and interpretation helps. A rule, existing application feature or smaller non-AI change may be the right recommendation.
These are prospective use cases to assess and test, not examples of completed client AI deployments.
- Request classification
- Suggest a category and destination for an incoming operational email, with unclear messages routed for review.
- Document data extraction
- Produce draft structured fields from a document, then check required values and references before any business record changes.
- Internal knowledge search
- Help staff find information in approved sources, respecting access permissions and showing references they can inspect.
- Draft operational updates
- Assemble a proposed update from authorised business data for someone to check before sending or saving.
- Bounded agent actions
- Let an agent propose a defined change using restricted tools, with approval before material writes to a business system.
Agree what a useful result looks like before building. Representative samples should include ordinary inputs, incomplete information and cases the model should decline or send to a person. Agree suitable test material and handling for sensitive data first.
Evaluate output quality, failure types, human-review workload, response time and operating cost. Acceptance thresholds depend on the task and are agreed during discovery. The pilot should show whether the workflow remains useful when inputs are difficult and the model is wrong.
Proposed controls include scoped access, restricted tools, output validation and approval gates for consequential actions. Incoming emails and documents are untrusted input; their contents must not override the workflow's instructions or permissions.
Give uncertain or invalid outputs an explicit route to a person who can correct, reject or complete the task manually. Appropriate records should show what was proposed, what was approved and what actually changed.
Implementation depends on the applications, approved data sources and interfaces available. Provider choice, hosting, retention and data handling need to be scoped against your requirements; they are not assumptions to leave until the end.
We assess native AI features and packaged products alongside a custom integration. Extending a tool you already use may be sufficient. LLM integration becomes custom work when the task, permissions or path back into your business systems requires it, with a clear owner for operation and ongoing evaluation.
Can AI work with our existing ERP, CRM or internal application?
Potentially, through supported interfaces and authorised data access. We first check what the task needs to read and write, and what the application permits.
Which tasks should remain rules-based?
Tasks with exact, stable rules: arithmetic, required-field checks, status synchronisation and approval authority. A workflow can use AI for interpretation and ordinary code for validation and execution.
What happens when the model is wrong or uncertain?
The proposed workflow should validate the output and route unsupported cases to review or a manual fallback. No model guarantees correct answers; evaluation needs to test errors as well as successful examples.
Can a person approve changes before they reach a business system?
Yes. We agree which actions require approval and design the workflow to enforce that decision before a material change is made.
What would a useful first pilot prove?
Whether one bounded task produces acceptable results with a workable review burden, response time and operating cost. It should also expose failure cases and support a decision to proceed, adjust the approach or stop.
Describe the input, the decision or action, the systems involved and the mistakes a person needs to catch. Start with a short description, without private documents or sensitive business records.
For workflows that span several tools, explore our custom business software.