Solutions and use cases
AI assistant app for reviewable tasks
An AI assistant app should help with a defined task and make its result easy to check. Establish the useful output, the data it may access and what users should do when the result is incomplete or wrong.
01
Begin with a specific task
Searching internal instructions, drafting an answer and extracting fields from a photograph are different workflows. Each needs its own data sources, acceptance criteria and error handling. Asking for chat inside an app is not enough to define the product.
Read more
Consider an assistant that retrieves relevant maintenance instructions and prepares a draft answer. This is a planning example, not a client reference. The design must address whether the document is current, whether the user may access it and what happens when no suitable source exists.
02
Typical features: search, chat and image recognition
Search may help users navigate an agreed document collection. Chat can clarify the question and display a draft with supporting passages. Access filtering needs to apply to retrieved content, not just the visible document list.
Read more
For image or document processing, show extracted fields and allow correction before saving. A blurred image or unsupported document must not silently produce a final record. Choose features using representative samples and acceptable error boundaries, rather than relying on an attractive demonstration.
03
When another approach may fit better
A structured search or explicit rule can be preferable when it reliably solves the task. AI introduces evaluation work, a provider dependency and potentially ongoing usage charges. If the source material has no owner or cannot be assessed, improving the information may be the necessary first step.
Read more
Consider the consequences of an incorrect suggestion. Start with limited, reversible actions and human approval where appropriate. Generating a suggestion and changing an external business record are separate capabilities with different access and review requirements.
04
Data flow and privacy decisions
Map what leaves the phone, what reaches the backend and what is sent to the selected processor. Provider credentials should remain outside the mobile client, with requests authorised and limited through the backend. Send only information needed for the task.
Read more
A selected passage or redacted sample may be sufficient instead of a full document. Check the chosen provider's concrete processing, retention and access arrangements with the responsible project stakeholders. Diagnostic logging should not automatically capture sensitive prompts or documents. Conversation history and deletion also need explicit decisions.
05
Evaluating quality and handling failure
Prepare usable test cases covering normal questions, incomplete context, ambiguous requests and missing answers. Define the acceptable result for each task. Fluent wording does not establish correctness, so assess whether outputs are supported by the intended sources.
Read more
Users need a clear route to revise, retry or continue without AI when processing fails. Treat instructions embedded in source material as untrusted input. The backend should restrict available operations independently of generated text. Revisit accepted cases when the model or processing configuration changes.
06
Delivery stages and decision points
Discovery needs the intended users, the task and redacted examples. Outputs should include scope boundaries, a data-flow plan and evaluation criteria. An early trial can assess whether the available sources support acceptable behaviour.
Read more
If they do not, narrow the task, improve the data or choose another approach. A UX prototype then follows the question, waiting state, result and correction route. Development reviews should distinguish mobile behaviour, system access and processing. Before release, test permission boundaries and failed provider responses alongside representative quality cases.
07
What affects the price?
Task complexity, source quality, integrations and evaluation requirements affect the effort. A document search tool differs from an image extraction workflow or an assistant performing actions across systems. Model choice, response-time expectations, data preparation and content updates matter as well.
Read more
Ongoing cost may include processing usage, hosting, storage and quality reviews. Agree request limits and failure handling early. A focused first release can use selected sources and question types with human-approved outputs, leaving wider autonomy and integrations for separate decisions.
08
Preparing an enquiry and a useful handover
Describe a real task and show what a useful, acceptable result would look like. Explain who owns the source information and who checks the outcome. Send redacted examples rather than confidential production material.
Read more
Handover can include source code, build instructions, processing configuration, evaluation cases and known limitations. Define who maintains the sources and reviews changes in model behaviour. The operational process should make it possible to investigate a failure without routinely exposing the content of user conversations or uploaded documents.
Frequently asked questions
Will the assistant always return a correct answer?
No. Incomplete and incorrect outputs must be considered. Appropriate sources, representative evaluation cases and human review belong in the product design.
Do we need to train our own model?
Not necessarily. Establish the task, available information and acceptable output first. Assess possible approaches before committing to a particular technology or training effort.
Can internal documents be sent to an external provider?
That requires checking the specific content, permitted use, provider arrangements and access boundaries with the responsible stakeholders. Use redacted samples for preparation.
Can the assistant change records in other systems?
That is a separate capability requiring explicit permissions, approval rules and action limits. The initial release can remain focused on reviewable suggestions.