AI tools and products
We develop AI tools and features for digital products. We turn good model responses into a work process with quality checks, clear costs and the ability to improve the tool with your team.
We develop AI tools for processing materials, searching and preparing answers. We test quality on your examples and show the team how to assess the results.
Define the requirements
Together with your specialist, we determine what makes an answer useful and which errors are unacceptable. We choose a task and decide what data may be used.
- From you
- Anonymised examples, specialist participation and data handling rules
- Step result
- Task, requirements for the result and permitted data
Test on examples
We try the selected approach on real cases. We assess errors, processing cost and the amount of review left to a person.
- From you
- Examples of good and incorrect answers, specialist participation
- Step result
- Test results and cost assessment
Integrate the tool
We connect data and the interface to the work process. We allow for human review and cases where there is too little information to answer.
- From you
- Access credentials, permitted data and answer review rules
- Step result
- Working tool with settings and a description of its limitations
Hand over and train
We show how to test new tasks and assess changes to the model or settings. We hand over control examples with the tool.
- From you
- The employee who will be responsible for the tool
- Step result
- Source code, settings, control examples and review instructions
Quality belongs to your task
We begin by describing a result a specialist can use in their work. Together, we decide what they must verify, which errors change the meaning of an answer and how to tell when the task is complete.
We collect examples with people who know the work and turn their experience into quality criteria. We use them to choose how to build the tool and to test later versions.
Important differences become testable
We check whether a cited source supports a conclusion and whether a prepared draft helps complete the task. When information is insufficient, the tool must show exactly what is missing.
We test the approach on an agreed set of tasks, including ambiguous cases. The results help us discuss what is ready for use, where a person must be involved and what requires a different solution.
What supports an answer?
Three training cases for searching documents. Choose the conditions and examine the criteria for an answer.
The answer rests on an accessible source
A link to the document is needed, and the conclusion must match the information found. A link alone does not prove the answer is good.
The tool identifies the gap
It reports missing data. How to refine the question or involve a specialist depends on the workflow; a guess must not be presented as a found fact.
The answer does not reveal restricted information
We check permissions and exclude unavailable data. The interface must not reveal the name or existence of a restricted document if that alone is confidential.
These distinctions must be included in the test set. Then the team can assess answers to ordinary, difficult and ambiguous questions.
| Situation | Expected action |
|---|---|
| A suitable source is found | Answer with a link to the document |
| Sources contain no answer | Report that information is missing |
| Employee cannot access a document | Answer without restricted information |
Fit the result into real work
We design the specialist’s role together with the rest of the tool. They need to understand where information came from, what to check, how to correct the result and how to continue.
We agree on access to sources and conditions for handling data. We determine which information may go to external services and what must remain inside the company. These conditions shape the choice of model, storage and interface.
Corrections can become knowledge
New observations appear during use. A repeated correction may reveal poor source materials, a weak rule or a limitation of the chosen approach. We examine such cases and add significant ones to the set of tests.
Then the specialist’s experience does not disappear into yet another corrected answer. The team gradually refines its understanding of the result it needs and gains a basis for the next change.
Compare the cost of an accepted result
We count processing, retries and specialist time until there is a result that can be used. This makes it possible to compare the full cost of working with different models.
When changing models, we repeat checks using saved criteria and assess whether the switch is justified. We examine integrations and data formats to determine the work involved and the cost of migration.
Hand over a way to improve quality
With the responsible specialist, we work through result assessment, adding a new control case and actions when quality worsens. Settings, integrations, source code, criteria and test materials remain with the team.
We document the external model provider’s terms and what materials are sent. You retain the tool’s source code, settings and task criteria for choosing and testing later alternatives.
The next version does not start from zero
You can expand the tasks, revise rules and test other models using the examples you have collected. A new version must show which work it does better and what changes for the specialist.
Along with the product, we hand over a way to check and improve quality consistently on your team’s tasks.
What the work may include
Knowledge-base search
Answers linked to sources the employee can access.
Specialist assistant
Document drafts based on source materials.
Document processing
Field extraction and a queue for disputed cases.
Enquiry analysis
Topics, repeated issues and questions for the product team.
How the work is organised
Define the requirements
Together with your specialist, we determine what makes an answer useful and which errors are unacceptable. We choose a task and decide what data may be used.
- From you
- Anonymised examples, specialist participation and data handling rules
- Stage outcome
- Task, requirements for the result and permitted data
Test on examples
We try the selected approach on real cases. We assess errors, processing cost and the amount of review left to a person.
- From you
- Examples of good and incorrect answers, specialist participation
- Stage outcome
- Test results and cost assessment
Integrate the tool
We connect data and the interface to the work process. We allow for human review and cases where there is too little information to answer.
- From you
- Access credentials, permitted data and answer review rules
- Stage outcome
- Working tool with settings and a description of its limitations
Hand over and train
We show how to test new tasks and assess changes to the model or settings. We hand over control examples with the tool.
- From you
- The employee who will be responsible for the tool
- Stage outcome
- Source code, settings, control examples and review instructions
What we check
Employee permissions
We check answers against access to their sources.
Data location
We choose models and storage according to company rules.
Full cost
We count processing, retries and manual result review.
What we hand over
- Source files and project repository
- Access credentials and a list of external integrations
- Update and recovery instructions
- Documents covering rights and licences used
- Control examples, quality criteria and rules for human involvement.
What task do you want to give AI?
Share anonymised input data, an appropriate result or your experiment. We will discuss quality criteria, data constraints and where to begin the work.
After agreeing the work
- Anonymised examples of input data
- Examples of suitable and incorrect results
- Responsible expert and data handling constraints
Review your answers
This brief is too large for a Telegram link. Continue with the bot and add the remaining answers there.
Copy the full text or attach the downloaded file to your email.
Saving is unavailable in this browser. Your answers remain available while the page is open.