04 / AI & data
Put your business knowledge to work.
Help customers find answers and give your team a better way to process information. Devekko builds AI applications and the data foundations behind them around your business workflows.
Start with one useful task and a clear way to measure it.
04 / The investment case
Turn repeat work into useful capacity.
A peer-reviewed study of 5,172 customer-support agents at one enterprise found a 15% average improvement in issues resolved per hour with AI assistance. Effects varied across workers. It is evidence from one deployment, not a small-business benchmark.
Source: Brynjolfsson, Li and Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025
How the illustrative model works
- Assumptions: 60 useful hours a month recovered after human review, valued at $35 an hour. The 60 hours are hypothetical, not derived from the study.
- Annual capacity value: 60 × $35 × 12 = $25,200.
- First-year ownership cost: $6,000 setup + $400 × 12 run and care = $10,800.
- Capacity-value ROI: ($25,200 − $10,800) ÷ $10,800 = 133%.
- Recovered time becomes savings only when spending actually falls, or value when it goes to useful additional work. It is not a payroll reduction.
If only half the benefit arrives: 30 useful hours a month recovered: $12,600 capacity value, 17% capacity-value ROI.
Measure useful hours recovered and cost per completed task, alongside accuracy and rework.
All amounts USD over the first 12 operating months. Not a Devekko client result, forecast, or price quote.
Choose a task worth improving
Answer questions from your own information
Build an assistant over approved documents, product information, or internal guidance. Show the sources behind an answer and give users a useful next step when the information is missing.
Turn documents into reviewed records
Extract fields from incoming documents, check them against rules, and route exceptions to a person. Connect the reviewed result to the system your team uses.
Help people find the right item
Use semantic search to match the meaning of a question with relevant content or products. Keep access permissions and current source information attached to the result.
Assist a repeatable workflow
Classify requests, summarize material, or prepare a draft response. Define where the process can proceed automatically and where a person must review it.
Give the application dependable data
Useful AI starts with information your team can maintain. We identify the sources, connect the relevant systems, and make ownership and update rules explicit.
Data work can include pipelines, cleanup, shared definitions, quality checks, and reporting. It can also stand alone as a project when your immediate need is reliable information.
Built around your workflow
Connect the AI feature to the information and tools the task requires. Define who can use it, which records it can access, and when a person needs to review the result.
For document-based features, we design search and retrieval around your approved sources. Add storage, background processing, and monitoring where the workflow needs them. Review usage and operating costs alongside answer quality.
The result is an application you can inspect and operate, with a clear path from its data sources to the task it performs.
Test a focused pilot before expanding
We define a task, representative examples, and success measures together. The pilot shows how the system handles ordinary requests, missing information, and difficult cases.
You can assess answer quality, review effort, response time, and operating cost before committing to a wider rollout. For document answers, we check source support. For extracted records, we check the fields against known examples.
A production scope adds the permissions, monitoring, failure handling, and maintenance process the application needs.
Use the AI pilot worksheet to outline the task, approved sources, test examples, and review criteria. It is a plain-text template you can edit in any text editor.
A few useful answers
Before we begin.
How is a project priced?
Pricing depends on the workflow, data sources, permissions, integrations, and evaluation needed. We agree the implementation scope and price in a proposal. Model usage and infrastructure costs are considered separately from the development work.
Do we need clean, organized data first?
We begin by assessing the sources you have. Data preparation can be part of the project, and the first pilot may use a small approved collection.
Can we use our existing tools and databases?
We assess their APIs, permissions, and data formats. The architecture can connect to existing systems where they meet the requirements.
Is our information used to train a model?
Data use depends on the services and configuration selected for the project. Before connecting business information, we review the providers' data-use terms and document permissions, storage, logging, and retention. The scope should state which data may be processed and for what purpose.
Will the AI always be right?
Model output can be wrong. The design uses source references, validation, evaluation, and human review appropriate to the task.
How much will it cost to run?
Running cost depends on the model, request volume, input and output size, storage, and indexing. We estimate the workload and measure the pilot before setting operating limits.
Can we start with data work alone?
Yes. Connecting sources, improving data quality, and building reporting can be a useful first project. Those foundations can support an AI feature later.
Your next project
Bring one workflow to the conversation
Describe the information people struggle to find or the task your team repeats. We can use it to define the first project.