AI prototypes & feasibility
Know within weeks whether your AI idea holds up.
We build working prototypes with your real data, directly on your systems. You decide based on a result you can try out yourself, instead of slides and promises.
We build the 80 percent quickly.
With today's tools, the core of an idea is in place in a short time: the search, the answer, the analysis, the interface. Enough to test value, quality and acceptance on real cases.
We are open about the remaining 20 percent.
Edge cases, security, integration, operations and running costs decide whether it goes into production. We show them in the results report, with effort and risk, before you invest.
Fast, but under control
Why fast AI prototypes often do not hold up.
With AI, software is built in a short time. Without clear rules, typical mistakes creep in that only show up late.
Two truths
The AI builds a second solution for something that already exists. Duplicated data and duplicated logic drift apart.
A different architecture
Without rules, the AI decides afresh with every task. The system breaks up into one-off solutions nobody can keep track of.
Plausible, but wrong
Results look finished even when data was made up or edge cases were missed.
Silent side effects
A change works, but breaks something elsewhere that used to run.
Our guardrails keep the AI on track.
- Binding architecture rules: every piece of information has exactly one place that owns it.
- Automated checks catch duplicate truths and rule violations before anything goes live.
- Test first, then code. A result only counts as done once it is proven in a real browser.
- A person decides on the architecture and approves every change.
Why a prototype first
Many AI initiatives fail late and expensively. A good prototype shows early whether an idea pays off.
- of companies without AI name hard-to-calculate costs as an obstacle (Bitkom, 2026)
- 63 %
- say they lack the technical know-how to implement it (Bitkom, 2026)
- 85 %
- of GenAI projects are forecast to be abandoned after proof of concept (Gartner)
- 30 %
- of agentic AI projects are forecast to be cancelled by the end of 2027 (Gartner)
- 40 %
Sources: Bitkom survey of 603 companies in Germany with 20 or more employees, September 2026 (companies not yet using AI); Gartner forecasts from 2024 and 2025.
Formats
Three stages, each with a clear result.
You can stop after any stage. Each one delivers a basis for decision that has value even without the next.
Stage 1
Idea sprint
1–2 daysWorkshop
- Collect ideas and measure them against your processes
- Briefly review data, systems and legal situation
- Weigh value against effort
Result: a prioritised list of three to five ideas and a recommendation on which one is worth it.
Stage 2
Feasibility study
2–3 weeksTechnical test
- Test the core question on real data
- Measure quality, with examples and error rate
- Project the cost per transaction and of operations
Result: a report recommending go, adjust or stop. A well-founded no is also a result.
Stage 3
Prototype / proof of concept
4–6 weeksWorking software
- A usable prototype for real users
- Connected to your CMS, shop or data source
- Tested with the business team and measured
Result: a prototype your team can try out themselves, and a plan for the remaining 20 percent.
Without the hype
Not every task needs an agent.
We choose the simplest technology that solves the problem reliably. That makes solutions cheaper, easier to trace and easier to run.
When the rules are clear
Classic automation
Rules, workflows, interfaces. Often the best solution, and without any AI.
When knowledge has to be found
Search and analysis
Semantic search across documents, products and manuals, with sources.
When language is processed
Language models with approval
Summarise, translate, simplify and check texts. A person approves wherever quality and liability require it.
When several steps should run on their own
Agents
Only where a task really requires several decisions and tools. With clear limits and a log.
Typical prototypes
Ideas that can be tested within a few weeks.
Knowledge from documents
Employees ask manuals, contracts or policies in their own words and get answers with sources.
Measured by: hit rate on 50 real questions.
Enriching product data
Descriptions, attributes and translations are generated from existing data, directly in the shop or PIM process.
Measured by: approval rate by the editorial team.
Reading documents
Details from orders, invoices, forms or tenders end up structured in your system.
Measured by: error rate per field compared with manual entry.
Customer service replies
Draft replies based on orders and the knowledge base. The team reviews and sends.
Measured by: time per request and share of drafts used.
Accessible content
Alt texts, plain language and clean structure for existing pages, with editorial approval.
Measured by: coverage and review effort per page.
Search in the shop
Customers find products by their needs instead of exact terms.
Measured by: click and zero-result rate in an A/B comparison.
Why trueinnovation
We build prototypes so that something can come of them.
On your systems
We know Magento and Adobe Commerce, AEM, Magnolia and AWS from many projects. The prototype works with your data and interfaces.
Prototypes become products
hub-OSX and our AI solution for accessibility started as prototypes themselves. We know what lies between a demo and operations.
Your data stays with you
With European providers and models where required. Classification under the EU AI Act is part of the result.
An honest result
We also recommend not pursuing an idea when the numbers speak against it.