The tag view groups articles by a concrete operational problem or implementation type, without exposing hundreds of raw frontmatter labels.
One RTX 3090 with 24 GB of VRAM was enough to run Qwen3.8-27B and fix a bug in Go code in 502.14 s, with 52 tool calls, about 33.88 Wh of energy and a score of 100/100 on our criteria. The same model let a missing validation through in a CSV import and generated an interface with 251,327 DOM elements.
An AI app makes sense when a person needs a new place to work and approves the result themselves. An automation carries out steps in the background that you can describe before the start. An AI agent chooses the next steps based on the content of the case, within the rules. At Syntalith each form starts from €3,000 net, and delivery takes 2 to 16 weeks.
You move an agent to a new environment after comparing it on a golden set: the same inputs, the expected decisions and a process owner's decision on every difference. In our demonstrator the set holds 336 behaviours, and 7 unresolved differences block the switch. A full run of the set costs about €0.76.
An AI system register is a list of the systems in use, with an owner, the company's role, the risk status and the missing documentation. The register's rules run before the model, and a person approves the risk classification. The transparency obligations in Article 50 of the AI Act apply from 2 August 2026. At Syntalith the application costs from €3,000 net. At 40 register entries a month of 90 minutes each, it pays back in 2.0 months.
AI for case law research gives, next to every claim, the exact passage of the judgment and the source record, and a claim without word-for-word support in the approved collection fails validation. A request for legal advice is refused. A lawyer judges the argument. At Syntalith the app costs from €3,000 net, and at 200 research questions a month of 45 minutes each it pays back in 1.0 month.
When a photo, a voice note and a job card from one call-out give different data, the AI system keeps each file separately, shows the conflicting values side by side and blocks closing the case until a coordinator decides. A convenient summary must not erase the difference between sources. At Syntalith the app costs from €3,000 net, and with 300 field call-outs a month at 15 minutes each it pays back in 1.6 months.
A team of agents in CrewAI is worth choosing only when it gives a better result than one agent on the same tasks, criteria and tools. In counterparty research, three roles usually use several times more model calls: at 200 files a month that is about €77.11 against €25.70 for a single agent. Scenario.
An agent that assembles a client file saves the state of the case after each arrival: documents received, gaps, conflicts and the decisions it is waiting for. When a corrected file arrives a week later, it resumes from the right step. Accepting the client stays with a person. At Syntalith an agent costs from €3,000 net. At 300 files a month, it pays back in 1.2 months.
At project kickoff the system links an email thread, a PDF and the note from the opening meeting into a plan in which every finding has a source. When two documents give two start dates, the conflict stops the plan from closing until a person decides. At Syntalith an automation costs from €3,000 net, and with 40 projects a month at 120 minutes each it pays back in 1.5 months.
You compare agent frameworks on the same set of cases, with the same answer format and one test endpoint. In our demonstrator six technologies, from LangGraph to plain Python, run 50 cases, including 12 required refusals. The result points to two candidates for a pilot on a company process.
AI invoice coding proposes an account and a cost centre only when it has a basis: an earlier entry for the same supplier and the same type of cost, or an order. An invoice without a basis goes to the accountant, who makes the decision. At Syntalith the automation costs from €3,000 net, and at 1,000 cost invoices a month of 3 minutes each it pays back in 2.3 months.
A legibility gate measures the brightness, contrast and sharpness of a photo of a transport document before it is read. When the image is unfit, the system records that no reading was attempted and asks for a new shot, in place of returning a full set of fields of unknown reliability. In our demonstrator, 5 of 7 image variants were stopped without calling the model. At Syntalith document reading automation costs from €3,000 net and pays back in 2.0 months.
A management report written with AI holds no made-up numbers when the figures and the commentary are kept apart. Every amount has an ID and an approved source, and the release check rejects any value outside that set. A conflict between sources stays visible until the controller decides. At Syntalith an automation costs from €3,000 net. At 12 reports a month of 240 minutes each, it pays back in 2.4 months.
You turn an order written in the customer's words into item codes and quantities with rules: matching to the catalogue, converting units and deciding on substitutes. Unclear lines go to an operator. An agent is needed only when the agreed path is not enough. At Syntalith an automation costs from €3,000 net, and with 1,000 orders a month at 6 minutes each it pays back in 1.1 months.
A model on your own server does not describe the whole isolation. With confidential documents you trace the path of the file, the query and the answer, check that there are no outbound connections and settle how a new release reaches a zone without internet. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. With 500 document questions a month at 8 minutes each, it pays back in 2.0 months.
An AI agent's access to a tool should not mean the right to every operation. A separate policy layer checks the identity, the resource and the context before each call and returns allow, deny or a requirement for a second approval, with a record of the rule applied. At Syntalith an agent costs from €3,000 net, and with 2,000 agent actions a month at 3 minutes each it pays back in 1.1 months.
Permission-aware RAG checks access before the model sees a document: the search filters by the employee's groups, and PostgreSQL enforces the row access rule again. At Syntalith an AI application on company documents costs from €3,000 net. At 400 questions a month that each take 10 minutes of searching, it pays back in 1.5 months.
Asking a database questions with AI makes sense when the definition of the measure and the SQL query that ran are visible under the result. In our demonstrator the system reads only four approved views with an account that cannot write, and a question costs $0.004400. At Syntalith the application costs from €3,000 net. At 300 data questions a month of 15 minutes each, it pays back in 1.5 months.
For a payment reminder, one model writes a draft by the company's communication rules, and a second with the opposite goal looks for threats, pressure and promises nobody made. Rules hold back disputed cases, and the operator sees the draft with the check result and decides alone whether to send it. At Syntalith the automation costs from €3,000 net, and at 300 reminders a month of 8 minutes each it pays back in 3.1 months.
Quoting a request for quotation with AI involves no guessing when the data schema allows only two results for a line: priced with evidence, or passed on for a decision with no price field. Known lines get a price from a precedent or a rule, and unusual machining goes to the process engineer with a specific question. At Syntalith an automation costs from €3,000 net. At 200 RFQs a month of 45 minutes each, it pays back in 0.8 months.
AI agents
The agent moves a case through its steps, and actions with consequences wait for approval from the person you name. Before you sign, we write down the goal and a test set built from your cases.
Times are shown in your own time zone. We work with clients across time zones.
See AI agentsOpenAI makes ChatGPT and Anthropic makes Claude. We belong to the partnership programs of both companies, so we work with their tools every day and learn from the creators themselves.
Claude Partner NetworkParticipation in programs does not mean that OpenAI or Anthropic endorses our services.