Decision Engine: Use cases and presets

The Decision Engine ships with ten ready-made presets, for example for ticket routing, inbox triage, lead qualification and AI agents. For each preset, this page shows what it decides, what input it expects, which answers are possible and when a human should review.

What a preset is

A preset is a question pack stored on our side, with fixed options, levels or yes/no. You send only the preset id and the state, meaning the text or JSON object to decide about. Because the questions live with us, they read the same on every call, which makes per-preset calibration possible later on. The preset's version number is echoed in every answer.

How to call a preset

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "ticket-routing-v1",
    "state": "Since this morning I cannot log in."
  }'

The full list with all questions and example states is returned by GET /v1/decisions/presets.

What applies to every preset

• The engine does not act on its own. It writes no reply to anyone, changes no text and executes nothing. Your workflow decides what follows from the result.
• Every answer carries a recommendation: recommendation: "act" (your application can act) or "review" (a human checks first), with the reasons in reasons. The default threshold gives act as long as there is no warning signal; review_only presents every answer to a human, autonomous none. Details under Automatic or human.
• The per-preset rules apply in addition. Whatever is listed under “Human review” below goes to a human even with act, and so does every answer with review.
• Every choice question has an abstention as its last option, such as unclear or other. If the model picks it, the recommendation is review (chose_abstain_option, except in autonomous mode). For “unsure, please check” that is more reliable than a percentage.
• Choice questions are double-checked: a second pass with reordered options, at twice the input tokens. If the answer flips, the engine recommends review (unstable_under_reordering), even at a probability close to 1. It does not catch misjudgements that come out the same in every order.
• Three presets expect a JSON object as state (Application complete?, Grounding check, Agent step gate). Send the object directly as state, not as a string; the questions refer to its fields.
• For two presets we recommend review_only or a stricter rule of your own: Publication release check and Agent step gate. The reasons are given there.

Ticket routing

Id: ticket-routing-v1 · For: customer service, helpdesk, shared service centres

Assigns a support message to a team and rates urgency, review need and language. The input is the message text from your ticket system. team picks the queue, urgency the position in it, language the template of the first reply.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "ticket-routing-v1",
    "state": "Seit heute früh kann ich mich nicht mehr anmelden. Außerdem wurde meine Rechnung für September doppelt abgebucht — bitte prüfen Sie das dringend."
  }'

Human review: unclear and every ticket with needs_human: true, before a reply goes out.

Adapting it: replace the three teams with yours. The description per option is the text the model reads: “invoices, payments, refunds” separates better than just “accounting”. The abstention unclear stays in last place.

Inbox document classification

Id: inbox-document-v1 · For: administration, back office, mail room, accounting

Pre-sorts incoming letters: invoice, reminder, contract, job application, official mail or advertising, plus deadline, need for action and language. The input is the OCR text from your DMS or mail server. doc_type picks the folder and the responsible unit, has_deadline sets the follow-up, action_level sorts the list.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "inbox-document-v1",
    "state": "Sehr geehrte Damen und Herren,\n\nzu unserer Rechnung Nr. 2026-0871 vom 12.08.2026 über 1.845,20 € haben wir bislang keinen Zahlungseingang festgestellt. Wir bitten um Ausgleich bis spätestens 30.09.2026. Sollte die Zahlung bereits erfolgt sein, betrachten Sie dieses Schreiben als gegenstandslos.\n\nMit freundlichen Grüßen\nBuchhaltung Müller & Sohn GmbH"
  }'

Human review: everything that is other, carries a deadline or has action level 3 to 4.

Adapting it: headers and footers may stay in the OCR text. Add your own document types (delivery note, order confirmation) as options with a clear boundary; other stays at the end.

Customer feedback analysis

Id: customer-feedback-v1 · For: customer service, customer success, product management

Rates a customer statement from a survey, shop review or social media: sentiment, churn risk, main topic and whether a personal response is needed. The input is the text, optionally with an NPS value or the source as JSON. The four values go to your CRM or BI system, for example as a task for customer success or as a topic breakdown in the dashboard.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "customer-feedback-v1",
    "state": "Die Kamera macht bei Tageslicht wirklich gute Bilder, aber nachts ist alles verrauscht, da hilft auch der Nachtmodus nicht. Der Support hat freundlich geantwortet, konnte aber nichts ändern. Für 349 € hätte ich mehr erwartet. Wenn das angekündigte Firmware-Update nichts bringt, geht sie zurück."
  }'

Human review: churn risk 2 to 3 and every needs_response: true entry.

Adapting it: for sentiment and churn_risk, expected_index gives a finer value than the chosen level, useful when charting trends over months. Use sovr-decision-v2 for that, whose probabilities are more finely graded.

Lead qualification

Id: lead-qualification-v1 · For: sales, inside sales, marketing automation

Rates a sales enquiry from a web form, e-mail or trade fair note by fit, timeline, budget signal and decision-making authority. The input is the enquiry text plus form fields such as company or headcount. The CRM turns fit and timeline into a priority: a strong fit with a short horizon goes straight to sales, a weak fit without a timeframe into the nurturing track.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "lead-qualification-v1",
    "state": "Betreff: Anfrage KI-Assistent\n\nGuten Tag, ich bin IT-Leiterin eines Maschinenbauunternehmens mit rund 850 Mitarbeitenden an drei Standorten. Wir suchen bis Ende Oktober eine DSGVO-konforme KI-Lösung für unsere 120 Konstrukteure, da unser bisheriger Anbieter zum 31.12. den Dienst einstellt. Das Budget für 2027 ist freigegeben. Können wir nächste Woche telefonieren?\n\nViele Grüße\nDr. Karin Vogt"
  }'

Human review: a strong fit with timeline 3 to 4 goes straight to a person. Every enquiry stays visible; sales decides about quote and contact.

Adapting it: send form fields most simply as a JSON object with the free text as one field. Add an industry question as a choice with an other option.

Public-authority inbox

Id: behoerden-posteingang-v1 · For: public administration

Classifies e-mails and scanned letters to a municipal administration: type of matter, responsible unit, deadline relevance and threat of escalation. With these four values the document management system puts the message into the right inbox, gives deadline matters a follow-up and presents escalations to the head of office.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "behoerden-posteingang-v1",
    "state": "Sehr geehrte Damen und Herren, gegen den Bußgeldbescheid vom 3. September 2026 (Az. 0815/26, Parken im eingeschränkten Halteverbot) lege ich hiermit fristgerecht Widerspruch ein. Mein Fahrzeug stand zum genannten Zeitpunkt nachweislich in der Werkstatt, die Rechnung liegt bei. Ich bitte um Aufhebung des Bescheids. Mit freundlichen Grüßen, Katrin Vogel"
  }'

Human review: unclear stays in the mail room. Objections and escalations are never filed without review by a case worker.

Adapting it: the five units are a typical, not a complete, structure. Adapt the options to your organisation chart and list the duties per option, so that the model assigns dog tax, for example, to the treasury and not to the public order office.

Application complete?

Id: antrag-vollstaendigkeit-v1 · For: public administration · State: JSON object

The formal pre-check of an online application (housing benefit, childcare place, basic income support): complete or not, what is missing, whether the data is consistent and which step follows. naechster_schritt drives the case system: nachfordern prepares a letter with the list of documents, bearbeiten hands the application to case processing.

The state is an object with the fields antragsart, eingangsdatum, pflichtfelder (object), erforderliche_unterlagen (list of names), eingereichte_unterlagen (list with name and lesbar), unterschrift_vorhanden and anmerkung. The field names are German because the questions refer to them.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "antrag-vollstaendigkeit-v1",
    "state": {
      "antragsart": "Wohngeld (Mietzuschuss)",
      "eingangsdatum": "2026-09-18",
      "pflichtfelder": { "name": "Anna Berger", "geburtsdatum": "1987-04-12", "anschrift": "Lutherstraße 4, 07743 Jena", "haushaltsmitglieder": 2, "iban": "DE89370400440532013000" },
      "erforderliche_unterlagen": ["Mietvertrag", "Einkommensnachweis", "Mietbescheinigung"],
      "eingereichte_unterlagen": [ { "name": "Mietvertrag", "lesbar": true }, { "name": "Einkommensnachweis", "lesbar": false } ],
      "unterschrift_vorhanden": true,
      "anmerkung": "Einkommensnachweis ist ein unscharfes Handyfoto."
    }
  }'

Human review: every request letter is approved before sending. ruecksprache and unclear always land with a case worker. The engine never decides on approval of the application.

Keep the double check on here. In our tests a complete application got bearbeiten (process) or nachfordern (request documents) depending on the order of the options, both times with a probability close to 1. The double check makes this case visible and recommends review (unstable_under_reordering) instead of sending the application back for documents for no reason.

Adapting it: the field names are part of the questions. If your portal uses other names, map them before the call or write your own questions that name your fields. Do not send the object as a string: it is serialised in a fixed way so that the same application always yields the same state.

Publication release check

Id: veroeffentlichung-check-v1 · For: communications, data protection and compliance, public administration

Checks press releases, posts, newsletters or public council documents before publication for third-party personal data, Art. 9 data and tone, and derives a release recommendation. Many data protection incidents are well-meant texts that name third parties' names, addresses or health details. The editorial system can show the recommendation as a traffic light or enforce it as a block, for every draft.

The state is the text to check; a first line Kanal: … only describes the channel and is not part of the text.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "veroeffentlichung-check-v1",
    "state": "Kanal: Facebook-Seite der Stadt\n\nWas für ein Fest! Über 2.000 Gäste haben gestern mit uns den Marktplatz gefüllt. Ein besonderer Dank geht an Familie Schmidt aus der Goethestraße 12, die den Kuchenstand organisiert hat, und an die Freiwillige Feuerwehr für die Absicherung. Oberbürgermeisterin Dr. Müller eröffnete das Fest um 14 Uhr."
  }'

Human review: freigeben is a recommendation to the editors, not an approval. nicht_freigeben and unclear always go to the data protection officer or legal department. ueberarbeiten goes back to the author with the two yes/no values as a hint.

Careful with automatic mode. In our tests the engine rated individual texts as freigeben with full confidence although the preset's rule would have sent them to the editors, for example an obituary containing details about third parties. Neither the probability nor the double check warned. If you want to publish without an editor, send "policy": { "mode": "review_only" } for this preset or treat freigeben as a suggestion only.

Adapting it: the channel line is the place for context the model should know without checking it, such as the audience or whether humour fits the channel. The preset does not replace a legal review.

Agent handoff

Id: agent-handoff-v1 · For: AI agents, platform development

Decides, before an AI assistant answers, which path it takes: answer directly, web search, company knowledge (RAG), ask back or hand over to a human. This happens as a separate, measurable step before the expensive answering model runs. action picks the tool, sensitive_data: true blocks web search for this request.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "agent-handoff-v1",
    "state": "Wie viele Urlaubstage habe ich laut unserer Betriebsvereinbarung im ersten Beschäftigungsjahr? Ich bin seit März dabei."
  }'

Human review: on human, on unclear and from risk 2 upwards. There is no automatic reply in those cases.

Adapting it: instead of a single text you can send the conversation so far as messages ([{role, content}]); the handoff then decides in context. Your own tools become your own options, with the name your agent uses internally.

Grounding check

Id: grounding-check-v1 · For: AI agents, IT, quality assurance · State: JSON object

Checks after answer generation whether a RAG answer is supported by its sources, whether it adds or contradicts anything, and whether it may be delivered. You can build the result into the pipeline as a hard rule.

The state is an object with question, sources (list of texts) and answer.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "grounding-check-v1",
    "state": {
      "question": "Wie lange gilt die Garantie auf das Modell X200?",
      "sources": ["Produktblatt X200: Herstellergarantie 24 Monate ab Kaufdatum. Akkus sind von der Garantie ausgenommen."],
      "answer": "Die Garantie auf das X200 beträgt 24 Monate ab Kaufdatum; Akkus sind von der Garantie ausgenommen."
    }
  }'

Human review: release with recommendation: "act" goes out automatically. revise and unclear land in post-processing, block is never delivered. Log every abstention and every review recommendation with question and sources; that becomes your test set later.

Release only via verdict. In our tests an answer with a fabricated core claim was rated “partially supported” by support. The verdict blocked it or sent it to review, so under the rule above it did not go out. Use support, unsupported_claims and contradiction only as reasons for post-processing.

Adapting it: the check only sees what is in sources. Send exactly the passages the answering model received. Otherwise it judges a different basis than the answer.

Agent step gate

Id: agent-step-risk-v1 · For: AI agents, automation, IT security · State: JSON object

Sits between planning and acting for an agent that operates a browser or a computer. It checks the planned step: is it reversible, what kind of effect would it have, how severe would a mistake be? Reading and scrolling pass through unimpeded, while "Place binding order" or "Delete permanently" stop, even though both look the same in the page structure.

The state is an object holding the agent's goal, where it currently is and which step it intends to take next.

curl https://sovrgpt.com/api/v1/decisions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "preset": "agent-step-risk-v1",
    "state": {
      "ziel": "Order a pack of printer paper from the office supplies shop",
      "seite": "Cart — checkout, step 3 of 3, payment on invoice",
      "schritt": "Click the button 'Place binding order'"
    }
  }'

Human review: irreversible: true, the effect kinds payment, destructive, credentials and external_message, the abstention unclear and risk level 3. Such steps never run automatically, only after confirmation. The normal case is irreversible: false with effect read, input or state_change: the agent proceeds.

If your agent runs without a human who could confirm, send "policy": { "mode": "autonomous" }. The recommendation is then always act, but the warning signals remain in reasons. On a warning signal the agent should skip the step or choose a reversible path.

Careful with unclear steps. Two cases from our tests: a dialog with unreadable text was confidently classified by sovr-decision-v1 as "submit a form" rather than unclear; only margin was conspicuously low. A "Continue" button without a recognisable effect was rated read with full confidence by the default runtime, although the rule asks for unclear. Under threshold an agent would click through. For agents with real consequences (payments, messages, deletion) we recommend "policy": { "mode": "review_only" } or your own rule that lets read pass only when the description is unambiguous and also evaluates margin.

The preset picks no element. For that it would need the page's elements as options, and a choice question has at most 26. In a sample of 20 real web pages the median was 27 clickable elements. The gate does not need that list; the description of the planned step is enough.

Designing your own questions

With questions instead of preset you define your own questions. What proved useful while building the ten presets:

• The abstention is the last option. Every choice question over free text gets an unclear or other at the end. Without a permitted way out, the question forces the model into an assignment nobody wants.
• At most 26 options, at most 10 levels. Every option gets a letter, every level a digit. A better description per option improves discrimination.
• Levels are ordered, lowest first. Only then do expected_index and normalized_score make sense.
• The state is data. Everything the model is supposed to do goes into instructions. An “ignore all rules” inside the state stays text.
• Order matters. Options are evaluated in your order, and a model may react to it. For choice questions the double check covers this (stability_check, on by default). Levels stay in your order; there the order is the meaning.
• Keep instructions short and decisive. Say what the decision is based on (“the main purpose, not subordinate clauses”) and when the abstention is right.
• v2 for graded probabilities. sovr-decision-v1 mostly answers with 1.0 or 0. For thresholds on max_probability or expected_index, use sovr-decision-v2.

The limits (questions, characters, images) and the meaning of the answer fields are in the API reference.

Trying it in the product

• Playground: logged in, via the sidebar → Decision Engine. Every preset runs there with its example state or your own text, with the full answer.
• In chat: the Entscheiden (decide) pill in the composer enables the decision_engine tool. That is how you try a preset on real texts.
• Via MCP: sovrgpt_decide and sovrgpt_list_decision_presets through the authenticated MCP server, scope decisions.

All paths run the same engine with the same presets.