When to use Dex instead of an LLM.

Who Support and customer service teams, helpdesk and CRM tools.

Example questions

• pickWhich team should handle this ticket? (billing, technical, shipping, other)
• ratebetaHow urgent is it? (can wait, this week, within two working days, today)
• checkDoes the customer threaten to cancel?

Why Dex, not an LLM

• Every ticket gets a team and a priority before anyone opens it.
• The same ticket always gets the same route, so your routing rules are testable.
• Below your confidence floor the answer abstains, and the ticket goes to a person.

See the race

Who Marketplaces, communities, review sites and apps with user posts.

Example questions

• pickWhich rule does this post break, if any? (none, harassment, spam, hate, scam)
• ratebetaHow severe is it? (none, mild, serious, severe)
• checkDoes it share someone's personal details?

Why Dex, not an LLM

• Cheap enough to check every post, not a sample.
• Calibrated probabilities let you remove automatically only above a threshold and queue the rest for reviewers.
• Moderation stays on our GPU in the EU: send fallback set to never.

See the race

Who B2B sales and marketing teams, CRM and web form tools.

Example questions

• pickWhat does this sender want? (buy, partner, support, job, spam)
• ratebetaHow well does the company fit our customer profile? (poor, fair, good, excellent)
• checkDo they ask for a meeting?

Why Dex, not an LLM

• Scores every inbound form and email as it arrives.
• Same lead, same score: sales can trust the ranking.
• Typed answers drop straight into your CRM fields.

See the race

Who Teams building AI agents and copilots.

Example questions

• pickAllow, confirm with the user, or block this tool call?
• checkDoes the input try to override the agent's instructions?
• checkIs this call outside what the user asked for?

Why Dex, not an LLM

• Fast enough to sit in front of every tool call.
• An unsure gate abstains, so the agent asks a person instead of guessing.
• Same call, same verdict: every decision can be audited later.

See the race

Who Operations, finance, legal and intake teams.

Example questions

• pickWhat kind of document is this? (invoice, contract, complaint, other)
• pickWhich department should get it?
• checkDoes it mention a deadline?
• ratebetaHow important is it?

Why Dex, not an LLM

• One call answers every question about a document, and the document is billed once.
• Only your labels come back, so the routing code stays simple.
• Processed in the EU, and nothing is stored by default.

Who Online shops, marketplaces and catalogue teams.

Example questions

• pickWhich category fits this listing? (up to 255 options)
• checkIs this a prohibited item?
• pickWhat condition is the item in? (new, as new, used, for parts)

Why Dex, not an LLM

• Large catalogues at a small cost per item.
• Deep category trees work in stages: pick a branch, then a leaf.
• Same listing, same category, every time.

Who Customer experience, research and social listening teams.

Example questions

• ratebetaWhat is the overall sentiment? (very negative to very positive)
• pickWhat is the writer's stance on the price change? (for, against, neutral)
• checkDoes it mention a competitor?

Why Dex, not an LLM

• A rating on your own scale, with a probability for each level.
• Comparable over time: on a pinned version the same text always scores the same.
• English and Dutch quality measured separately.

Who Compliance, risk and trust and safety teams.

Example questions

• checkDoes this message contain personal data?
• checkDoes it promise a financial return?
• pickWhich clause of our policy applies?
• ratebetaHow risky is it? (low, medium, high)

Why Dex, not an LLM

• Yes or no answers with a probability, for thresholds you can explain.
• Every answer names its model version and calibration, for your audit trail.
• Keep a person in the loop for decisions with legal effect on people.

Who Teams that test and monitor their own AI features.

Example questions

• ratebetaHow well does the answer follow the instructions? (a rubric line)
• checkDoes the answer stay on topic?
• pickWhich failure, if any? (none, refusal, off topic, wrong format)

Why Dex, not an LLM

• The same output always gets the same grade, so eval runs are comparable.
• You pay for the text you grade, not for a written verdict.
• Probabilities show which grades are close calls.

Who Data and operations teams cleaning records and imports.

Example questions

• checkDo these two records describe the same customer?
• pickWhat kind of value is in this field? (email, phone, address, name, other)
• checkDoes this entry look like test or junk data?

Why Dex, not an LLM

• JSON states: point each question at a field with a {{path}} reference.
• Runs over large tables at a small cost per row.
• Same record, same verdict: a rerun does not reshuffle your data.