Niches · Decision models

Route a message or request to a queue, team, intent or model tier

Four or more companies' docs show a full fit.

5companies, full fit
18companies, fit with a limit
30companies carry the tag
22share of the category's work
12first-hand accounts

What counts as done

One label from the caller's list with a probability per option, in one call, with low-probability answers routable to a human by threshold.

Measured, for our own tests: under 500 ms hosted or under 1 s local. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: which team should handle this ticket, is this billing, shipping, technical or spam, which model tier does this prompt need, what is the intent of this message

Does the job 6

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Decision 3.0 B
vLLM Semantic Router project
choice with a probability per option on /v1/systemone through vLLM, Engine mode or d3_server.py; 54 ms median text on d3 and 18 ms on d3-mini (MI325X). First on the Decision Index text board at 66.3. Languages not stated. source
Liquid d1 C
Liquid AI, Inc.
POST /decisions/v1/systemone returns a choice, a probability per option and confidence; $0.02 per million input tokens with $2 free credit. Arize measured p50 0.20 s and Dataiku about 86 percent on routing. d1-2B's card lists 16 languages. source
Microsoft-Decision-1 E
Microsoft
choice with a probability per option on Foundry's /providers/microsoft/v1/systemone or OpenRouter's Decisions API at $0.042 per million input tokens; OpenRouter shows 0.42 s latency and Dataiku about 86 percent. Microsoft's perturbation test found zero flips on reversed or shuffled options. source
OpenAI Decisions API A
OpenAI OpCo, LLC
choice with probabilities and confidence, a documented 'other' fallback, and answers that can come back as refusal; Arize measured p50 0.16 s and Dataiku 91 percent, the best in its test. GPT-6 Luna needs the Build tier ($5 of credit). source
pplx-decider BB
Perplexity AI, Inc.
choice with 1 to 255 options returns a probability per option and confidence on /v1/decisions in Jev's state-and-questions shape, with an 'other: null' option in the example; $0.02 per million, 10 requests a second per organisation by default. Dataiku about 86 percent. Card lists no languages. source
pplx-decider-flash B
Perplexity AI, Inc.
Same /v1/decisions shape at $0.01 per million with one question answered on its own; Decider for Mac runs it at 127.0.0.1 with no key for private text. Decision Index 54.81, 10.74 behind the 27B. Card lists no languages. source

Does the job, with a limit 20

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Celeris-1 Decision D
Marqo Inc
Choice answers carry a probability per option and a rescaled confidence, up to 512 options, on /v1/systemone or /v1/decisions. Workspace activation still runs a capacity waitlist and credit needs a card with a $5 minimum. No latency figure is published for the decision model. source
Clef B
Cloudflare, Inc.
choice with a probability per option, 1 to 64 questions a call, 10,000 free neurons a day. Vendor median 209 ms (Clef) and 39 ms (Clef-flash); Arize measured p50 0.50 s on RAGTruth-length inputs. Dataiku had Clef about 86 percent, Clef-flash 77. Languages unstated. source
Laya BB
Convai Innovations Pvt. Ltd.
laya-serve answers /v1/systemone in 33 ms on a T4 with a Router across 100+ languages, but the base checkpoints need fine-tuning: 0.425 on Banking77, weak past about 20 options and 37 percent on Hebrew intent (yuv.ai), Portuguese fell to 0.500 after 0.4.0 (issue 1453). source
Decider BB
Mark Marosi (Mapika)
/v1/systemone with 2 to 255 options and a probability per option, 50 ms median on CUDA and 133 ms on an M1 Pro. English only by the README's limits; issue 9's Chinese cases failed, and the 2B needs bucket-like option names. source
Drex Max v1 C
Nace AI, Inc.
choice returns a probability per option through /v1/systemone on the Python loader, Kapitsa, MLX or Nace's llama.cpp fork. The Open RAIL-M licence bars commercial use above $1 million in revenue or funding, and the card lists no languages. source
elephant-1 D
Baseten Labs, Inc.
choice with a probability per option from one shared-prefix prefill, hosted on Baseten's /v1/systemone at $0.035 per million input tokens or locally through qev's DecisionEngine. The card's language is English only, and self-hosting needs OneJev at a pinned commit plus serve.patch. source
GLiClass C
Knowledgator Engineering Ltd.
One forward pass scores every label with a softmax (classification_type single-label) up to 128 labels, locally on CPU or GPU. No Jev shape (issue 47 asks for one), F1 only on the cards, and issue 12 reports every label scoring 0.99. source
Granite Switch 4.1 B
International Business Machines Corporation
The decision aLoRA returns a softmax over single-token option labels, read through vLLM chat completions with logprob_token_ids and max_completion_tokens 1; the caller writes the prompt, tokenises labels and chunks past 128 options. Decision Index 46.72 on the 8B (IBM's run). source
Kev BB
Jared Palmer
/v1/systemone with 1 to 255 options and a probability per option; Kev-4B answers six questions in 41.5 ms on an L40S and Arize had Kev 9B at 48 of 52 on routing. English only by the card, so non-English tickets are a blocker. source
JEV-27B-VL C
AutoTrust AI Lab
POST /v1/decide takes one question a request with 2 to 256 options and returns a probability per option, about 264 ms a row on an RTX PRO 6000; CLINC150 with 150 intents zero-shot is on the card. Needs an 80 GB GPU and the card's vLLM development build. source
joey-1 C
Baseten Labs, Inc.
choice with a probability per option from one shared-prefix prefill, hosted on Baseten's /v1/systemone at $0.01 per million input tokens or locally through qev's DecisionEngine. The card's language is English only, and self-hosting needs OneJev at a pinned commit plus serve.patch. source
mammoth-1 D
Baseten
choice with a probability per option from one shared-prefix prefill, hosted on Baseten's /v1/systemone at $0.10 per million input tokens or locally through qev's DecisionEngine. The card's language is English only, and self-hosting needs OneJev at a pinned commit plus serve.patch. source
Drex 1.5 BB
Nace AI, Inc.
choice returns a probability per option and confidence on /v1/systemone, but drex-v1.5 is a retired id: the OpenAPI lists it among ids answered by drex-max-v1.1 at Max's limits and $0.03 per million. The $5 signup credit needs a Google or GitHub sign-in or a card. No latency figure published. source
RightNow System Zero B
RightNow AI, Inc.
boolean, choice and score answers carry a decision, confidence and ranked alternatives, but every answer says calibrated: false in this release, latency headers come with no guarantee, and starting credit has its own rate and spend limits. source
ScaleDown E
Carbon ScaleDown, Inc.
POST /v1/scaledown mirrors the Jev shape with a probability per option summing to 1, but /classify caps a request at 20 labels (16 multi-label), the key comes from sales, and no latency is published. source
SPEV D
Sprag AI, Inc.
choice with 2 to 255 options returns a probability per option and confidence on an OpenAI-style /v1/decisions ($0.09 per million text tokens), with 60 requests a minute per organisation by default, a 402 without a payment method, and no published latency. Card says multilingual without a language list. source
kettei C
Sqwish Ltd, trading as Sqwish Labs
POST /v1/decide (or the Jev shape on /v1/systemone) returns a probability per outcome for up to 128 decisions about one context, with a 'none' outcome and an abstain_below policy worked through in the ticket-routing example. New accounts may start on a waitlist, and no latency figure is published. source
Strands Decider 2B B
Amazon Web Services, Inc.
/v1/systemone locally at 115 ms median on an RTX 3090 and 153 ms on an M3 Pro, but AWS Builder found German recall 16 percent, a 55 percent first-answer preference flipping on swap, routing AUC 0.51, and Dataiku had it at 65 percent against 86 for the leaders. source
Jev B
TypeSafe AI, Inc.
choice up to 255 options with probabilities and confidence, a companion-noul pattern for none-of-these and 0.7/0.3 threshold guidance; OpenRouter's test had 59 of 60 intents at 0.194 s p50 and Arize p50 0.08 s. Early access runs behind a waitlist with no free tier found. source
Vela 2.0 BB
vLLM Semantic Router project (with KR Labs)
choice over 2 to 255 options with a distribution through vela2_serve.py's /v1/systemone; MASSIVE unseen-scenario intent 0.727 on the 4B and a Jev Decision Index of 31.63, below its Decision 2.0 base. The 0.3B runs on CPU. source

1 listings in the category were added after this reading and aren't marked yet.

What stops agents finishing

  • waitlist or no free tier for a hosted model
  • label set over the model's option cap (Laya about 20, Jev 255)
  • non-English text with an English-only checkpoint
  • hosted endpoint behind a bot check that refuses the SDK's default user agent

Around it

Before: email.read, support.fetch-ticket, chat.receive-message. After: inference.llm.draft-reply, support.assign-queue, human-review.

Where this comes from

Weight 22 of 100 in Decision models: 12 of 41 accounts; the most common job in the HN threads, the AWS Builder posts and the Dataiku and yuv.ai tests

The same as data: /api/v1/use-cases/decision-models.route-to-queue.json, or the MCP tool get_use_case with id decision-models.route-to-queue. Other jobs in this category: /api/v1/niches?category=decision-models.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.