How decision models compare in the contact centre

From one decision model to a comparison

Part 1 showed what a decision model is: a model that makes one quick call, such as which team takes a request or whether a person must step in, working as the fast System One beside a language model's slower reasoning. Part 2 showed how we tested one. The next question from every buyer was the obvious one: which decision model, and is it good enough? Since then OpenAI, Microsoft, Perplexity and Cloudflare have released or listed their own. So we built a bigger test: eight decision-model configurations against a general-purpose model at two effort settings, including the false alarms and manipulated routing Part 2 said were still to come.

The question matters wherever a request has to reach the right team quickly: in the contact centre, across digital channels such as chat and messaging, and on the service desks that support staff. In large enterprises, heads of AI, customer experience (CX), digital and the contact centre often share these decisions, so we report the results in their terms: the right queue, the right handoff and the right priority.

At a glance

The models we tested

We tested ten model configurations from nine models. The eight decision-model configurations include Jev 1.13.0 and Jev 1.14.0; Claude Haiku 5.5 supplies the general-purpose comparison at two effort settings.

For the general-purpose comparison we used Claude Haiku 5.5, because the GPT and Nova models from Parts 1 and 2 were building our answer key this time, and a model shouldn't mark its own work.

Basis: Product capabilities checked through 25 October 2026; image scores use supported inputs only. TypeSafe says Jev's opt-in image input is not yet ready for production.

Results at a glance

The eight decision-model configurations matched 91% to 98% of routes and 87% to 98% of handoffs on new text cases. We built the harness, ran every configuration five times per scenario, and logged and scored every call. The result is a practical shortlist for contact centre routing and handoff.

Basis: Accepted-route and exact handoff matches from five runs per scenario. Source: Humint Labs test, 25 October 2026. Method and full record in Part 5.

Basis: Five runs per scenario. Route uses 445 answers (89 scenarios) and includes allowed alternatives; handoff uses 485 (97 scenarios) and requires the right yes/no answer. Costs are model calls. p95 comes from the 15 October 2026 timing run: one run of 68 text scenarios, successful calls only, including 66 for Claude higher effort and 64 for low effort. 95 in 100 calls were this fast or faster. Jev 1.14.0 was not separately timed.

Perplexity Decider had the highest observed text scores and lowest cost, US$0.016 per 1,000 calls. If text routing and call cost drive your choice, start there, then check processing location and terms.

Claude Haiku 5.5 at low effort cost less than higher effort, but matched fewer routes and handoffs. Test the setting you plan to operate.

How fast they answered

OpenAI Decisions had the fastest text p95 at 322 ms, followed by Jev 1.13.0 at 370 ms. Both are strong choices for a quick decision step. In the contact centre the decision is one step in a longer path, so time speech recognition, retrieval, rules and transfer as part of the complete service path.

Urgency is the softer result

Urgency was less reliable on the half-level measure used in Part 1. Jev 1.13.0 was within half a level on 59% of new answers, compared with 50 of 60 on the earlier cases. It was within one level on every new answer, but that broader check can hide a queue-changing error. Across the eight decision-model configurations, within-half results ran from 55% to 72%; Part 5 has the full urgency table. Test the exact line that changes your service action as well as the average error.

Why these numbers differ from Part 1

Part 1 asked one or two questions about shorter scenarios, about 412 input tokens per call. These new text calls ask three questions over longer prompts, about 1,107 input tokens. The token price did not rise: each new call does more work. On the original 46 scenarios, the cost per call stayed in line with Part 1. Part 5 sets out the arithmetic.

We timed every call as a full round trip from Australia, from sending the request to receiving the answer, and logged each one in Langfuse. The response times come from our 15 October 2026 run over the same scenarios and test harness, and the accuracy and cost figures from the full five-repeat study on 25 October. The 25 October run had a network delay that added about 0.6 seconds per call, so we use the clean 15 October timings. On text, Jev 1.13.0 answered in a median of 289 ms, and its 95th-percentile time was about ten times faster than Claude Haiku 5.5 at its higher effort setting.

What the answers looked like

Here are three public synthetic messages and every recorded answer: agreement, a route split and an image handoff boundary. The angry bill is a customer experience case. The booking outage and the VPN request come from staff, the employee experience (EX) side of the same service, where a service desk keeps reception and remote workers running.

In each table, ×5 means the same displayed value on all five runs, and a count such as ×3 shows how many runs gave that value. Urgency shows one decimal and probabilities two; decisions use the unrounded records, so a displayed 0.50 can sit below the decision line.

Where each one fits

Use the results to choose candidates for a trial on your own contact centre and digital service rules. The same model can fit a text queue and create too much review work on images.

Basis: Observed fit in this synthetic test, with product information checked through 25 October 2026. The two Claude settings are reported separately above.

Then make three checks before you choose:

• Images: can it accept the image, then make the right handoff?
• Cost and speed: compare the full call mix and time the complete route.
• Processing location and terms: confirm the actual deployment and contract.

Operating pattern the results support

• Define the action: agree what changes the queue, review path or priority before choosing a model.
• Keep review available: test a proposed review band around the handoff line on your own cases. A band such as 0.45 to 0.55 is illustrative, not a measured safety guarantee. Jev 1.14.0 crossed 0.5 on two of five runs of the VPN case.
• Check the original input: show the message and image beside the suggestion so a contact centre agent or reviewer can correct a wrong route or urgency.
• Repeat when things change: keep the test cases for a model-version or policy change, and time the complete service path.

Next steps

Choose one contact centre or digital service decision and test a shortlist on your own agreed cases, including the requests close to its handoff or priority line. Contact Humint Labs to build and run that test. Part 4 shows where to put the result to work.