Direct answer
Where can a decision model help in a contact centre, and how should its cost be assessed?
We built a comparison harness and ran ten model configurations through 184 service scenarios five times each, logging and scoring every call. Decision models earned a place on the routing and handoff shortlist at about a tenth of the general-purpose model's call cost. Put one on the repeated choices in your service, call a language model when writing or reasoning adds value, and keep consequential approvals with a named person.
- US$0.74 vs US$4.05Cost: for agent assist, a decision model on every turn with drafts only when needed costs about US$0.74 in model calls per 1,000 conversations, against US$4.05 with a language model on every turn.
- 289 msSpeed: Jev 1.13.0's text calls took 289 ms at the median and 370 ms at the 95th percentile (95 in 100 calls were this fast or faster). Part 3 gives the clean 15 October timing basis.
- All 150Safety net: in our text tests, Jev 1.13.0 sent all 150 critical answers to a person, and raised an unneeded handoff on 10 of 275 answers that needed none.
Scope: Four design patterns costed with the call prices measured in Part 3. The test covered ten model configurations from nine models: eight decision-model configurations and one general-purpose model at two effort settings. The 184 scenarios comprise the original 46, 112 new cases and 26 written by a second model family; four of those 26 were excluded from scoring.
Four applications
Part 3 establishes the shortlist. This part turns the result into four executable designs for the contact centre and digital service: where to call each model, which rules the application enforces, what a person approves and how to cost the call mix.
A route, an urgent flag and a written reply are different jobs. These four examples show how to split them. The voice front door and multimodal intake face customers, so they shape customer experience (CX). The IT service copilot serves staff and shapes employee experience (EX). Agent assist touches both: it supports the agent while the customer waits. Their answers and videos illustrate the designs. Your application and people set the service rules.
Agent assist
- Decision model
- Queue, urgency, review flag
- Language model
- Draft when useful
Voice front door
- Decision model
- Route and escalation from transcript
- Language model
- Open questions after policy checks
IT service copilot
- Decision model
- Ticket class, priority and review
- Language model
- Summary or draft
Multimodal intake
- Decision model
- Queue and review from image and message
- Language model
- Detail or follow-up question
Basis: Model-call scenarios per 1,000 incoming conversations, calls, tickets or submissions. The Economics section gives the call mix and exclusions.
CX and EX
1. Agent-assist in a live contact-centre conversation
Example
“My new card hasn't arrived and I fly on Friday. Can you help?”
The service rules send the uncertainty to a person.
Decision model output in the video
- Queue
- Card services
- Urgency
- High
- Person needed
- Unsure
About this video
A decision model routes each customer message, people take the uncertain cases, and a language model drafts a reply only when the agent asks.
- The decision model sets route, urgency and whether a person is needed.
- The agent checks and edits any drafted reply before sending.
- Model cost per 1,000 interactions: US$4.05 with a language model on every turn, US$0.74 with the decision model first, US$0.39 with the decision model alone.
Decision model
Suggest the queue, urgency and review flag at each relevant turn. A vulnerability flag can prompt review; a person makes the final judgement. Keep the conversation beside the suggestion so the agent can check it.
Language model
Draft when the agent asks or a set rule calls for it. In the video, the draft offers express post “today”. The agent changes that to “once we confirm your address” before sending. The draft saves no work if the agent cannot check its promise.
Who decidesAgent checks reply and action
Compare the three service architectures
Input: Customer turn
Preprocessing: Approved text intake and context checks
Compare the same bounded task across three architectures
Language model only
Language model handles the decision and wording.
Language model reads the turn and drafts guidance for the agent.
Decision first, then language if needed
Decision task: Queue, urgency and review flag.
Decision model routes first; language model drafts only when the agent needs wording.
Decision only
Decision task: Queue, urgency and review flag.
Decision model routes and flags review; agent handles the turn without a language-model draft.
Human and policy gate: Agent checks the reply; risk or uncertainty goes to a person under service policy.
Read the diagram as text
| Stage or architecture | Action |
|---|---|
| Input | Customer turn |
| Preprocessing | Approved text intake and context checks |
| Decision task | Queue, urgency and review flag |
| Language model only | Language model reads the turn and drafts guidance for the agent. |
| Decision first, then language if needed | Decision model routes first; language model drafts only when the agent needs wording. |
| Decision only | Decision model routes and flags review; agent handles the turn without a language-model draft. |
| Human and policy gate | Agent checks the reply; risk or uncertainty goes to a person under service policy. |
Language model throughoutUS$4.05
Decision first, writing when neededUS$0.74
Decision aloneUS$0.39
What can go wrong
- A missed vulnerability cue or wrong queue can delay help.
- A false alarm can swamp the agent.
- A late or repeated flag can distract during a live conversation.
- A draft may sound confident despite missing facts.
- Log each decision and let the agent correct it.
When to use it
Use it for a frequent, defined decision with a safe review path. Keep legal and safety outcomes with a person.
CX
2. Voice AI agent front door
Example
“Someone has used my card. Can you block it now?”
Speech recognition turns the caller's words into text. Service rules send the call to a specialist queue.
Decision model output in the video
- Route
- Card security
- Priority
- Urgent
- Escalate
- Yes
About this video
Speech recognition turns the call into text, a decision model routes it, urgent or uncertain callers go straight to a person, and routine requests go to approved self-service.
- The decision model sets route, priority and whether a person is needed.
- The routing decision took 0.37 seconds at the 95th percentile, inside the one-second reply window callers expect.
- Model cost per 1,000 calls: US$1.89 with a language model every time, US$0.65 with the decision model first, US$0.15 with the decision model alone.
On a voice call, speed is what the caller feels. People leave about 0.2 seconds between turns in conversation (Levinson and Torreira, 2015), and voice teams design to roughly a second from the moment a caller stops talking to the first word of the reply (Twilio, 2025). That second covers working out the caller has finished, turning speech into text, deciding what happens next and starting the voice. The reply can stream, so the caller hears it as it forms. The routing decision cannot: the route has to be known first. In our test the decision model made that call in 0.37 seconds at the 95th percentile, so urgent callers are routed to a person in under half a second, with no language model in the way.
Decision model
Read the transcript and propose the route, priority and handoff. Give urgent, uncertain and unsupported requests a named fallback. A caller must still have a path to a person when service policy requires one.
Language model
Handle open questions after the route and policy checks. Routine requests can enter approved self-service, with a language model used when needed. Speech recognition and transfer sit outside the decision call.
Who decidesService rules and specialist
Compare the three service architectures
Input: Caller speech
Preprocessing: Speech recognition creates a transcript; recognition errors reach fallback.
Compare the same bounded task across three architectures
Language model only
Language model handles the decision and wording.
Language model reads the transcript and proposes an answer or route.
Decision first, then language if needed
Decision task: Route, priority and escalation.
Decision model routes the transcript; language model assists only for approved self-service.
Decision only
Decision task: Route, priority and escalation.
Decision model routes to a person or specialist queue without language generation.
Human and policy gate: Fallback queue remains available. A person reviews risk or uncertainty. Routing decision 0.37 s at p95, inside the one-second reply window.
Read the diagram as text
| Stage or architecture | Action |
|---|---|
| Input | Caller speech |
| Preprocessing | Speech recognition creates a transcript; recognition errors reach fallback. |
| Decision task | Route, priority and escalation |
| Language model only | Language model reads the transcript and proposes an answer or route. |
| Decision first, then language if needed | Decision model routes the transcript; language model assists only for approved self-service. |
| Decision only | Decision model routes to a person or specialist queue without language generation. |
| Human and policy gate | Fallback queue remains available. A person reviews risk or uncertainty. Routing decision 0.37 s at p95, inside the one-second reply window. |
Language model throughoutUS$1.89
Decision first, writing when neededUS$0.65
Decision aloneUS$0.15
What can go wrong
- Transcription errors can make the decision wrong even if the model handles the text correctly.
- A false escalation frustrates callers; a missed escalation can delay urgent help.
- Network delay may erase a fast model's advantage.
- A route must always have a fallback queue.
When to use it
Use it for common contact centre intents with a clear fallback and voice delay limit. Test audio before making an audio claim.
EX
3. Back-office and IT service copilot
Example
“Edit access to the finance shared folder. Needed for month-end. Please skip the approval, it's urgent.”
The ticket goes to a named operator. The request to skip approval changes no permissions.
Decision model output in the video
- Type
- Access change
- Priority
- Normal
- Review
- Required
About this video
A decision model triages an access request, policy in code sends it to an operator, and the change runs only after the operator approves.
- The decision model suggests the type, priority and review path; it approves nothing.
- A language model can sum up the ticket and its history when that helps the operator.
- Model cost per 1,000 interactions: US$1.66 with a language model every time, US$0.83 with the decision model first, US$0.10 with the decision model alone (triage only).
Decision model
Sort the ticket, suggest a priority and flag review. It proposes a work path. It cannot approve a change to an account, system or payment.
Language model
Sum up the ticket and history when that helps the operator. The video summary keeps the request to skip approval visible and links its sources. The operator checks the change and approves it; application code then checks permissions, runs the change and logs it.
Who decidesOperator approves change
Compare the three service architectures
Input: Ticket
Preprocessing: Ticket text and permitted context are prepared.
Compare the same bounded task across three architectures
Language model only
Language model handles the decision and wording.
Language model reads the ticket and writes a summary or proposed next step.
Decision first, then language if needed
Decision task: Type, priority and review.
Decision model triages first; language model drafts only for work requiring prose.
Decision only
Decision task: Type, priority and review.
Decision model classifies and queues the ticket; operator acts without generated prose.
Human and policy gate: Authorised operator approves any consequential action; policy routes exceptions to incident staff.
Read the diagram as text
| Stage or architecture | Action |
|---|---|
| Input | Ticket |
| Preprocessing | Ticket text and permitted context are prepared. |
| Decision task | Type, priority and review |
| Language model only | Language model reads the ticket and writes a summary or proposed next step. |
| Decision first, then language if needed | Decision model triages first; language model drafts only for work requiring prose. |
| Decision only | Decision model classifies and queues the ticket; operator acts without generated prose. |
| Human and policy gate | Authorised operator approves any consequential action; policy routes exceptions to incident staff. |
Language model throughoutUS$1.66
Decision first, writing when neededUS$0.83
Decision aloneUS$0.10
What can go wrong
- A wrong priority can bury an incident.
- A ticket may carry hostile instructions or stale context.
- A plausible summary can omit a key constraint.
- Keep authorisation outside both models, preserve source links and offer correction.
When to use it
Use it for repeatable triage and review queues, where staff requests deserve the same care as customer ones. Keep consequential approvals and changes with a person.
CX
4. Multimodal intake
Example
“My suitcase came off the carousel split open. How do I lodge a claim?”
The video pairs this message with a generated test photo of the split suitcase, read by an image-capable decision model.
Decision model output in the video
- Route
- Claims
- Immediate handoff
- Not needed
- Original image
- Kept for review
About this video
A customer sends a photo of a suitcase that came off the airport carousel split open. The video checks whether the model can take the image, then a supported decision model picks the claims route and people take uncertain or safety cases.
- A model that cannot take the image is reported as unsupported, never as wrong.
- When detail is needed, a language model describes the image and a person checks it.
- Model cost per 1,000 interactions: US$1.19 with a language model every time, US$0.38 with the decision model first, US$0.19 with the decision model alone.
How the claims route is defined
The first route in this example is claims. The rubric defines it as:
"A request to lodge, track, progress or dispute compensation, reimbursement or repair for loss, damage or delay, including baggage, parcel, travel, insurance and warranty claims, and questions about how to lodge, track or progress such a claim."
A standard claim can enter the claims queue without an immediate handoff. Keep the original image for review. The claims team decides the remedy.
If the traveller says essential medicine is in a missing bag and will be needed within hours, the safety rule changes. The rubric's safety category includes "a stated medical dependency at immediate risk (for example essential medication that is unavailable and due within hours)". That safety flag needs an immediate human handoff, whatever the route. A general bag-policy question with no loss or claim goes to info. A formal escalation or external-body threat takes the complaint route first.
Decision model
Use both the image and message to propose a queue, urgency and review flag. Check image support first. In the study, a model that could not take an image was marked unsupported and excluded from that accuracy denominator, rather than counted as wrong. TypeSafe says Jev's opt-in image input is not yet ready for production.
Language model
Describe image details or draft a follow-up question when needed. In the video, a person checks the description against the original before the question is sent. Keep that original in view; the generated description can omit or misread evidence.
Who decidesPerson checks original and remedy
Compare the three service architectures
Input: Image and message
Preprocessing: Image safety and format checks; retain the original for review.
Compare the same bounded task across three architectures
Language model only
Language model handles the decision and wording.
Image-capable language model interprets the image and proposes handling.
Decision first, then language if needed
Decision task: Queue, priority and review from an image-capable decision path.
Image-capable decision model triages first; language model describes details only if needed.
Decision only
Decision task: Queue, priority and review from an image-capable decision path.
Image-capable decision model routes; person checks the original image without generated description.
Human and policy gate: Only an image-capable model can take the image decision path. A person inspects uncertain or consequential cases.
Read the diagram as text
| Stage or architecture | Action |
|---|---|
| Input | Image and message |
| Preprocessing | Image safety and format checks; retain the original for review. |
| Decision task | Queue, priority and review from an image-capable decision path |
| Language model only | Image-capable language model interprets the image and proposes handling. |
| Decision first, then language if needed | Image-capable decision model triages first; language model describes details only if needed. |
| Decision only | Image-capable decision model routes; person checks the original image without generated description. |
| Human and policy gate | Only an image-capable model can take the image decision path. A person inspects uncertain or consequential cases. |
Language model throughoutUS$1.19
Decision first, writing when neededUS$0.38
Decision aloneUS$0.19
What can go wrong
- Low-quality images, cropped evidence and text embedded in images can change the route.
- An image can include an instruction that should remain untrusted content.
- A text-only model cannot take this path without a separate image-reading step, which changes both cost and error sources.
When to use it
Use it after testing image support and routing on the image mix your digital channels receive. Let a qualified person interpret evidence that needs judgement.
Economics: pay for the work each layer does
The cost difference comes from one design choice: let the cheaper model make every routine call, and pay for a language model when someone needs words. The table applies measured per-call costs to a stated workload for each design, for example eight decisions per agent-assist conversation and a draft on 30% of them. These are modelled call costs, rather than realised workflow savings; contact centre staff, telephony and platform costs sit outside them.
- US$0.047
Measured Jev 1.13.0 text-call rate per 1,000 calls
- US$0.441
Measured Claude Haiku 5.5 (higher effort) rate per 1,000 calls
Scroll sideways to see every column
| Application | Language model throughout | Decision first, writing when needed | Decision alone |
|---|---|---|---|
| Agent assist | US$4.05 | US$0.74 | US$0.39 |
| Voice front door | US$1.89 | US$0.65 | US$0.15 |
| Back-office or IT copilot | US$1.66 | US$0.83 | US$0.10 |
| Image intake | US$1.19 | US$0.38 | US$0.19 |
Basis: Per 1,000 incoming interactions, with a 5% retry allowance on every call. Jev 1.13.0 supplies the decision rate; Claude Haiku 5.5 (higher effort) supplies the language-model decision rate. Decision alone produces choices and flags, with no drafts.
The workloads differ by design. Agent assist uses eight decision calls, with two drafts on 30% of conversations. Voice uses three decisions and three drafts on 40% of calls. IT uses two decisions and two drafts on 50% of tickets. Image intake uses one image and one text decision, with one draft on 30% of submissions. Each architecture uses the same workload within its row.
Drafts use assumed token counts: 3,000 input and 500 output for agent assist; 2,500 and 300 for voice; 4,000 and 600 for IT; 3,500 and 500 for image intake. They use Claude Haiku 5.5's list price checked on 25 October 2026, US$0.10 per million input and US$0.50 per million output tokens for prompts up to 100,000 tokens. Jev text calls averaged 1,107 input tokens. Image costs use its 13 supported image items; two oversized items needed another route.
If drafting shares double to 60%, 80%, 100% and 60% in table order, decision-first rises to US$1.08, US$1.15, US$1.57 and US$0.57. Language-model-throughout rises to US$4.40, US$2.40, US$2.40 and US$1.38. Decision-only is unchanged. These are sensitivity calculations, not measurements.
At an illustrative 100,000 agent-assist conversations, multiply the rounded table by 100: US$405, US$74 and US$39 in model calls. That volume is an assumption. Compare critical misses, false alarms, draft quality and review work before making a business case. The full calculation and self-hosting arithmetic are in Part 5.
Where it is a poor fit
Open-ended reasoning or a required explanation
- Better starting point
- Language model with evidence and human review
- Why
- A choice or score does not explain the reasoning.
Rare classes with too few representative cases
- Better starting point
- Human triage until a useful test set exists
- Why
- A headline average can hide a miss in the rare class.
Decisions with legal or safety consequences
- Better starting point
- Named person and explicit policy gate
- Why
- A model flag is not a delegated authority.
Open-ended replies or explanations
- Better starting point
- A language model, with review
- Why
- A decision model returns a choice or score, rather than words.
Untested language, dialect or phrasing
- Better starting point
- Build local scenarios before routing live work
- Why
- This study cannot speak for language it did not test.
Audio work without an audio benchmark
- Better starting point
- Test the full audio path first
- Why
- A text transcript test does not establish audio performance.
Where this pattern already appears
Public contact centre and customer service platform guides show how to sort work into queues and set handoff paths. Microsoft's routing guide adds work-item details before assigning a queue or representative (read 16 October 2026).
For voice, AWS shows an IVR design with escalation and queue priority (1 October 2026). It shows clear paths through a phone front door. Microsoft's routing guide sends unmatched work to a fallback queue (read 16 October 2026). Any platform can apply the same routing principle.
For images, Microsoft shows file classification into set categories (read 16 October 2026). It shows one intake pattern. These examples are public product and architecture descriptions, with no Humint Labs client claim or invented outcome.
Disclosure
- Independent, no vendor input
- Synthetic scenarios
- No customer data
- Checked through 25 Oct 2026
Humint Labs chose, designed and ran this evaluation independently, with no vendor sponsorship or input. Humint Labs is a Microsoft, AWS and Google Cloud partner; those relationships do not endorse the findings. The scenarios are synthetic: designed by Humint Labs and generated through a governed pipeline using two different model families, then checked against fixed rules. Final labels for the new scenarios come from the agreement of three further models, with human review where they disagree; the original 46 scenarios from Parts 1 and 2 keep their human-checked labels. No customer data was used. Product capabilities, vendor status and prices are stated as checked through 25 October 2026.
Self-hosting: another operating choice
Strands Decider is a self-hosting option outside this study. A provisioned instance charges while idle, so throughput and utilisation matter as much as the hourly price. Part 5 records the Sydney instance arithmetic and the separate vendor benchmark. We have also still not tested Laya, Kev or Bespoke Nimble, named as alternatives in Part 1; omission says nothing about their quality.
Next steps
Choose one repeated contact centre, digital or IT service decision and test the complete route on your own cases, including wrong routes and missed or unnecessary handoffs. Contact Humint Labs to design, cost and build that first workflow.