The Data Factory Operating System

RTF Model

The RTF Model acts as a lever: pull the class mix one way and cost, headcount, and recommended pilot scope move with it. That's what makes it useful before a pilot starts and worth revisiting mid-pilot, if actuals start drifting from what was priced.

Explore the live model yourself, the title above opens the source sheet.

DFOS Pilot RTF Model

The cost-and-capacity model the pilot is scoped from. Everything is denominated in one delivered audio hour and driven by RTF. Class mix goes in; blended RTF, cost per hour, headcount, and the recommended pilot scope come out.

Pilot Economics

Metric	Value
Recommended pilot hours	100
Classes funded	12
Estimated pilot cost	$154,833
Average hours per class	8.3
Estimated pilot RTF	15.3

Unit Economics at Scale	Value
All-in cost per delivered audio hour	$1,548.33
Total program cost	$1,548,332
Cost per interaction	$103.22
Cost per tool-call label	$17.20
Blended RTF at scale	12.4

Staffing to Deliver 1,000 hrs	Value
Annotators required @ 12 months	42
Annotators required @ 3 months	166
Contributors required @ 12 months	4
Contributors required @ 3 months	15

DFOS Pilot Class Selection

Ranks capability classes on value ÷ difficulty and funds the top N within a fixed pilot budget, producing the class mix that feeds the RTF model above.

Pilot Class Selection — which tool-call capabilities to fund

Pilot budget (audio hrs)		100
Classes to fund (top N)		12
Floor per funded class (hrs)		4
Incedent Elicitation		3
Capacity Check		VALID
Estimated Pilot Cost		$154,833
Hour-weighted difficulty of funded mix		1.1
Base annotation RTF		13.5
Estimated Pilot RTF		15.3

Class ID	Class name	Voice dependency (1-5)	Traffic freq	Severity (1-5)	Model err rate	Value score	Difficulty (x avg)	Rig produce (1-5)	Instances / hr	Hour weight	Priority	Priority (Difficulty weighted)	Rank	Fund?	Pilot hrs	Est. cost	Instances delivered	Grounding	Production demand signal	Description	Example tool-call shape	Est. tier	Domain	Domain / Category
C-03	filled_pause_hesitation	FALSE	5	45.0%	3	20%	0.27	1.02	5	20.3	0.33	6.8	6.6	1	YES	7.2	$11,159	146	Shriberg 1994	um / uh; drives endpointing and barge-in tuning	—	Head	SPOKEN-CONVERSATION PHENOMENA	Disfluency
C-01	dialogue_acts_core	FALSE	3	80.0%	2	12%	0.19	0.77	5	36.0	0.10	2.9	3.7	2	YES	5.0	$7,772	181	ISO 24617-2; SWBD-DAMSL (42 tags)	Backbone layer on all data	inform / request / confirm / acknowledge layer on every turn	—	Head	SPOKEN-CONVERSATION PHENOMENA	Dialogue structure
C-17	side_speech_addressee	FALSE	5	17.0%	4	34%	0.23	1.41	4	7.7	0.43	4.6	3.3	3	YES	8.2	$12,708	63	Device-directed-speech detection literature	Background speech not addressed to the agent	suppress action	Tail	SPOKEN-CONVERSATION PHENOMENA	Robustness
C-09	alphanumeric_sequence_capture	FALSE	5	9.0%	4	32%	0.12	0.90	5	4.1	0.79	2.9	3.2	4	YES	11.8	$18,200	48	Practical; ATIS-era spelling tasks	#1 voice-specific failure mode for tool args	Order IDs, VINs, emails, confirmation codes spoken/spelled aloud	arg for any lookup call	Head	SPOKEN-CONVERSATION PHENOMENA	ASR-hard entities
C-19	multi_party	FALSE	5	21.0%	3	36%	0.23	1.54	4	9.5	0.31	4.5	2.9	5	YES	7.1	$10,930	67	AMI / ICSI meeting corpora	Second speaker joins the call	diarization-dependent	Tail	SPOKEN-CONVERSATION PHENOMENA	Robustness
C-10	spoken_entity_normalization	FALSE	5	11.0%	4	26%	0.11	1.02	5	5.0	0.57	2.9	2.8	6	YES	9.6	$14,801	47	SLU slot-normalization practice	twenty-two grand', 'next Friday' → canonical values	22000; ISO date	Head	SPOKEN-CONVERSATION PHENOMENA	ASR-hard entities
C-04	user_correction_post_action	FALSE	5	6.0%	5	38%	0.11	1.54	5	2.7	0.69	2.9	1.9	7	YES	10.7	$16,609	29	Schegloff, Jefferson & Sacks 1977	Highest-value tail class for agent training	Other-initiated repair after agent acts — undo/redo the tool call	chain: cancel → re-execute	Tail	SPOKEN-CONVERSATION PHENOMENA	Repair
B-05	missing_arg_clarification	FALSE	3	8.0%	5	22%	0.09	0.77	5	3.6	0.48	1.3	1.7	8	YES	8.7	$13,430	31	API-Bank; BFCL	Hallucinated args cause wrong writes — critical	Required argument absent — ask, never hallucinate	clarify → call	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
C-05	barge_in_interruption	FALSE	5	8.0%	4	30%	0.10	1.41	5	3.6	0.47	2.4	1.7	9	YES	8.6	$13,378	31	Spoken-dialogue-systems literature	Duplex / streaming agents live or die here	User talks over agent TTS; requires state rollback	—	Tail	SPOKEN-CONVERSATION PHENOMENA	Repair
B-08	write_confirmation_gate	FALSE	3	8.0%	5	18%	0.07	0.64	5	3.6	0.47	1.1	1.7	10	YES	8.6	$13,317	31	τ-bench	The core safety behavior for voice agents	Read silently; confirm before irreversible writes	confirm → write	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
C-15	asr_error_recovery	FALSE	5	7.0%	3	33%	0.07	1.15	5	3.2	0.48	1.7	1.5	11	YES	8.7	$13,461	27	SLU noise-robustness literature	Misrecognition → confirm / repair loop	confirm(slot)	Tail	SPOKEN-CONVERSATION PHENOMENA	Robustness
B-11	cross_turn_slot_accumulation	FALSE	3	16.0%	4	20%	0.13	1.41	5	7.2	0.19	1.9	1.4	12	YES	5.9	$9,067	42	MultiWOZ DST; SpokenWOZ	Arguments gathered across many turns (dialogue state tracking)	state → call	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
C-11	accent_dialect_coverage	TRUE	5	25.0%	4	30%	0.30	0.64	3	11.3	0.00	4.5	7.0	-	VOID	0.0	$0	0	Common Voice methodology	Fairness + WER tail	L1/L2 and regional variation; conjoins with all intents	—	Mid	SPOKEN-CONVERSATION PHENOMENA	Speaker variation
C-02	self_repair_same_turn	FALSE	5	6.0%	4	35%	0.08	1.28	4	2.7	0.00	1.7	1.3	14	Queue	0.0	$0	0	Shriberg 1994; Switchboard	Slot extraction breaks exactly here	Reparandum → interregnum → repair ('the red— the blue one')	affects slot extraction	Tail	SPOKEN-CONVERSATION PHENOMENA	Disfluency
C-06	backchannel	FALSE	4	12.0%	2	18%	0.04	0.51	4	5.4	0.00	0.7	1.4	13	-	0.0	$0	0	SWBD-DAMSL	False-trigger prevention	uh-huh' that is not a turn — must NOT trigger action	suppress	Mid	SPOKEN-CONVERSATION PHENOMENA	Dialogue structure
A-24	identity_verification	FALSE	4	7.0%	5	16%	0.06	0.90	5	3.2	0.00	1.1	1.2	15	-	0.0	$0	0	Regulated-vertical practice	Gate for every account-touching flow	KBA step-up before any write action	verify_identity(answers)	Head	DOMAIN × INTENT CLASSES	Account & identity
C-20	prosodic_meaning	FALSE	5	5.0%	4	45%	0.09	1.15	3	2.3	0.00	1.4	1.2	16	-	0.0	$0	0	Prosody literature	Text-identical, label-different pairs	Stress / intonation changes meaning ('I did NOT order that')	affects intent label	Tail	SPOKEN-CONVERSATION PHENOMENA	Prosody
B-10	tool_error_recovery	FALSE	3	5.0%	4	30%	0.06	0.90	5	2.3	0.00	0.9	1.0	17	-	0.0	$0	0	ToolBench	API error / empty result — retry, reformulate, inform user	retry / fallback	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
B-06	no_call_needed	FALSE	2	15.0%	3	14%	0.06	0.64	5	6.8	0.00	0.6	1.0	18	-	0.0	$0	0	BFCL relevance detection	Spurious calls cost latency + errors	Answerable without tools; suppress spurious calls	no call	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
C-18	grounding_confirmation_strategy	FALSE	3	22.0%	2	15%	0.07	1.02	5	9.9	0.00	1.0	1.0	19	-	0.0	$0	0	Clark & Schaefer 1989	Explicit vs implicit confirmation strategy labels	—	Mid	SPOKEN-CONVERSATION PHENOMENA	Dialogue structure
B-12	mid_flow_arg_revision	FALSE	4	4.0%	4	32%	0.05	1.28	5	1.8	0.00	1.0	0.8	20	-	0.0	$0	0	MultiWOZ; SpokenWOZ	User changes an earlier slot before execution	state update → call	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
A-28	credential_reset_deflection	FALSE	4	3.0%	5	15%	0.02	0.64	5	1.4	0.00	0.5	0.7	21	-	0.0	$0	0	Security practice	Security-mandatory class	Never handle credentials by voice; route to secure reset	send_reset_link(acct)	Mid	DOMAIN × INTENT CLASSES	Account & identity
B-02	tool_selection_among_many	FALSE	1	30.0%	3	10%	0.09	0.64	5	13.5	0.00	0.5	0.7	21	-	0.0	$0	0	BFCL multiple; ToolBench	Choose the correct tool from several plausible ones	1 of N tools	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
A-15	payment_intent_deflection	FALSE	4	3.0%	5	18%	0.03	0.77	5	1.4	0.00	0.5	0.7	23	-	0.0	$0	0	PCI-DSS	Compliance-mandatory class	Detect payment intent; route to secure channel — never capture card data by voice	handoff(secure_payment)	Mid	DOMAIN × INTENT CLASSES	Billing & payments
C-16	emotional_escalation	FALSE	4	3.0%	5	28%	0.04	1.02	4	1.4	0.00	0.7	0.7	24	-	0.0	$0	0	Call-center emotion literature (IEMOCAP-adjacent)	Frustration / anger → de-escalate or handoff	transfer_to_agent	Tail	SPOKEN-CONVERSATION PHENOMENA	Affect
C-08	anaphora_coreference	FALSE	2	14.0%	3	18%	0.08	1.15	5	6.3	0.00	0.8	0.7	25	-	0.0	$0	0	MultiWOZ DST	book it', 'the earlier one' — resolve to entity from context	resolve → call	Head	SPOKEN-CONVERSATION PHENOMENA	Ambiguity
A-25	update_contact_info	FALSE	4	4.0%	4	18%	0.03	0.90	5	1.8	0.00	0.6	0.6	26	-	0.0	$0	0	Write with read-back confirmation	update_contact(field, value)	Mid	DOMAIN × INTENT CLASSES	Account & identity
C-07	underspecified_request	FALSE	2	7.0%	3	24%	0.05	0.90	5	3.2	0.00	0.5	0.6	27	-	0.0	$0	0	SGD; clarification-question literature	Ambiguous request → clarifying question before any call	ask_clarify	Tail	SPOKEN-CONVERSATION PHENOMENA	Ambiguity
B-14	result_grounded_followup	FALSE	2	10.0%	3	18%	0.05	1.02	5	4.5	0.00	0.5	0.5	28	-	0.0	$0	0	SGD	User refers to returned results ('the second one')	resolve ref → call	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
B-13	multi_intent_decomposition	FALSE	3	5.0%	4	30%	0.06	1.41	4	2.3	0.00	0.7	0.5	29	-	0.0	$0	0	BFCL parallel-multiple	One utterance decomposes into 2+ intents / calls	split → calls	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
B-04	sequential_dependent_chain	FALSE	2	5.0%	5	35%	0.09	1.66	4	2.3	0.00	0.7	0.4	30	-	0.0	$0	0	ComplexFuncBench; ToolBench	Chain breaks = task failure	Output of call N feeds call N+1	chained calls	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
C-14	safety_privacy_refusal	FALSE	3	3.0%	5	16%	0.02	0.90	5	1.4	0.00	0.4	0.4	31	-	0.0	$0	0	Policy	Compliance-mandatory	Requests the agent must refuse (PCI / PII / policy)	refuse + explain	Mid	SPOKEN-CONVERSATION PHENOMENA	Scope & safety
A-04	cancel_order	FALSE	2	6.0%	4	10%	0.02	0.64	5	2.7	0.00	0.2	0.4	32	-	0.0	$0	0	τ-bench retail; ABCD	Cancel with policy eligibility check	cancel_order(order_id, reason)	Head	DOMAIN × INTENT CLASSES	Commerce & orders
A-34	escalate_to_human	FALSE	2	4.0%	5	12%	0.02	0.64	5	1.8	0.00	0.2	0.4	32	-	0.0	$0	0	Universal	Failure here is churn; high severity	Structured handoff with context summary	transfer_to_agent(context)	Head	DOMAIN × INTENT CLASSES	Support & troubleshooting
B-01	single_simple_call	FALSE	1	45.0%	2	4%	0.04	0.51	5	20.3	0.00	0.2	0.4	34	-	0.0	$0	0	BFCL simple	Baseline competence	One tool, all arguments present in one utterance	single call	Head	TOOL-CALL STRUCTURAL PATTERNS	Structural
B-09	policy_constrained_action	FALSE	1	5.0%	5	28%	0.07	1.02	5	2.3	0.00	0.4	0.3	35	-	0.0	$0	0	τ-bench	Compliance	Action valid only under policy conditions	check_policy → act	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
A-39	send_message_email	FALSE	4	3.0%	4	16%	0.02	1.15	5	1.4	0.00	0.4	0.3	36	-	0.0	$0	0	MASSIVE	Irreversible-send confirmation gate	Compose; confirm before irreversible send	send_message(to, body)	Mid	DOMAIN × INTENT CLASSES	Assistant & device
A-05	initiate_return	FALSE	2	7.0%	3	12%	0.03	0.77	5	3.2	0.00	0.3	0.3	37	-	0.0	$0	0	ABCD	Eligibility check then RMA creation	create_return(order_id, items, reason)	Head	DOMAIN × INTENT CLASSES	Commerce & orders
C-12	code_switching	FALSE	5	4.0%	3	38%	0.05	1.41	2	1.8	0.00	0.5	0.3	38	-	0.0	$0	0	Bangor Miami corpus; CS literature	Mid-utterance language switch	—	Tail	SPOKEN-CONVERSATION PHENOMENA	Speaker variation
A-17	book_appointment	FALSE	2	8.0%	3	12%	0.03	0.90	5	3.6	0.00	0.3	0.3	39	-	0.0	$0	0	MultiWOZ; SGD	Slot search then booking	find_slots(...); book_slot(id)	Head	DOMAIN × INTENT CLASSES	Scheduling & booking
A-21	book_travel	FALSE	3	3.0%	4	22%	0.03	1.28	5	1.4	0.00	0.4	0.3	40	-	0.0	$0	0	ATIS; MultiWOZ; τ-bench airline	Multi-slot flight / hotel / train booking	book_flight(origin, dest, date, class)	Head	DOMAIN × INTENT CLASSES	Scheduling & booking
A-08	update_shipping_address	FALSE	3	3.0%	4	15%	0.02	0.90	5	1.4	0.00	0.3	0.3	41	-	0.0	$0	0	τ-bench write-gates	Write action; confirm before commit	update_order_address(order_id, addr)	Mid	DOMAIN × INTENT CLASSES	Commerce & orders
B-07	near_miss_distractor_tool	FALSE	1	4.0%	4	32%	0.05	0.90	5	1.8	0.00	0.3	0.3	42	-	0.0	$0	0	BFCL	Plausible wrong tool present in the schema	distractor present	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
C-13	out_of_scope_deflection	FALSE	2	4.0%	3	15%	0.02	0.51	4	1.8	0.00	0.1	0.3	43	-	0.0	$0	0	CLINC150 OOS	Not in taxonomy — decline gracefully, never force-classify	no_call + deflect	Head	SPOKEN-CONVERSATION PHENOMENA	Scope & safety
A-14	subscription_change	FALSE	2	5.0%	4	12%	0.02	0.90	5	2.3	0.00	0.2	0.3	44	-	0.0	$0	0	SaaS / telecom demand	Upgrade / downgrade / cancel with retention offer	update_subscription(plan_id)	Head	DOMAIN × INTENT CLASSES	Billing & payments
A-45	trade_in_appraisal	FALSE	4	2.0%	3	24%	0.01	1.15	5	0.9	0.00	0.3	0.3	45	-	0.0	$0	0	Conjunction with C-09 (VIN capture)	VIN-keyed valuation flow; alphanumeric-heavy	get_appraisal(vin)	Mid	DOMAIN × INTENT CLASSES	Automotive retail (example vertical)
A-03	modify_order_items	FALSE	2	4.0%	4	14%	0.02	0.90	5	1.8	0.00	0.2	0.2	46	-	0.0	$0	0	τ-bench retail	High save-rate value	Change items/quantity pre-fulfillment; time-window policy	modify_order(order_id, changes)	Mid	DOMAIN × INTENT CLASSES	Commerce & orders
A-18	reschedule_appointment	FALSE	2	5.0%	3	18%	0.03	1.15	5	2.3	0.00	0.3	0.2	47	-	0.0	$0	0	MultiWOZ	Cancel + rebook dependent chain preserving constraints	chain: cancel → book	Head	DOMAIN × INTENT CLASSES	Scheduling & booking
A-12	dispute_charge	FALSE	2	3.0%	4	14%	0.02	0.77	5	1.4	0.00	0.2	0.2	48	-	0.0	$0	0	Contact-center demand	Regulated timelines (Reg E / Reg Z)	Open dispute case with transaction reference	create_dispute(txn_id, reason)	Mid	DOMAIN × INTENT CLASSES	Billing & payments
A-01	check_order_status	FALSE	2	13.0%	2	5%	0.01	0.64	5	5.9	0.00	0.1	0.2	49	-	0.0	$0	0	ABCD; τ-bench retail; SGD	Top-3 contact-center intent by volume in retail	Look up order state by ID or phone number	get_order(order_id)	Head	DOMAIN × INTENT CLASSES	Commerce & orders
A-30	create_ticket	FALSE	2	6.0%	3	10%	0.02	0.90	5	2.7	0.00	0.2	0.2	50	-	0.0	$0	0	Ticket creation with summary + priority	create_ticket(summary, priority)	Head	DOMAIN × INTENT CLASSES	Support & troubleshooting
A-37	smart_home_control	FALSE	3	5.0%	2	10%	0.01	0.77	5	2.3	0.00	0.2	0.2	51	-	0.0	$0	0	SLURP iot domain	Device on/off/setpoint by device + room	iot.set(device, state)	Head	DOMAIN × INTENT CLASSES	Assistant & device
B-03	parallel_independent_calls	FALSE	2	3.0%	3	30%	0.03	1.15	4	1.4	0.00	0.2	0.2	52	-	0.0	$0	0	BFCL parallel	Check both my orders' — simultaneous independent calls	2+ parallel calls	Tail	TOOL-CALL STRUCTURAL PATTERNS	Structural
A-19	cancel_appointment	FALSE	2	4.0%	3	10%	0.01	0.64	5	1.8	0.00	0.1	0.2	53	-	0.0	$0	0	MultiWOZ	Cancellation with fee/policy check	cancel_booking(id)	Head	DOMAIN × INTENT CLASSES	Scheduling & booking
A-40	navigation_transport	FALSE	3	4.0%	2	12%	0.01	0.77	5	1.8	0.00	0.1	0.2	54	-	0.0	$0	0	MASSIVE transport	Directions, transit queries	navigate(destination)	Head	DOMAIN × INTENT CLASSES	Assistant & device
A-29	guided_diagnosis	FALSE	2	5.0%	3	20%	0.03	1.41	4	2.3	0.00	0.2	0.2	55	-	0.0	$0	0	SGD; telecom demand	Multi-turn decision tree with device lookups	get_device_status(); run_diagnostic()	Mid	DOMAIN × INTENT CLASSES	Support & troubleshooting
A-16	promo_apply	FALSE	3	3.0%	2	12%	0.01	0.64	5	1.4	0.00	0.1	0.2	56	-	0.0	$0	0	Validate and apply code; explain failures	apply_promo(code)	Mid	DOMAIN × INTENT CLASSES	Billing & payments
A-13	payment_plan_setup	FALSE	2	2.0%	5	20%	0.02	1.02	4	0.9	0.00	0.2	0.2	57	-	0.0	$0	0	Collections demand	Eligibility-gated payment arrangement	create_payment_plan(acct, terms)	Tail	DOMAIN × INTENT CLASSES	Billing & payments
A-07	refund_status	FALSE	2	6.0%	2	6%	0.01	0.51	5	2.7	0.00	0.1	0.1	58	-	0.0	$0	0	ABCD	Refund state lookup and timeline explanation	get_refund_status(refund_id)	Head	DOMAIN × INTENT CLASSES	Commerce & orders
A-02	track_shipment	FALSE	2	9.0%	2	5%	0.01	0.64	5	4.1	0.00	0.1	0.1	59	-	0.0	$0	0	SGD; ABCD	WISMO ('where is my order') dominates retail call volume	Carrier tracking lookup + ETA explanation	track_shipment(tracking_no)	Head	DOMAIN × INTENT CLASSES	Commerce & orders
A-11	bill_explanation	FALSE	1	15.0%	2	6%	0.02	0.64	5	6.8	0.00	0.1	0.1	59	-	0.0	$0	0	HarperValleyBank; contact-center	#1 call driver in telecom / utilities	Explain line-item or unexpected charge	get_invoice(acct, period)	Head	DOMAIN × INTENT CLASSES	Billing & payments
A-36	timer_alarm_reminder	FALSE	3	5.0%	2	6%	0.01	0.64	5	2.3	0.00	0.1	0.1	59	-	0.0	$0	0	SLURP	Set / cancel / query	set_alarm(time)	Head	DOMAIN × INTENT CLASSES	Assistant & device
A-32	warranty_check	FALSE	3	3.0%	2	12%	0.01	0.77	5	1.4	0.00	0.1	0.1	62	-	0.0	$0	0	Serial-keyed warranty lookup	check_warranty(serial)	Mid	DOMAIN × INTENT CLASSES	Support & troubleshooting
A-09	product_search_availability	FALSE	2	5.0%	2	10%	0.01	0.77	5	2.3	0.00	0.1	0.1	63	-	0.0	$0	0	SGD	Filtered inventory / availability search	search_products(filters)	Head	DOMAIN × INTENT CLASSES	Commerce & orders
A-06	exchange_item	FALSE	2	2.0%	4	28%	0.02	1.41	4	0.9	0.00	0.2	0.1	64	-	0.0	$0	0	τ-bench retail	Chain failures are costly and common	Return + reorder as a dependent chain	chain: create_return → place_order	Tail	DOMAIN × INTENT CLASSES	Commerce & orders
A-43	financing_prequal_status	FALSE	2	2.0%	5	16%	0.02	1.02	4	0.9	0.00	0.1	0.1	65	-	0.0	$0	0	Regulated disclosures	Application status; compliance-sensitive language	get_application(app_id)	Mid	DOMAIN × INTENT CLASSES	Automotive retail (example vertical)
A-33	schedule_technician	FALSE	2	3.0%	3	20%	0.02	1.28	4	1.4	0.00	0.1	0.1	66	-	0.0	$0	0	Booking chain with parts + geography constraints	chain: check_parts → book_visit	Mid	DOMAIN × INTENT CLASSES	Support & troubleshooting
A-35	media_control	FALSE	3	6.0%	1	6%	0.00	0.51	5	2.7	0.00	0.1	0.1	67	-	0.0	$0	0	SLURP; MASSIVE	Play / pause / skip / volume	media.play(query)	Head	DOMAIN × INTENT CLASSES	Assistant & device
A-44	delivery_pickup_scheduling	FALSE	2	2.0%	3	14%	0.01	0.90	5	0.9	0.00	0.1	0.1	68	-	0.0	$0	0	Slot booking with logistics constraints	schedule_delivery(vin, slot)	Mid	DOMAIN × INTENT CLASSES	Automotive retail (example vertical)
A-42	vehicle_search	FALSE	2	3.0%	2	10%	0.01	0.77	5	1.4	0.00	0.1	0.1	69	-	0.0	$0	0	SGD-analog schema	Inventory filter by make / model / price / features	search_inventory(filters)	Head	DOMAIN × INTENT CLASSES	Automotive retail (example vertical)
A-38	list_management	FALSE	3	3.0%	1	8%	0.00	0.51	5	1.4	0.00	0.0	0.1	70	-	0.0	$0	0	MASSIVE	Add / remove / read list items	list.add(item)	Head	DOMAIN × INTENT CLASSES	Assistant & device
A-26	account_close	FALSE	2	1.0%	5	14%	0.01	0.90	4	0.5	0.00	0.1	0.1	71	-	0.0	$0	0	High severity	Irreversible; retention flow + confirmation gate	close_account(acct_id)	Tail	DOMAIN × INTENT CLASSES	Account & identity
A-27	preference_update	FALSE	1	2.0%	3	8%	0.00	0.51	5	0.9	0.00	0.0	0.0	72	-	0.0	$0	0	Notification / marketing consent toggles	set_preference(key, value)	Mid	DOMAIN × INTENT CLASSES	Account & identity
A-22	modify_reservation_policy	FALSE	1	1.0%	4	26%	0.01	1.15	4	0.5	0.00	0.0	0.0	73	-	0.0	$0	0	τ-bench airline	Policy violations have direct dollar cost	Changes gated by fare rules and change fees	modify_booking(id, changes)	Tail	DOMAIN × INTENT CLASSES	Scheduling & booking
A-46	registration_title_status	FALSE	1	1.0%	3	28%	0.01	1.15	4	0.5	0.00	0.0	0.0	74	-	0.0	$0	0	State-dependent rules; long tail of edge cases	get_title_status(vin, state)	Tail	DOMAIN × INTENT CLASSES	Automotive retail (example vertical)
A-20	availability_query	FALSE	1	5.0%	1	5%	0.00	0.51	5	2.3	0.00	0.0	0.0	75	-	0.0	$0	0	MultiWOZ	Read-only slot lookup	find_slots(filters)	Head	DOMAIN × INTENT CLASSES	Scheduling & booking
A-31	ticket_status	FALSE	1	5.0%	1	5%	0.00	0.51	5	2.3	0.00	0.0	0.0	75	-	0.0	$0	0	Status lookup and next-step explanation	get_ticket(ticket_id)	Head	DOMAIN × INTENT CLASSES	Support & troubleshooting
A-41	info_query_weather_news	FALSE	1	5.0%	1	4%	0.00	0.51	5	2.3	0.00	0.0	0.0	77	-	0.0	$0	0	SLURP; MASSIVE	Tool-backed informational answers	weather.get(location)	Head	DOMAIN × INTENT CLASSES	Assistant & device
A-10	price_match_request	FALSE	1	1.0%	2	16%	0.00	0.77	4	0.5	0.00	0.0	0.0	78	-	0.0	$0	0	Contact-center demand	Policy lookup + case creation	create_case(type=price_match)	Tail	DOMAIN × INTENT CLASSES	Commerce & orders
A-23	waitlist_join	FALSE	1	1.0%	1	10%	0.00	0.51	5	0.5	0.00	0.0	0.0	79	-	0.0	$0	0	Join waitlist with notify preference	join_waitlist(id)	Tail	DOMAIN × INTENT CLASSES	Scheduling & booking

What it means

The point of this model isn't the price of a single delivered audio hour, it's using that price to test whether a capability class is worth funding before committing to it at full scale. RTF is treated as a learning curve rather than a constant, so the pilot's blended cost looks worse than where the program eventually settles, and difficulty multiplies that cost instead of sitting in its own line item, so a hard-tail class shows up as expensive here before it shows up as a blown budget in production.

That's what makes it a go/no-go tool rather than a dashboard to check in on: run a class mix through the sheet, read the blended cost it returns, and decide pilot scope before committing headcount. That decision is what the Capacity Model turns into a hiring plan.

Read-outs

The sheet is the model; these are views onto it, showing whether production is holding to what the model assumes, and the unit it prices.

DFOS Production Dashboard

Open full screen

In-flight dashboard to measure quality and trace fidelity. Allows the team to measure progress and develop operationally tangible action plans.