AI cost tracking for brokerages: knowing what each conversation actually costs

A brokerage adopting AI in their sales stack inherits a new cost category: AI inference. Voice transcription costs roughly $0.006/minute. Language model calls cost roughly $0.001–$0.05 per lead conversation, depending on the model and the length. Vector embedding for the knowledge base costs essentially zero per call but accumulates at scale. Image and PDF processing each have their own meters.

Most AI products bundle this into a flat subscription fee and obscure the underlying economics. The vendor is making margin on the difference between the price and the actual inference cost. The brokerage owner sees a monthly invoice and no insight into what's driving it.

This is the wrong default for a regulated, audit-sensitive industry. AI cost should be transparent — per tenant, per feature, per lead — and visible to the brokerage in real time.

What costs what

A brokerage running a serious AI stack is paying for five categories of inference.

Language model calls. The AI BDR's conversation responses, the briefing composition, the voice memo extraction, the knowledge base retrieval Q&A. These are priced per input + output token, with newer/better models costing 5–20× the older/smaller ones.

Voice transcription. Inbound voice notes from leads, producer post-showing memos, voice calls if transcribed. Priced per minute, typically $0.006–$0.015 depending on the provider.

Image processing. OCR of pre-approval letters, vision-based description of property photos, KYC document scans. Priced per image, typically $0.001–$0.01.

Vector embeddings. Each chunk in the knowledge base gets embedded once; each retrieval query gets embedded each time. Priced per token, very low ($0.00001 per chunk roughly).

Reranking and search. Some retrieval pipelines use a separate reranking model. Priced per ranked pair, very low.

These five compound differently. Language model calls are the largest line item in almost every deployment. Voice transcription is the second. The other three are rounding errors at most scales.

What a real cost meter tracks

A working AI cost meter inside a brokerage tracks four dimensions for every inference call.

Tenant. Which brokerage account incurred the cost. In multi-tenant SaaS, this is essential — the vendor needs to know what to bill each customer and each customer needs visibility into their own spend.

Feature. Which product feature triggered the call. Sara responding to a WhatsApp lead is one feature. The producer's voice memo extraction is another. The morning briefing composition is a third. Aggregating costs by feature is how the brokerage knows which features are paying for themselves.

Lead. Which lead's conversation generated the cost. This is where the real economics surface. A high-touch lead with 80 messages over 90 days might cost $3 in AI inference. A cold lead that never engaged might cost $0.05. The brokerage can compute ROI per lead.

Time. Hour-level resolution. Costs concentrate around inbound times (mornings, evenings, weekend afternoons) and the meter should show this. It also surfaces anomalies — an unexpected cost spike at 3am is signaling something wrong, possibly an automated loop.

A meter that tracks all four dimensions is a meter you can ask real questions of. "What's our cost per lead this month, by source?" is answerable. "Which feature has the highest cost-per-conversion?" is answerable. "Why did our AI bill jump last Thursday?" is answerable.

The "ai_usage_metric" table

The data structure that supports this is mundane. A single table, one row per inference call. Roughly:

Column	Type	Notes
id	uuid	Primary key
tenant_id	uuid	The brokerage account
feature	text	sara_response, voice_extract, briefing, etc.
lead_id	uuid (nullable)	Set when the call relates to a specific lead
model	text	claude-opus-4-7, whisper-1, gpt-4o, etc.
input_tokens	int (nullable)	For language model calls
output_tokens	int (nullable)	For language model calls
duration_sec	numeric (nullable)	For audio transcription
image_count	int (nullable)	For vision calls
cost_usd	numeric	Computed at write-time
created_at	timestamptz	Indexed

Every inference call writes one row. Costs roll up by tenant, feature, lead, and time window. A 90-day window for a mid-sized brokerage might have 50–200k rows — small enough to query directly without a warehouse.

This structure is not novel. The discipline is consistently writing to it from every code path that calls a model.

What the brokerage sees

On a dashboard the brokerage owner opens once a week:

Spend this month. Total AI cost, broken down by feature. Sara responses, voice extraction, briefings, knowledge base retrieval, image processing. With trend vs. last month.

Cost per lead. Distribution of AI spend across leads. Median, p90, p99. The expensive leads named, with reasons (very long conversation, lots of multimedia, etc.).

Cost per won deal. AI cost as a share of commission. If the AI cost $4.20 to close a $1.4M deal with a $42k commission, that's 0.01% — easy decision. If it cost $4,200, that's a different decision.

Cost-per-conversion by feature. Which features generated leads that closed. Sara's outbound vs. inbound widget responses vs. WhatsApp cadence — each measured against the conversions it drove.

Anomalies. Days when AI cost was 2× normal, with what feature drove it. This is how operational issues surface — a runaway loop, a misconfigured prompt running 10× the normal token count, a stuck retry queue.

The brokerage owner reading this is no longer evaluating AI on faith. They have line-item economics.

The vendor incentive question

Most AI products don't expose this dashboard because it's against the vendor's interest. The flat-fee subscription model relies on the vendor capturing the margin between what the brokerage pays and what the inference actually costs. Showing the meter exposes the margin.

The honest framing is: brokerages should be willing to pay vendor margin, because the vendor is providing more than inference. They're providing the integration, the engineering, the support, the compliance work, the product surface. But the margin should be explicit, not hidden.

A vendor that shows the meter and prices transparently is signaling that they're confident in the value they add beyond reselling tokens. A vendor that obscures the meter is signaling that their margin depends on the customer not knowing the underlying cost.

This is the same dynamic Stripe made standard in payments. Stripe shows you exactly what each charge cost and what their fee was. Before Stripe, payment processors made their margin in the dark. After Stripe, the model shifted. AI in vertical SaaS is in the pre-Stripe phase right now.

What "real-time" means here

Per-call metering, written synchronously into the usage table, is what makes the rest of this possible. Three properties matter.

Sub-second latency to record. The inference call returns, the cost row is written, before the user-facing response is sent. No batch jobs, no nightly ETL. The brokerage's dashboard at 10am reflects calls made at 9:59am.

No off-by-one accounting. Token counts come straight from the model provider's response. They're not estimated. They're not rounded. They're the same numbers the vendor sees on their own invoice.

Cost-at-write resolution. The cost is computed and stored at the time of the call, using the price schedule in effect that moment. Price changes don't retroactively rewrite history.

This makes the meter trustworthy. The brokerage can audit it, the vendor can audit it, and external auditors (for SOC 2, for fair-billing inquiries) can audit it. There's no opacity.

Where this is heading

The transparency standard is going to become a requirement, not a feature. Brokerages buying AI products in 2027 will expect the same per-feature, per-lead, per-tenant visibility they get from Stripe on payments and from cloud providers on infrastructure. Products that don't ship it will lose deals to products that do.

The discipline isn't novel. It's the same metering and reporting any other metered infrastructure does. Bringing it to AI in real estate is just the application.

---

The Closi AI usage dashboard is built directly on the ai_usage_metric table — per-tenant, per-feature, per-lead resolution, real-time, with anomaly detection. Closi's security and transparency posture covers the data ownership and audit guarantees.