Capacity & Scaling Analysis

Pre-work for Session 1 | Prepared: 2026-02-17

The Scaling Story (3-sentence summary for Danny)

Envo is built on a serverless architecture (Vercel + Supabase) that scales automatically with demand — we do not need to buy bigger servers as we grow. The primary cost drivers are usage-based: LLM calls per tenant conversation, Twilio per WhatsApp/SMS message, and VAPI per voice minute — meaning costs grow proportionally with tenant activity, not just property count. At 3,000 properties our estimated platform cost is GBP 1.50—4.50 per property per month (depending on tenant interaction volume), and the BYOAK model lets us shift the most expensive costs (LLM, Twilio, voice) to larger customers.

Infrastructure Stack

Plain-English version: Everything runs in the cloud, managed by third parties. No servers to maintain.

Layer	Service	Why It Scales
Web app + API	Vercel (Next.js)	Serverless — spins up capacity automatically per request, scales to zero when idle
Database	Supabase (PostgreSQL + pgvector)	Managed cloud database; we can upgrade compute independently of storage
Auth	Supabase Auth	Included; handles up to 100K monthly active users on Pro plan
File storage	Supabase Storage	Pay-per-GB; stores issue photos, compliance documents
AI brain	Claude (Anthropic) with OpenAI fallback	Pay-per-token; no capacity limits, multi-provider fallback chain
Embeddings	OpenAI text-embedding-3-small	Pay-per-token; one-time cost per document, near-zero per query
WhatsApp + SMS	Twilio	Pay-per-message; no infrastructure to manage
Voice	VAPI	Pay-per-minute; includes STT + TTS + telephony
Email	SendGrid	Pay-per-email; Essentials plan from GBP 16/mo

Key point: Nothing in this stack has a hard capacity ceiling. Every layer is pay-as-you-go or upgradeable. The question is not “can it handle 3,000 properties?” but “what does it cost at 3,000 properties?”

Cost Drivers

1. LLM Generation (Claude / OpenAI)

What it is: Every tenant conversation triggers AI calls — intent classification, RAG retrieval, response generation, and tool use (e.g. creating issues).
How it scales: Per-message (each tenant message = 1+ LLM call). A typical conversation might be 4-8 messages.
Current pricing: Claude Sonnet 4 = $3/ M in p u tt o k e n s,$ 15/M output tokens. A single query-response cycle uses roughly 1,500 input tokens (system prompt + context + retrieved chunks) and 300 output tokens.
Estimated cost per query: ~$0.01-0.03 (from RAG Pipeline doc; validated by token pricing).
Estimated cost per conversation (6 messages avg): ~$0.06-0.18.

Scale	Assumed conversations/mo	Monthly LLM cost
100 properties	200	GBP 12-36
500 properties	1,000	GBP 60-180
1,000 properties	2,000	GBP 120-360
3,000 properties	6,000	GBP 360-1,080

Assumption: ~2 conversations per property per month (conservative for HMOs; could be higher). Pricing converted at $1 = GBP 0.80 approximately.

2. Embeddings (OpenAI)

What it is: Converting property documents (house rules, manuals, lease info) into vectors for AI search.
How it scales: One-time cost per document upload + near-zero per query.
Current pricing: $0.02/ Mt o k e n s . A 10 - p a g e d oc u m e n t 5, 000 t o k e n s =$ 0.0001.
Per-query cost: ~$0.0001 (negligible).
At scale: Even at 3,000 properties with 20 documents each = 60,000 documents. Initial embedding cost ~ GBP 5-10 total. Ongoing query costs negligible.

Verdict: Not a meaningful cost driver. Effectively free.

3. Twilio — WhatsApp

What it is: Sending and receiving WhatsApp messages to/from tenants.
How it scales: Per-message. Both inbound and outbound are charged.
Current pricing: ~ $0.005 p er m ess a g e (Tw i l i o ma r k u p) + M e t a co n v ers a t i o n f ees . U t i l i t y m ess a g es w i t hin 24 h rser v i ce w in d o w a re f reeo f M e t a^{'} s t e m pl a t e f ee . F orser v i ceco n v ers a t i o n s t h e t o t a l i sro ug h l y$ 0.01-0.03 per message depending on template vs free-form.
Estimate per conversation (6 messages avg): ~$0.06-0.18.

Scale	Assumed messages/mo	Monthly WhatsApp cost
100 properties	1,200	GBP 10-17
500 properties	6,000	GBP 48-86
1,000 properties	12,000	GBP 96-173
3,000 properties	36,000	GBP 288-518

Assumption: 12 messages per property per month (both directions combined). WhatsApp is the primary channel.

4. Twilio — SMS

What it is: SMS notifications to landlords, vendors, and tenants (e.g., issue updates, vendor assignments).
How it scales: Per-message.
Current pricing: ~$0.046 per outbound SMS (UK).
Estimate: 3-5 SMS per issue (landlord notification, vendor assignment, tenant update).

Scale	Assumed SMS/mo	Monthly SMS cost
100 properties	300	GBP 11
500 properties	1,500	GBP 55
1,000 properties	3,000	GBP 110
3,000 properties	9,000	GBP 331

Assumption: ~1 issue per property per month with 3 SMS notifications per issue.

5. VAPI — Voice AI

What it is: AI-powered phone line for tenants to call and report issues verbally.
How it scales: Per-minute of call time.
Current pricing: $0.13-0.31 per minute all-in (VAPI platform + STT + LLM + TTS + telephony).
Estimate per call: 3-5 minutes average = $0.39-1.55 per call.

Scale	Assumed calls/mo	Monthly voice cost
100 properties	20	GBP 6-25
500 properties	100	GBP 31-124
1,000 properties	200	GBP 62-248
3,000 properties	600	GBP 187-744

Assumption: ~20% of properties generate 1 voice call per month. Voice is a secondary channel behind WhatsApp.

6. Hosting & Database

What it is: The platform itself — web app hosting and database.
How it scales: Mostly fixed costs with step-ups at certain thresholds.
Supabase Pro: $25/ m o ba se . I n c l u d es 8 GB d a t aba se, 100 K M A U s . C o m p u t e u p g r a d e l ik e l y n ee d e d a t sc a l e ($ 110/mo for dedicated 2-core at 1,000+ properties).
Vercel Pro: $20/mo per team member. Usage-based beyond included credits.

Scale	Supabase	Vercel	Total hosting
100 properties	GBP 20	GBP 16	GBP 36
500 properties	GBP 20-40	GBP 16-32	GBP 36-72
1,000 properties	GBP 88	GBP 24-48	GBP 112-136
3,000 properties	GBP 88-200	GBP 40-80	GBP 128-280

Note: These are the costs that scale most slowly. They are largely fixed until we hit database compute limits (~8GB RAM threshold).

7. SendGrid — Email

What it is: Email notifications and potentially email-based issue intake.
How it scales: Per-email. Free tier retired May 2025.
Current pricing: Essentials plan from $19.95/mo for up to 50K emails.
At scale: Even at 3,000 properties, monthly email volume is likely under 50K. Stays on Essentials.

Scale	Monthly email cost
All scales (up to 3,000)	GBP 16

Verdict: Flat cost, not a scaling concern.

Cost-Per-Property Estimate

This table combines all cost centres into a per-property monthly estimate. All figures in GBP.

Cost Centre	Per-Property (100 props)	Per-Property (500 props)	Per-Property (3,000 props)	Scales With	Confidence
LLM (Claude + OpenAI)	0.12-0.36	0.12-0.36	0.12-0.36	Messages	Medium — depends heavily on conversation volume
WhatsApp (Twilio)	0.10-0.17	0.10-0.17	0.10-0.17	Messages	Medium — depends on channel mix
SMS (Twilio)	0.11	0.11	0.11	Issues	Medium — depends on notification design
Voice (VAPI)	0.06-0.25	0.06-0.25	0.06-0.25	Call minutes	Low — voice adoption is uncertain
Hosting (Vercel)	0.16	0.03-0.06	0.01-0.03	Fixed/stepped	High — known pricing
Database (Supabase)	0.20	0.04-0.08	0.03-0.07	Fixed/stepped	High — known pricing
Email (SendGrid)	0.16	0.03	0.005	Fixed	High — known pricing
Embeddings (OpenAI)	~0	~0	~0	Documents	High — negligible
TOTAL	0.91-1.41	0.49-1.06	0.43-0.99

Key Takeaway

At 3,000 properties, the estimated platform cost is GBP 0.43-0.99 per property per month before any margin. The fixed costs (hosting, database) become negligible per-property at scale. The variable costs (LLM, messaging, voice) dominate and scale linearly.

Important caveats:

These estimates assume moderate tenant engagement (~2 conversations/property/month). HMOs with 6+ tenants per property could be 3-5x higher.
Voice costs are the most uncertain — if voice adoption is high, it becomes the single largest cost.
All per-message costs are approximate and subject to provider pricing changes.
No cost is included for team salaries, support overhead, or Envo’s own operational costs.

Scaling Limits & Bottlenecks

What breaks first?

Priority	Bottleneck	Trigger Point	Fix	Cost
1	Database compute	~500-1,000 properties (complex queries with RLS joins)	Upgrade Supabase compute tier	GBP 88-200/mo
2	pgvector search latency	~50,000+ document chunks (semantic search slows)	HNSW index tuning, or migrate to dedicated vector DB (Qdrant/Pinecone)	GBP 50-200/mo
3	WhatsApp throughput	Meta rate limits on shared Twilio number	BYOAK (customer uses own Twilio) or multiple Twilio numbers	Customer-funded
4	LLM rate limits	High concurrent conversations	Multi-provider fallback already designed (Claude → OpenAI → Kimi → GLM)	No extra cost
5	Single-DB multi-tenancy	Compliance requirements from enterprise clients	Schema-per-tenant migration (ADR-001 acknowledges this risk)	Significant engineering

What does NOT break?

Vercel — Serverless, auto-scales. No action needed.
Supabase Auth — 100K MAU included. We will not hit this with 3,000 properties (landlords + staff only; tenants do not authenticate).
SendGrid — 50K emails/mo on Essentials is plenty.
Embedding storage — pgvector handles millions of vectors with proper indexing.

Pre-scaling investments needed

Database indexing strategy — Need proper indexes on org_id, property_id, status columns before scaling past ~500 properties. (Designed but should be load-tested.)
Connection pooling — Supabase’s built-in PgBouncer, but may need tuning for concurrent serverless function calls.
Cost monitoring — No spend tracking is currently in place (noted as a gap in Costs & Usage doc). Need alerts before scaling.
Load testing — System has not been deployed to production yet. Need baseline performance data.

BYOAK Opportunity

ADR-014 defines a three-tier white-label model (Basic / Premium / Partner) where larger customers bring their own API keys.

Which costs can be shifted?

Cost Centre	Shiftable via BYOAK?	Available From	Impact
LLM (Claude/OpenAI)	Yes	Premium tier	Shifts the largest variable cost (~35% of total)
SendGrid (email)	Yes	Premium tier	Minor savings but adds customer control
Twilio (WhatsApp + SMS)	Yes	Partner tier	Shifts second-largest variable cost (~25% of total)
VAPI (voice)	Yes	Partner tier	Shifts most uncertain cost
Supabase (database)	No	—	Shared infrastructure, cannot split
Vercel (hosting)	No	—	Shared infrastructure, cannot split

What this means for pricing

At the Premium tier (BYOAK for LLM + SendGrid), Envo’s cost-per-property drops to roughly GBP 0.25-0.55 because the LLM spend — our biggest variable cost — moves to the customer’s own Anthropic/OpenAI account.

At the Partner tier (BYOAK for everything), Envo’s cost-per-property drops to approximately GBP 0.03-0.10 — essentially just our share of hosting and database.

The BYOAK pricing implication

BYOAK customers cost Envo significantly less to serve, but they also expect a lower platform fee since they are paying their own API bills. The pricing model needs to reflect this:

Basic tier: Envo absorbs all costs. Higher per-property fee. Simpler for small landlords.
Premium/Partner tier: Customer pays own API bills. Lower platform fee. But Envo still captures value through the software itself.

This is a strong scaling story: as customers get larger, the most expensive costs shift to them, and Envo’s margin on infrastructure improves.

Open Questions

These need answers before we can confidently set pricing:

Must-answer for pricing

What is the actual conversation volume per property per month? Our estimates assume ~2 conversations/property/month. HMOs with 6+ tenants could be much higher. We need real data from a pilot.
What is the channel mix? The cost difference between WhatsApp-only and WhatsApp+Voice is significant. Voice is 5-10x more expensive per interaction than chat.
What does “included” mean? If we do a hybrid model (base fee + usage), what is a reasonable “included” volume before overage kicks in?
SendGrid free tier is gone. The Costs & Usage doc assumed a free tier for email. SendGrid retired it in May 2025. Minimum cost is now ~GBP 16/mo (Essentials). Minor, but needs updating.

Should-answer before scaling

No production deployment yet. All performance assumptions are theoretical. We need to deploy and load-test before committing to scale targets.
No cost tracking in place. We cannot currently measure actual per-conversation or per-property costs. This should be instrumented before pilot.
Database compute upgrade threshold. At what property count does Supabase Pro base compute become insufficient? Need load testing with realistic data volumes.
VAPI vs Retell decision. The docs reference both. Pricing and capability differ. Need a final decision on voice provider before costing.
Multi-region. All current architecture assumes a single Supabase region. If we expand beyond the UK, we need to consider data residency and latency.
Tenant density per property. An HMO with 8 tenants generates fundamentally different load than a single-let. Pricing per-property may need a “per-unit” modifier.

Nice-to-know

Competitive pricing. What do Fixflo, Arthur Online, and Goodlord charge per property? This caps what we can charge regardless of our costs.
LLM price trajectory. Claude and OpenAI prices have been falling. Our per-query cost could halve within 12 months. Should we price aggressively now?

Summary Table: Cost at Scale

	100 properties	500 properties	1,000 properties	3,000 properties
Total monthly platform cost (GBP)	91-141	245-530	530-1,130	1,290-2,970
Per-property (GBP/mo)	0.91-1.41	0.49-1.06	0.53-1.13	0.43-0.99
Per-property with BYOAK Premium (GBP/mo)	—	—	0.30-0.55	0.25-0.55
Per-property with BYOAK Partner (GBP/mo)	—	—	0.05-0.15	0.03-0.10
Biggest cost driver	Hosting (fixed)	LLM + Twilio	LLM + Twilio	LLM + Twilio

Bottom line: The infrastructure costs are low. Even at the high end, GBP 1/property/month gives us excellent margin room if the per-property price is GBP 5-15+. The real question is not “can we afford to scale?” but “what conversation volume should we include in the base price?”

Prepared for Session 1: Product Capability & Packaging. All estimates based on published provider pricing as of February 2026 and assumptions documented inline. No production usage data is available yet.

EHQ Brain

Explorer

capacity-analysis