Agente IA 24/7 = custo explode (R$ 10k → R$ 50k/mês sozinho)

Agente IA 24/7 = custo explode (R$ 10k → R$ 50k/mês sozinho)

Você tem SaaS.

Seu SaaS: agente IA (atendimento ao cliente, automação).

Você pensa:

"Agente IA é bom (handles conversations).

Agente IA deveria rodar 24/7 (sempre disponível pra customers).

Agente 24/7 = melhor UX (customer can chat anytime).

Agente 24/7 = melhor ROI (handle more conversations).

Agente 24/7 = competitive advantage (my competitor doesn't have this).

I should deploy agente 24/7 (all the time, every day).

Customers win (always have support).

I win (handle more volume, lower support cost).

Everyone wins (right?)."

You deploy agente 24/7:

"Agente is always on (running in background).

Customers can chat anytime (3am, Sunday, holiday).

Agente responds instantly (no waiting for support team).

I'm happy (competitive advantage).

Customers are happy (always available).

What could go wrong?"

Month 1 bill arrives:

"API calls last month: 1 million

Cost per 1k tokens: R$ 0.01 (OpenAI GPT-4)

Average response: 500 tokens

Total tokens: 500 million

Total cost: R$ 5k

That's reasonable (expected, acceptable)."

Month 2 bill:

"API calls: 2 million (volume grew)

Total tokens: 1 billion

Total cost: R$ 10k

Hmm, cost doubled (but volume grew, so proportional)."

Month 3 bill:

"API calls: 5 million (volume exploded)

Total tokens: 2.5 billion

Total cost: R$ 25k

Wait, that's R$ 25k on JUST API calls.

Adding infrastructure (hosting, monitoring): +R$ 5k

Total agente cost: R$ 30k/month

My agente is EXPENSIVE.

Monthly customers paying me: R$ 100k

Agente cost: R$ 30k (30% of revenue)

Agente margin: 70% (acceptable? or too high cost?)

Wait, if agente goes down, I lose customers (competitive necessity).

So agente cost is FIXED (can't turn it off).

If I turn agente off 24/7, customers leave.

So I'm stuck (paying R$ 30k whether I like it or not).

This is a problem."

Recent discovery (Google I/O 2026):

"Google launches Gemini Spark (24/7 AI assistant).

"Feature: Agente is always-on (runs continuously, handles tasks anytime).

"Review: Agente is useful (handles summaries, planning, tasks).

"But: No one talks about cost (what does 24/7 agente cost to run?).

"Hidden truth: Always-on agente = always-on cost (R$ 10k+ per month for small SaaS)."

You realize:

"Google can afford 24/7 agente (has infinite money, absorbs cost).

But can I (small SaaS)?

If agente costs R$ 30k/month, and I make R$ 100k/month:

I can afford it (30% margin is tight but survivable).

But if I have 10 agentes (for 10 products):

10 × R$ 30k = R$ 300k/month

My revenue: R$ 100k

I'm bankrupt (cost > revenue).

So 24/7 agente doesn't scale.

At some point, agente cost becomes unsustainable.

How do I solve this?"

---

O problema (always-on agente = always-on cost)

Why 24/7 agentes are expensive (and costs explode)

COST BREAKDOWN: 24/7 Agente (always-on)

1. API CALLS (OpenAI GPT-4)

   • Agente runs continuously (24 hours/day)
   • Even if zero customers chatting, agente is polling (checking for messages)
   • Polling cost: ~R$ 0.001 per check
   • 10,000 checks/day (every 9 seconds) = R$ 10/day = R$ 300/month
   • Customer conversations: +1,000 conversations/month = R$ 5k/month
   • Background tasks: +R$ 1k/month
   • Total API cost: ~R$ 6.3k/month

2. INFRASTRUCTURE (hosting, monitoring, db)

   • Server to run agente: R$ 500/month (small server)
   • Monitoring (uptime tracking, alerts): R$ 200/month
   • Database (logging conversations, metrics): R$ 300/month
   • Total infrastructure: ~R$ 1k/month

3. HUMAN OVERSIGHT (review, debugging, optimization)

   • QA person (10 hours/week reviewing agente): R$ 2k/month
   • On-call engineer (if agente crashes at 3am): R$ 2k/month
   • Total human cost: ~R$ 4k/month

TOTAL AGENTE COST: R$ 6.3k + R$ 1k + R$ 4k = R$ 11.3k/month

But volume grows:

• Month 1: 1k conversations = R$ 6.3k API + R$ 1k infra + R$ 4k human = R$ 11.3k
• Month 2: 5k conversations = R$ 15k API + R$ 1k infra + R$ 4k human = R$ 20k
• Month 3: 20k conversations = R$ 50k API + R$ 1k infra + R$ 4k human = R$ 55k
• Month 4: 50k conversations = R$ 100k API + R$ 1k infra + R$ 4k human = R$ 105k

At month 4, agente ALONE costs R$ 105k/month.

If your SaaS revenue is R$ 150k, agente is 70% of revenue (unsustainable).

---

WHY ALWAYS-ON IS EXPENSIVE:

1. Polling cost (even with zero customers)

   • Agente checks for messages every few seconds
   • Polling = API call
   • API call = cost
   • Example: Check 10,000 times/day = R$ 10/day = R$ 300/month
   • That's just IDLE cost (no customers chatting)

2. Response gets longer (agente adds features)

   • Month 1: Agente responds with 100 tokens (cheap)
   • Month 3: Agente responds with 500 tokens (5x more expensive)
   • Why? You added features (context, formatting, links)
   • Result: Cost per response grows (even if volume stays same)

3. Background tasks accumulate

   • Agente summarizes conversations (background task)
   • Agente sends alerts (background task)
   • Agente updates customer profile (background task)
   • Each task = API call
   • Tasks accumulate (cost grows)

4. Human overhead doesn't scale

   • 1 agente needs 1 person to monitor (R$ 2k/month)
   • 10 agentes needs... still ~1 person (but busier, less reliable)
   • At some point, you need 2 people (R$ 4k/month)
   • Human cost grows with complexity, not linearly with volume

5. Agente becomes mission-critical

   • Agente is always-on (customers depend on it)
   • If agente goes down, customers notice (bad UX)
   • You need redundancy (backup agente running, just in case)
   • Redundancy = 2x cost (two agentes running at same time)

---

REAL-WORLD EXAMPLE: SaaS with Always-On Agente

Scenario:

• You're a small SaaS (R$ 100k/month revenue)
• You deploy agente 24/7 (to handle customer support)
• Agente handles 5k conversations/month (30% of support volume)

Month 1:

• API cost: R$ 5k (conversations)
• Polling cost: R$ 300 (idle checks)
• Infrastructure: R$ 1k
• Human monitoring: R$ 2k
• Total: R$ 8.3k/month
• Profit impact: -8.3% of revenue
• Verdict: Acceptable (agente saves support team R$ 10k/month, so net positive)

Month 6 (volume grew):

• API cost: R$ 15k (conversations grew)
• Polling cost: R$ 300 (same)
• Infrastructure: R$ 2k (scaled up to handle load)
• Human monitoring: R$ 2k (still one person, but overworked)
• Total: R$ 19.3k/month
• Profit impact: -19.3% of revenue
• Verdict: Getting expensive (agente is still saving R$ 20k/month, but margin is tighter)

Month 12 (volume exploded):

• API cost: R$ 50k (conversations exploded)
• Polling cost: R$ 300 (same)
• Infrastructure: R$ 3k (scaled more)
• Human monitoring: R$ 4k (now needs 2 people)
• Total: R$ 57.3k/month
• Profit impact: -57.3% of revenue
• Verdict: EXPENSIVE (agente is now 50% of revenue, unsustainable)
• Problem: You can't turn agente off (customers expect it)
• Stuck: Agente costs too much, but you can't remove it

A solução (manage always-on cost)

Strategy 1: Don't run 24/7 (scheduled agente)

OPTION: Run agente only during peak hours (not 24/7)

Setup:

• Agente runs 9am-6pm (business hours) = 9 hours/day
• Agente is offline 6pm-9am (after hours) = 15 hours/day
• After-hours: Customers get auto-reply ("support team is sleeping, we'll respond tomorrow")
• Cost: 9/24 = 37.5% of 24/7 cost

Benefit:

• Cost: Reduce agente cost by 62.5%
• Example: R$ 50k/month 24/7 → R$ 18.75k/month scheduled
• Simplicity: Easier to manage (no night-shift monitoring needed)
• Reliability: Simpler = fewer bugs (less complexity)

Disadvantage:

• UX: After-hours customers don't get support (worse experience)
• Competition: Competitor with 24/7 agente has advantage
• Lost sales: After-hours inquiries might go to competitor
• Customer frustration: "Why is support offline at 6pm?"

When to use:

• Your customers are business hours (9-6 timezone)
• After-hours volume is low (<5% of total)
• Cost is critical (need to cut agente expense)
• Can afford to lose after-hours customers

Example:

Your SaaS: HR software (business customers in Brazil)

Business hours (9am-6pm): 95% of conversations After-hours (6pm-9am): 5% of conversations

Strategy: Run agente 9am-6pm, offline after-hours

Cost:

• 24/7 agente: R$ 50k/month
• Scheduled agente (9-6): R$ 18.75k/month
• Savings: R$ 31.25k/month (62.5%)

Downside:

• After-hours customers get auto-reply
• Some after-hours customers might switch to competitor
• Estimated lost revenue: R$ 5k/month (10% of after-hours)
• Net savings: R$ 31.25k - R$ 5k = R$ 26.25k/month

Strategy 2: Smart polling (reduce idle cost)

OPTION: Don't poll constantly, use webhooks (event-driven)

Setup:

• Old way: Agente checks every 5 seconds ("any messages?") = 17,280 checks/day
• New way: Customer sends message → webhook triggers agente (event-driven)
• Agente only activates when needed (not idle polling)
• Cost: 50-90% reduction in idle cost

Benefit:

• Cost: Reduce idle polling cost by R$ 300→R$ 50/month
• Efficiency: Agente only runs when there's work (not wasting resources)
• Scalability: Webhook-based scales better (no overhead)

Disadvantage:

• Complexity: Need to implement webhooks (dev work)
• Latency: Small delay when agente starts (milliseconds)
• Depends on platform (some platforms don't support webhooks)

When to use:

• Have developer resources (to implement webhooks)
• Idle polling is significant cost (check your bill)
• Can afford small latency increase
• Want to scale without increasing cost

Example:

Before (polling):

• Agente checks every 5 seconds (17,280 checks/day)
• Each check = 1 API call = R$ 0.000001 each = R$ 0.02/day
• 30 days = R$ 0.60/month (negligible)

Wait, polling is actually cheap? Let me recalculate...

Actually, "polling" usually means checking your infrastructure (not API calls). So polling cost is infrastructure (server cycles), not API cost.

Better example:

Before (agente running constantly):

• Agente server running 24/7 = R$ 500/month
• Agente is idle 80% of time (no conversations)
• Wasted cost: R$ 400/month (on idle capacity)

After (event-driven webhooks):

• Agente server spins up only when customer messages (serverless)
• Pay only for actual execution time
• Cost: R$ 100/month (only active time)
• Savings: R$ 400/month

Strategy 3: Tiered agente (expensive for complex, cheap for simple)

OPTION: Use cheap agente for simple tasks, expensive agente for complex

Setup:

• Cheap agente (GPT-3.5): Handles FAQ, routing = R$ 0.0005 per 1k tokens
• Expensive agente (GPT-4): Handles complex issues = R$ 0.01 per 1k tokens
• Conversation flow: Customer question → cheap agente → "Is this simple?" → route to cheap or expensive
• Cost: 70% cheap agente (FAQ) + 30% expensive agente (complex) = 40% cost savings

Benefit:

• Cost: Use cheap agente for 70% of conversations (huge savings)
• Quality: Complex issues still get good agente (GPT-4)
• Hybrid: Best of both (cost-effective + quality)

Disadvantage:

• Complexity: Need routing logic (which questions go to cheap vs expensive)
• Quality risk: Cheap agente might fail on edge cases
• Latency: Routing adds delay (milliseconds)

When to use:

• Have high conversation volume (savings justify complexity)
• Most conversations are simple (FAQ, routing)
• Complex issues are minority (can afford expensive agente for them)
• Want to scale without cost explosion

Example:

Your SaaS: e-commerce platform

Conversation types:

• "What's your return policy?" (simple FAQ) = 50% of conversations
• "How do I track my order?" (simple FAQ) = 30% of conversations
• "I want to return my item but there's an issue" (complex) = 20% of conversations

Strategy:

1. Customer asks question
2. Cheap agente (GPT-3.5) tries to answer (FAQ)
3. If confidence < 50%, route to expensive agente (GPT-4)
4. Expensive agente handles complex issue

Cost:

• 80% conversations use cheap agente: 80% × R$ 0.0005 = R$ 0.0004 per token
• 20% conversations use expensive agente: 20% × R$ 0.01 = R$ 0.002 per token
• Average: R$ 0.0024 per token
• 24/7 with GPT-4 only: R$ 0.01 per token
• Savings: 76% cost reduction

Example numbers:

• GPT-4 only (24/7): R$ 50k/month
• Tiered agente (cheap+expensive): R$ 12k/month
• Savings: R$ 38k/month

Strategy 4: Hybrid (scheduled + tiered + smart polling)

OPTION: Combine all three strategies

Setup:

1. Run 24/7 but scheduled (business hours full agente, after-hours minimal)
2. Use tiered agente (cheap for FAQ, expensive for complex)
3. Use event-driven webhooks (no idle polling)

Result:

• Business hours (9am-6pm): Full-featured agente (tiered, cheap+expensive)
• After-hours (6pm-9am): Minimal agente (only very cheap, FAQ-only)
• Idle cost: Eliminated (event-driven)

Cost:

• 24/7 GPT-4 agente: R$ 50k/month
• Hybrid approach: R$ 8k/month
• Savings: 84%

Trade-offs:

• After-hours customers get limited support (FAQ only)
• Business hours customers get full support (cheap + expensive agente)
• Overall cost is low (sustainable)
• Overall quality is good (best of both)

Conclusão: Always-on agente is expensive (plan carefully)

**O que você precisa saber:

1. 24/7 agente sounds great (customer wins) but costs explode (you lose)

   • Always-on agente = always-on cost (R$ 10k+ minimum per month)
   • Cost grows exponentially (as volume grows, cost grows faster)
   • At some point, cost > revenue (unsustainable)
   • Google can afford 24/7 agente (infinite money)
   • But can you? Probably not
   • Lesson: 24/7 agente is competitive move (but financially dangerous)

2. Agente cost has multiple components (not just API calls)

   • API calls: ~60% of cost (GPT-4 responses)
   • Infrastructure: ~10% of cost (servers, databases, monitoring)
   • Human oversight: ~30% of cost (monitoring, debugging, optimization)
   • Total: R$ 10k-50k+ per month (depending on volume)
   • Lesson: You can't just look at API bills (human cost hides)

3. Idle cost is real (agente running when no customers)

   • Even with zero conversations, agente is running
   • Monitoring, polling, background tasks = cost
   • Idle cost = R$ 500-1k per month (minimum)
   • Lesson: Always-on = cost, even at idle

4. You can reduce cost by 50-84% with smart strategies

   • Strategy 1: Scheduled agente (only business hours) = -62.5% cost
   • Strategy 2: Event-driven (no polling) = -10-20% cost
   • Strategy 3: Tiered agente (cheap for FAQ) = -40% cost
   • Strategy 4: Hybrid (all three) = -84% cost
   • Lesson: Cost is manageable (not destiny, requires planning)

5. 24/7 agente is a choice (not a requirement)

   • Customers PREFER 24/7 (but might tolerate business hours)
   • You can start with scheduled agente (cheaper)
   • Upgrade to 24/7 later (when revenue justifies cost)
   • Or hybrid (24/7 but tiered/smart polling)
   • Lesson: Start lean (scheduled agente), scale when you can afford it

Na OpenClaw, ajudamos SaaS a:

• ASSESS agente cost (how much is agente REALLY costing?)
• PLAN agente strategy (24/7? Scheduled? Hybrid?)
• IMPLEMENT cost-saving tactics (tiered agente, webhooks, smart polling)
• MONITOR agente cost (track, alert when costs spike)
• OPTIMIZE agente ROI (measure revenue per agente cost)
• SCALE sustainably (grow agente without bankruping company)

Resultado: Seu agente IA é SEMPRE-ON quando precisa + ECONOMICO (custo controlado) + ESCALAVEL (cresce com receita) + PROFITAVEL (ROI positivo).

Seu agente 24/7 custa R$ 50k/mês?

Ou você já otimizou pra scheduled + tiered + webhooks?

Otimizar custo do agente IA sempre-on →