Bot, AI agent, and human support: when to use each (and how much it costs)

Every time someone asks me "is it worth putting AI on my WhatsApp?", I throw the question back: AI for what?

Because "AI" became such a wide umbrella it lost meaning. People are calling "AI agent" a tiny robot that replies with dressed-up if/else, and people are selling "100% human support" while the rep copies and pastes a template saved in Notion. Both exist, but they're not opposites — they're different points on the same spectrum.

This text is for you, trying to decide how to handle WhatsApp in 2026 without burning money. We'll break down:

• What each mode is (bot, AI agent, human) — no technical fluff
• When each makes sense and when it's silly
• How much each truly costs, with real numbers
• How to combine the three — because the right answer almost never is "just one"
• How AI hubs like Glama change the cost equation

If you only have 30 seconds: bot for the queue, AI for intelligence, human to close. The rest is detail.

The three modes, demystified

Before comparing, let's align what each one is. In practice, not in conference theory.

Bot (traditional chatbot)

It's a tree flow: the customer clicks "1 - Sales", "2 - Support", "3 - Other". Each path leads to another menu, or to a canned reply, or to a transfer to a human.

Real advantages:

• Practically zero cost after configured (you pay the platform, not each message)
• Instant response — no model latency
• 100% predictable behavior — tested once, you know what'll happen

Honest limitations:

• Freezes off-script. Customer asked something you didn't anticipate? Back to menu, or asks for a human.
• Frustrates if the script is bad. Everyone has hated a "press 7 for other options" menu.
• Doesn't scale in variety. If your business gets 200 different questions a day, you can't map all of them in a tree.

AI agent (LLM with RAG and Function Calling)

It's a language model (like GPT-4, Claude, Gemini) with two important extensions:

• RAG (Retrieval-Augmented Generation): the AI consults your knowledge base (manual, FAQ, warranty policy, catalog) before replying. This eliminates the classic "hallucination" problem — the AI doesn't invent, it cites what's in your documents.
• Function Calling: the AI doesn't just chat, it performs actions. "Check my order status", "schedule a meeting tomorrow at 3pm", "generate a duplicate invoice". The AI calls a function in your system and brings the result back to the conversation.

Real advantages:

• Handles variety. The customer asks something you've never seen? The AI gets it.
• Scales 24/7 without getting tired or grumpy at 11pm on a Friday.
• Personalizes by context. If the customer is a new lead, it pulls a sales script. If a paying customer, it pulls support. All in the same agent.

Honest limitations:

• Costs per token (chunk of word that goes in or out of the model). Long conversation = bigger bill.
• Has latency. 2 to 5 seconds per response. To an impatient customer, it feels slow.
• Can be wrong. Even with well-built RAG, in 1-2% of interactions the AI gives a reasonable-but-wrong reply. You need a process to catch and correct.

Human support

It's what it always was: a flesh-and-bone person reading a message and replying.

Real advantages:

• Genuine empathy. An angry customer wants to talk to a person, not a calm robot.
• Negotiation. Hammering price, offering a one-off discount, making creative concessions — humans handle this.
• Closing. Complex sales conversation, with objections, requires reading between the lines.

Honest limitations:

• Expensive. Salary, training, payroll taxes, weekend shifts.
• Doesn't scale well. Each rep handles X simultaneous conversations. To double output, you double headcount.
• Inconsistent. Humans get tired, forget, have bad days. Service quality varies.
• Not 24/7. A customer at midnight on Sunday waits until Monday.

When to use each one (the honest table)

Summarizing where each one shines — and where it's waste:

Situation	Bot	AI Agent	Human
Initial greeting and triage	✅ Ideal	⚠️ Expensive for this	❌ Waste
Standard FAQ (price, hours, link)	✅ Ideal	✅ Works	❌ Waste
Varied question about product/service	❌ Freezes	✅ Ideal	⚠️ Expensive
System lookup (order status, duplicate invoice)	⚠️ Possible but rigid	✅ Ideal (function calling)	⚠️ Slow
Price/terms negotiation	❌ Can't	❌ Shouldn't	✅ Ideal
Complaint or serious problem	❌ Worsens	⚠️ Detects and transfers	✅ Ideal
24/7 support outside business hours	✅ Works	✅ Ideal	❌ Unfeasible
Closing a complex sale	❌ No	❌ No	✅ Ideal
Post-sale for paying customers	⚠️ Frustrates	⚠️ Acceptable	✅ Ideal

Notice the pattern: all three columns have green marks. None of the three is best at everything. The right question isn't "which one to use?", it's "which one to use when?".

How much each one costs (with real numbers)

I'll estimate costs for a common scenario: SMB receiving 3,000 conversations per month on WhatsApp, averaging 8 messages per conversation (4 from the customer, 4 from the company).

Cost of a traditional bot

• Chatbot platform: $20 to $100/month fixed, regardless of volume
• Initial setup: 5 to 20 hours of work (you or a marketing person)
• Maintenance: 2 to 5 hours/month tweaking flows

Cost per conversation: practically $0 after setup. The bot is the cheapest agent that exists — as long as the script fits its capabilities.

Cost of an AI agent (without optimization)

This depends a lot on the model. Two examples:

Scenario A — premium model only (GPT-4 or Claude Opus):

• ~3,000 tokens per conversation (input + output)
• Average price: ~$0.03 per 1,000 tokens (input and output combined) — current numbers in OpenAI's official pricing and Anthropic's official pricing
• Cost per conversation: ~$0.09
• Total monthly cost: 3,000 × $0.09 = $270/month

Scenario B — cheaper model (GPT-3.5, Claude Haiku, Gemini Flash):

• Same structure, but ~10x cheaper
• Cost per conversation: ~$0.01
• Total monthly cost: $30/month

A cheap model answers 90% of questions as well as the premium one. The difference shows in complex reasoning, creative negotiation, long text generation. If the AI is just to qualify and answer questions, the cheap model handles it.

Cost of an AI hub (Glama and similar)

Here's the part many people don't calculate. When you integrate AI through a single platform, you pay only for usage without signing a contract with each provider.

The Glama AI Hub — which CRM Whats Pro uses — works like this:

• Unified API for 100+ models (OpenAI, Anthropic, Google, open source models)
• You pay the same price as each model, with no hidden markup (in most cases)
• No need to open accounts at 5 places or manage 5 API keys
• Automatic fallback: if a model goes down, the hub routes to another automatically — zero downtime
• Switch models without changing code: today you're on GPT, tomorrow you want to test Claude. Flip a switch, not start a project.

The big play of using a hub isn't the price per token — it's the flexibility of combination. You can:

1. Cheap model for 80% of messages (greeting, FAQ, simple qualification)
2. Premium model for the 20% that matter (negotiation, objection handling, closing message)

CRM Whats Pro lets you make this choice in the panel — you define which model to use for each type of agent. Practical result: AI bill drops by half or more, without losing quality where it matters. (Step-by-step setup of agents, model selection via Glama, and RAG configuration is documented at help.crmwhatspro.com, with videos on our YouTube channel.)

Cost of human support

• Junior salesperson: ~$700 + benefits = ~$1,100/month
• Capacity: 200 to 400 conversations/month per rep (depends on complexity)
• Cost per conversation: $2.75 to $5.50

Compare with AI: a human is 300 to 500 times more expensive per conversation. But a human is the only one who closes complex sales. That's why the equation isn't "who's cheaper", it's "who should handle what".

The combination that works (and that CRM Whats Pro implements)

The setup that delivers most results for SMBs is a triple layer:

Customer sends message
        ↓
[1] Bot: greeting and triage ($0)
        ↓
[2] AI Agent: qualification, FAQ, automatic actions ($0.01–$0.09)
        ↓
[3] Human: only when AI detects closing signal or problem ($2.75–$5.50)

The AI sits between the bot and the human. It does three things:

1. Covers what the bot can't handle (varied questions, context)
2. Filters before the human — only passes to the rep what's actually going to close
3. Cites what's in the base (RAG) and executes actions (function calling) without needing a human

And here Glama enters as a silent optimizer: each agent can use the right model for each task.

• Initial qualification agent: cheap model (GPT-3.5/Haiku/Gemini Flash) — spends little, replies fast
• Objection/closing agent: premium model (GPT-4/Claude Opus) — worth it for the message that decides if the sale happens
• Post-sale agent: mid-tier model (Claude Sonnet/GPT-4-mini) — balance between cost and quality

Each of these agents acts at a different stage of the funnel — to better understand how a lead moves between them, see the 7 stages of the WhatsApp sales funnel.

With this configuration, it's reasonable to expect:

• 80% of conversations resolved without a human
• Humans only see hot leads or complex problems — where they're irreplaceable
• AI cost in the range of $40 to $120/month for 3,000 conversations/month
• First response time < 10 seconds (because it's AI)
• 24/7 availability

The most common mistakes when implementing AI on WhatsApp

I'll list what I see go wrong — so you don't fall into the same holes.

Mistake 1: turning on AI before organizing human support

AI on top of a messy process amplifies the mess. If your team already doesn't know who answers what, throwing AI in the middle just makes it worse — because now there's another "rep" (the AI) writing messages no one reviews. First fix the shared inbox, distribute leads, define the routine. Then turn on AI. (The complete guide on how to sell more on WhatsApp in 2026 covers this organization-of-the-basics part.)

Mistake 2: using the premium model for everything

GPT-4 or Claude Opus answering "what are your business hours?" is throwing money away. Each conversation that could cost $0.01 ends up costing $0.09 — multiplied by 3,000 conversations/month, that's $240 wasted. Set the cheap model as default, premium only where it makes a difference.

Mistake 3: not setting up RAG properly

The AI "hallucinates" because RAG is bad or nonexistent. When you turn on AI without giving it access to your documents (manual, FAQ, pricing policy), it makes things up based on what it learned during training — and almost always misses details about your business. Build RAG first: dump all your institutional content into the knowledge base before activating the AI.

Mistake 4: letting the AI decide when to escalate to a human without clear rules

The AI needs to know exactly when to transfer. Without that, it tries to solve everything (and sometimes makes it worse) or transfers too much (and wastes humans). Good rules:

• Customer mentions "cancel", "complaint", "lawyer" → human immediately
• Customer has received 3 replies and still has questions → human
• Customer says "I want to buy" and the agent is for qualification → human (rep) with full context
• Conversation goes past 15 messages without resolution → human

Mistake 5: not measuring what the AI is doing

You can't optimize what you can't see. Minimum dashboard for AI:

• How many conversations did it handle alone?
• How many did it transfer? Why?
• Average cost per conversation
• Resolution rate (customer left satisfied without needing a human)
• Customer rating (post-conversation NPS, if possible)

Without this, you'll never know if the investment is worth it.

The question that matters: does it make sense for **your** business?

Summarizing the decision:

• You get fewer than 100 conversations/month? Basic bot + human support solves it. AI is overkill.
• You get between 100 and 500/month? Bot + simple AI (cheap model) + human just to close. You can have fast ROI.
• You get 500 to 5,000/month? Triple setup (bot + AI with Glama + human), with models optimized per agent. Big gain here.
• You get more than 5,000/month? AI is mandatory, humans are the exception (negotiation and serious problems). Without AI, your operation collapses before the month ends.

And always, in any range: start small, measure, adjust. AI isn't a magic button — it's an iterative process. The teams that win most are the ones that started with 1 simple agent, measured the result, and refined.

AI won't replace your human salespeople. It will free them from conversations that don't need them — so they can focus on the ones only they can close. And that's where results show up.