AI agents for business: what they can do today, honestly
What an AI agent reliably does for a business in 2026, what still needs a person, what one costs to build and run, and how to spot a demo.
By Sohail Hussain29 October 20268 min read
As of October 2026, an AI agent can reliably answer questions from your own documents, collect and qualify enquiries, look up and update records in your systems, book appointments, route requests and draft replies for a person to approve. It cannot reliably make judgement calls with money or legal consequences, handle an angry customer, or work from documents that contradict each other. The useful ones are built narrowly, tested against real conversations, and hand over to a person the moment they are unsure.
That is the honest state of the technology. This post sets out what agents do well, where they fail, what one costs to build and run, and how to tell a working agent from a demo, drawn from the ones we build as part of our AI agents and chatbots service.
What is an AI agent, as opposed to a chatbot?
A chatbot takes a message and returns a message. An agent takes a message, decides what it needs, calls tools to get it (search your help content, query the order system, check a calendar), acts if it is allowed to (create a ticket, book a slot, send a confirmation), and then replies. The model does the understanding and the deciding; the tools you give it set the limits of what it can do.
That distinction matters for two reasons. An agent is more useful because it can finish a task rather than describe how to. And an agent needs more care because it can do things with consequences, so what it may do, and when it must stop and ask, has to be designed. Chatbot vs AI agent covers the distinction in more depth.
What can an agent do reliably today?
| Task | Reliability as of October 2026 | What it needs |
|---|---|---|
| Answer questions from your documents, policies and product content | High, when the content is current and consistent | Retrieval over your own content; see RAG explained |
| Collect and qualify enquiries (name, need, budget, timing) | High | A clear script and a handover rule |
| Look up order, booking or account status | High | Read access to the system through its API |
| Book, reschedule and cancel appointments | High within set rules | Calendar access and rules on what it may change |
| Create tickets, tasks and records | High | Write access, limited to specific fields |
| Draft replies to emails and messages for a person to send | High for routine, medium for sensitive | A review step |
| Triage and route requests to the right person | High | Categories and examples |
| Summarise long threads, calls and documents | High | Nothing special |
| Explain pricing and quote from a price list | Medium; fine for list prices, poor for negotiated ones | Hard limits on discounts |
| Handle complaints and upset customers | Low | Immediate handover |
| Make refund, credit or exception decisions | Low; should not | A person |
| Give legal, medical or financial advice | Should not | A person, and disclaimers |
The pattern in the table: agents are strong where the answer exists in your content or your systems and the action is routine, and weak where judgement, emotion or exceptions are involved. Which conversations should an AI handle, and which should it not gives the full list we use when scoping.
Where do agents go wrong?
Making things up. A model asked a question it cannot answer from the content it has will sometimes produce a plausible answer anyway. The fix is structural: retrieve relevant content first, instruct the model to answer only from it, say "I don't know" otherwise, and test. How to stop an AI chatbot making things up covers the techniques; none of them is perfect, which is why the next point exists.
Not handing over. An agent that keeps trying when it should pass to a person is worse than no agent. Handover rules should cover: the customer asks for a person, the agent is unsure, the topic is on the restricted list, the customer is upset, or the same question has been asked twice. Human handover sets out the rules and the mechanics.
Inconsistent source content. If your help centre says one thing and your PDF price list says another, the agent will say both, on different days. Agents expose content problems that were already there. Fix the content before the launch, not after.
Too much access. An agent with write access to everything will, eventually, write the wrong thing. Give it the narrowest access that lets it do its job: specific fields, specific actions, limits on amounts, logging of every action. AI agent security covers data, prompts and permissions.
No testing. Demos work because the person demonstrating asks questions the agent answers well. Real customers do not. Before launch, run the agent against two or three hundred real past conversations and score the answers. How to test an AI agent before launch describes the method.
What does an agent cost to build and run?
Build. As of October 2026, a support or enquiry agent on your website or WhatsApp, answering from your content and taking one or two actions, costs ₹2,00,000 to ₹10,00,000 in India or $5,000 to $25,000 elsewhere from a studio, over four to ten weeks. Most of that is content preparation, integration with your systems, testing and the handover flow, not the model. Agents with several integrations, internal knowledge across many document sources, or voice run higher. How much does an AI chatbot cost to build and run breaks the figure down.
Model usage. Models are priced per million tokens of input and output. Across the major providers, as of October 2026, list prices run from roughly $0.10 per million input tokens for the smallest models to $10 per million for the largest, with output tokens priced several times higher; Anthropic's current range, for example, is listed from $0.10 to $10 per million input tokens depending on the model. A support conversation that reads a few pages of your content and replies three times uses a few thousand tokens. At a few thousand conversations a month, the bill is usually ₹5,000 to ₹50,000 or $60 to $600. Our AI API cost calculator turns tokens per request, requests per month and model prices into a monthly figure; the token counter shows how many tokens your content and prompts use; how to estimate AI API costs explains the method. Prompt caching, offered by the main providers, cuts the cost of re-reading the same content substantially.
Channel and hosting. WhatsApp Business API conversations carry per-message charges set by Meta and vary by country; website chat costs nothing beyond hosting. WhatsApp AI chatbot: how it works and what it costs covers that channel.
Is it worth it? Work it out rather than assume. The chatbot ROI calculator takes tickets deflected, handling time and agent cost and returns a monthly saving. For most businesses with a few hundred repetitive enquiries a month, the agent pays back within a year. For a business with fifty enquiries a month, it usually does not, and a good contact page and a fast human reply are the better investment.
Which model should you use?
As of October 2026, OpenAI, Anthropic and Google each offer models at small, medium and large tiers, and all three are capable enough for the tasks in the table above. The choice between them matters less than the choice of tier: a small model handles routing, classification and simple questions at a fraction of the cost; a large model is worth it for complex reasoning over long content. Most of our agents use a small model for most turns and a larger one when a question is hard. OpenAI vs Anthropic Claude vs Google Gemini for business apps compares them on the things that matter in production: data handling, reliability, tool calling and price. The providers' own documentation is the source for current prices and data policies; Anthropic's is at platform.claude.com/docs.
How do you tell a real agent from a demo?
Ask five questions of anyone proposing one.
- What content does it answer from, and what happens when the answer is not there? The right answer involves retrieval over your content and an explicit "I don't know" path.
- What can it do, and what is it forbidden to do? There should be a short list of actions with limits, not "anything you need".
- When does it hand over to a person, and how? The handover should carry the conversation so far, and reach a person who will act.
- How was it tested, and can we see the results? Scored results on real past conversations, not a demo.
- What does it log, and what can we see a month in? Every conversation, every action, resolution rate, handover rate and the questions it could not answer. What a month of AI agent logs taught us shows what that looks like.
Vague answers to any of these mean the agent will embarrass you. Clear answers mean it will take the repetitive half of your enquiries within a month of launch, which is where the value is.
If your first use case is support, how to build a customer support AI agent is the step-by-step guide. If it is internal, internal knowledge assistant: a build guide covers answering staff questions across policies and documents, which is usually the lowest-risk place to start. And if you are weighing a build against Intercom, Tidio or similar, website chatbot: build vs buy sets out when the off-the-shelf option is the right one, because for a small site it often is.
Was this guide useful?