AI agents vs chatbots comes down to one line: a chatbot answers the question, and an AI agent then does the work the answer implies, updating the booking, raising the invoice or changing the page. The reply was never the expensive part, which is why Gartner’s 2023 survey of 5,728 customers found only 14% of customer service issues are fully resolved in self-service.
Every US dollar figure below converts at LKR 328.75, the Central Bank of Sri Lanka’s indicative USD/LKR spot rate for 10 September 2026. This is commercial information, not legal advice.
The parcel question, and the one it could not answer
Picture the owner of a small homeware brand in Colombo 7, three staff, orders arriving from Instagram and the website. She is a hypothetical, not a client. She buys an AI chatbot, and for a fortnight she is delighted. It knows the delivery charge to Kandy. It knows the opening hours. It tells people where their parcel is, because somebody wired it to the courier’s tracking page.
Then a customer types: I gave the wrong address, can you send it to my office instead.
The chatbot writes a polite paragraph saying it will pass the request to the team. That paragraph lands in the same inbox everything landed in before, except now it arrives at half past ten at night with a customer already expecting it to have been sorted. Nothing was removed. A queue was renamed.
She did not buy the wrong product. She bought a product that answers, when the job she wanted done was a job that changes something.
What is an AI agent for business, and what does a lab mean by the word
The word has been stretched until it means whatever is on the invoice, so it helps to take a definition from people who build the models rather than resell them.
Anthropic’s engineering team, in Building effective agents, splits the category in two. “Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.” Both are agentic systems. Only one of them decides anything.
Read that against what most Sri Lankan buyers are shown in a demo. A menu of four options, a knowledge base, a fallback that says a member of our team will get back to you. That is not even a workflow in the sense above. It is retrieval with a friendly voice.
For a business, the practical test is narrower than the lab’s and more useful. Does the thing have permission to write into a system you care about, and is there a named person who approves the write. A chatbot reads. An agent reads, then writes, under supervision. Everything else is packaging, which is the same complaint we made about buying capacity without a brief in You Bought the Contractor. You Never Wrote the Brief.
Four scenarios, and where the work actually sits
Take four jobs that occur in almost every business here, and follow them past the reply.
| Scenario | What a chatbot does | What an AI agent does | Who approves |
|---|---|---|---|
| Customer wants to book or move an appointment | Gives available times from a page, asks the customer to call or fills a form | Reads the live diary, moves the slot, writes the booking, sends the confirmation | Support lead, or nobody once the rule is proven |
| Customer queries an invoice | Explains the payment terms and where to find the invoice | Opens the account, checks what was billed and paid, drafts the credit note or the chase | Partner or manager, every time |
| Customer asks whether an item is in stock | Repeats what the website says, which may be two days old | Reads the stock file, answers with the real number, flags the reorder | Nobody for the answer, owner for the reorder |
| Support ticket arrives about a recurring fault | Answers the ticket, closes it, files nothing | Answers, tags the pattern, adds it to the build list with the count | Product owner |
Three of those four rows involve money, a promise or a commitment, which is why the approval column exists at all. The rule we write into every specification is that nothing generated reaches a customer, an account or a repository until a named person has approved either that exact output or the template it came from. How that gets measured is a separate discipline, and we set it out in what an AI agent’s approval rate means.
What you are actually paying for in a chatbot quote
specshop.dev is an AI consultancy and AI agency in Colombo, Sri Lanka, led by Janaka Ediriweera, Principal AI and Product Management Consultant. Here is the part buyers get wrong on price, and it is not the direction you expect.
Chatbot platforms are cheap to start and priced by conversation. Intercom’s Fin publishes US$0.99 per outcome, with a minimum of 50 outcomes a month, where an outcome is a resolution, a configured handoff to a human, a disqualification, or a qualified prospect at US$9.99. Add the agent-assist seat at US$35 per user a month and the Pro add-on from US$99 a month. So a small Colombo operation might start at about US$49.50 a month, roughly LKR 16,273, and that is genuinely good value for what it is.
Note what a resolution means there: no further help is requested after the last answer. That is a real and reasonable definition, but it is a measure of the conversation ending, not of your work being done. The homeware owner’s address change would count as a handoff, billed, and then done by hand.
An agent is priced the other way round. On our roster you hire an AI agent at US$1,225 a month, LKR 402,700, for the ops and finance, support and sales, web or product lead posts, or US$3,000 a month for the delivery engineer, with a one-off recruitment fee of US$9,500, LKR 3,123,125, covering deployment, connection to your inbox, WhatsApp and tools, a 30-day probation with weekly scorecards, and the first month’s salary. If the approval rate has not improved from its day-one baseline by day 30, you do not confirm the hire and the balance is never invoiced.
Those are not competing quotes for the same thing. One buys deflection at the top of the funnel. The other buys a role: the support and sales agent takes first conversations on WhatsApp and the website around the clock and then asks for the review or the repeat order, the ops and finance agent keeps every VAT, corporate tax and licence date in view and chases the unpaid invoice, the web agent stages price changes and new pages before publishing, the product lead reads every support request and finds the pattern, the delivery engineer builds to a signed specification and opens the pull request.
When a chatbot is enough, honestly
Often. If your incoming questions are dominated by hours, location, delivery charge, price list and where is my parcel, and if the answers live on a page that is already accurate, a chatbot is the correct purchase and an agent is an expensive way to do the same job.
Three conditions make it enough. The answers exist in writing and are kept current. Nothing the customer asks for requires a change in a system. And there is a visible route to a person within two messages, which matters more than any of the technology, because Ren and Zhang, writing in California Management Review in April 2026, report that 53 to 77% of survey respondents have had a bad or frustrating experience with chatbots, and that Gartner found in July 2024 that 64% of customers would prefer companies did not use AI for customer service at all.
If you are on WhatsApp, the same rule applies with sharper edges, and the flows customers accept are narrower than most vendors admit, as we set out in WhatsApp automation in Sri Lanka.
How to test a vendor’s agent in one question
Ask this, exactly: after it answers, what does it change, in which system, and who approves the change before it happens.
A vendor selling a chatbot will answer with channels and languages and a knowledge base. A vendor selling an agent will name a system, a write action and a reviewer, in one sentence, without checking. There is no third answer. If they say it integrates, ask whether the integration reads or writes, because almost every integration in a demo reads.
Two follow-ups close it. What happens when it is not confident, and does that go to a named person or a generic queue. And what does it do that I could not do with a good FAQ page.
What to do this week
Do not shop. For five working days, keep a tally of every incoming message in two columns: answered by information, and needed somebody to change something. Retail owners are usually surprised by how heavily the second column loads after four in the afternoon.
If column one is the bigger pile, buy a chatbot, write the answers properly, and put a human exit two messages in. If column two is the bigger pile, no chatbot will help you, because the reply was never the bottleneck. In that case write one page of rules for the single heaviest task in column two, with three worked examples and the line where a named person must approve. If you would rather have the counting done for you, our workflow mapping sprint is US$2,300, LKR 756,125, for two video sessions and a written report in 48 hours, credited in full against the recruitment fee if you hire within 30 days. We do this as an AI agency in Colombo, and the first thirty-minute call is free, including the answer that you do not need us yet.
Questions people actually ask.
What is the difference between an AI agent and a chatbot?
A chatbot produces a reply and stops, while an AI agent produces the reply and then takes the action it implies, such as moving the appointment, drafting the credit note or updating the price on the page. Anthropic’s engineering guidance draws the same line technically: workflows run through predefined code paths, whereas agents dynamically direct their own processes and tool usage while keeping control of how they accomplish the task.
Is ChatGPT an AI agent?
On its own, as a chat window you type into, it is a very capable assistant rather than an agent for your business, because it has no permission to write into your diary, your accounts or your stock file. It becomes agentic only when it is connected to those systems with defined tools, a written rule for each task and a named person approving the actions that carry money or a promise.
Do I need a chatbot or an AI agent for my business?
Count your incoming messages for a week and split them into questions that information answers and requests that need something changed in a system. If the first pile is bigger, a chatbot is the right purchase and platform pricing starts around US$0.99 per resolved conversation, and if the second pile is bigger you need an agent, because answering faster does not remove the work.
What can an AI agent do that a chatbot cannot?
It can write into the systems you already run: move a booking in the diary, raise or credit an invoice, read the live stock file rather than the website copy of it, stage a page change, tag a recurring fault into the build list. Each of those write actions should sit behind a named approver, which is what separates a supervised agent from an automation that surprises you on a Friday.
How much more does an AI agent cost than a chatbot?
A chatbot platform such as Intercom’s Fin starts at about US$49.50 a month, roughly LKR 16,273, at its 50-outcome minimum of US$0.99 each, while an AI agent on our roster is US$1,225 a month, LKR 402,700, or US$3,000 for the delivery engineer, plus a one-off recruitment fee of US$9,500, LKR 3,123,125. The gap is large because you are buying different things: deflection at the top of the funnel in one case, and a role with a reviewer and a measured 30-day probation in the other.
Sources
- Building effective agents, Anthropic engineering, 2024 — “Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks”, and the grouping of both under “agentic systems”. Fetched 10 September 2026.
- Fin pricing, Fin AI (Intercom), 2026 — US$0.99 per outcome with a minimum of 50 outcomes a month; outcomes defined as resolution (“No further help is requested after Fin’s last answer”), procedure handoff and disqualification, with qualification at US$9.99; Copilot at US$35 per user a month; Pro from US$99 a month; no charge when a conversation is simply passed to the team without an outcome. Fetched 10 September 2026.
- Chatbot Frustration is Real: Hidden Costs and Best Practices, Yuqing Ren and Rongjin Zhang, California Management Review, 28 April 2026 — “only 14% of customer service issues are fully resolved in self-service”, from Gartner’s survey of 5,728 customers conducted in December 2023 and reported in August 2024; “53-77% of survey respondents have had a bad or frustrating experience of interacting with chatbots”; Gartner’s July 2024 finding that 64% of customers preferred that companies didn’t use AI for customer service. Fetched 10 September 2026.
- Daily Indicative USD/LKR Spot Exchange Rates, Central Bank of Sri Lanka, 2026 — the indicative rate of LKR 328.75 per US$1 on 10 September 2026, used for every conversion in this post. Fetched 10 September 2026.