Last quarter a prospect showed us three quotes for what their one-page brief called an AI assistant for customer support. The lowest was $4,000. The highest was $70,000. All three proposals used the same words: AI, assistant, intelligent, automation. Their question for us was reasonable: is the expensive vendor gouging, or is the cheap one lying?
Usually neither. The vendors are quoting different systems that happen to share a vocabulary. There is a capability ladder in this market with four distinct rungs, and until you know which rung a proposal is standing on, the price tells you nothing. Here is the ladder, what each rung can and cannot do, and how to work out which one your problem actually needs.
The short answer: a chatbot answers; an AI agent acts. A chatbot — even one grounded in your own documents — only reads: it can tell a customer what the refund policy says. An agent reads and writes: it looks up the order, checks eligibility, and issues the refund. That one difference is why agent projects at AGI Software Solutions carry permissions engineering, evaluation suites and approval guardrails a chatbot never needs — and why honest quotes for the two differ by 10×.
Why quotes for the same brief differ by 10x
The words chatbot, AI assistant, and AI agent have no agreed technical meaning in a sales conversation. A vendor reselling a $30/month FAQ widget calls it an AI agent. A team building a system that reads your order database and issues refunds also calls it an AI agent. Both are being roughly honest by their own definitions, and the buyer is left comparing a bicycle quote against a truck quote because both said vehicle.
The real difference between the rungs is not how smart the model is. It is two capabilities: whether the system can read your data, and whether it can act on your systems. Each capability you add multiplies the engineering effort, because each one adds a whole category of failure that has to be designed against before launch. That is where the 10x spread comes from, and it is why comparing quotes without naming the rung is a waste of everyone’s time.
Rung 1: the scripted bot (no AI at all)
A rung-one bot is a decision tree with a chat skin. Buttons, menus, keyword matching, canned answers. There is no language model anywhere in it, which sounds like an insult in 2026 and is not: a decision tree cannot hallucinate, cannot be prompt-injected, and costs almost nothing to run.
It is the right buy when a narrow set of questions makes up most of your volume. Where is my order. What are your hours. How do I reset my password. If your support inbox is 60% the same eight questions, a scripted bot with a clean escalation path clears them for a cost that rounds to zero, and it does so with perfect consistency.
Typical shape: an off-the-shelf widget, a few days of configuration, a low one-time fee or a small subscription. The failure mode is just as predictable: the ninth question. Anything outside the tree produces a loop of unhelpful menu options, which is exactly why these bots earned the reputation they have. Buy this rung for the eight questions, never for open-ended conversation.
Rung 2: an LLM chatbot that knows your business
Rung two adds a language model plus retrieval over your own content, the pattern the industry calls RAG. The bot answers in natural language, and instead of following a script it draws on your documents: help articles, policies, product specs, past tickets. Ask it something phrased in a way nobody anticipated and it can still find the relevant policy and answer sensibly.
What it cannot do is touch anything. It can tell a customer what your refund policy says; it cannot issue the refund. It reads, it never writes.
This is where most businesses should start, because most of what gets described to us as a support problem is really a findability problem: the answer exists, buried on page 14 of a PDF nobody opens. The engineering at this rung is well understood by now. The work that remains is unglamorous: cleaning the source documents, deciding what the bot must refuse to answer, and testing it against a few hundred real customer questions before launch. In our experience the rung-two projects that fail, fail on document quality, not on the AI.
Cost profile: the projects we scope at this rung usually land in the low five figures and take three to six weeks, plus a modest monthly bill for model usage. Predictable, low-risk, and an honest vendor will say so.
Where "AI virtual assistant" sits on this ladder
Worth pinning down, because it is the vaguest term in the category and vendors use it at every rung. An AI virtual assistant almost always means rung two: it converses in natural language, it knows your business through retrieval, and it hands off when it reaches its limit. What separates the term from "chatbot" is usually scope and persona rather than capability, so an assistant is often a chatbot pointed at a wider set of documents with a friendlier name.
Where the label gets expensive is when it is used for rung three. If a proposal describes an AI virtual assistant that will "handle" bookings, refunds or account changes, that is not rung two with a nicer personality, it is an agent with write access, and it needs the permissions, evaluation suite and approval design described below. Ask one question to place any proposal on this ladder: can it change something in one of our systems? Yes puts it at rung three regardless of what the deck calls it.
Rung 3: an agent that can actually do things
Rung three is where the word agent starts to mean something. The system does not just answer questions about an order; it looks the order up in your live systems, checks eligibility against policy, and triggers the refund, the rebooking, or the address change. It reads and acts. We have written before about why we think of these as automation workflows with judgment rather than chat interfaces, because the chat window is the least important part.
The price jump between rung two and rung three surprises buyers, and it should not. Everything that made rung two safe disappears the moment the system can write:
- Permissions. The agent needs scoped credentials for your order system, CRM, or calendar. Deciding exactly what it may touch, and proving it cannot touch anything else, is a real security project, not a checkbox.
- Evals. A wrong answer at rung two mildly annoys a customer. A wrong action at rung three refunds the wrong order. Before launch you need an evaluation suite that runs the agent through hundreds of scenarios, measures how often it does the right thing, and re-runs on every change.
- Guardrails and approvals. Which actions execute automatically, which queue for human sign-off, what the spend and volume limits are, and what the agent does when it is unsure.
That work is most of the invoice. When a rung-three quote arrives at rung-two prices, the evals and guardrails are what got deleted, and you will pay for them after launch instead of before, at a worse exchange rate. Honest rung-three projects, in our experience, run from the mid five figures upward depending on how many systems the agent touches and what a mistake costs.
Rung 4: multi-agent systems (you probably do not need this)
Rung four is several specialized agents coordinating on a genuinely multi-step operations problem: an order-to-cash flow where one agent validates, another reconciles, a third handles exceptions; a claims process; a procurement loop. The orchestration, state management, and failure handling are a serious engineering effort in their own right, and we covered what that actually involves in our piece on multi-agent orchestration.
We list this rung mostly to arm you against upsell. The overwhelming majority of companies never need it. If a vendor proposes a multi-agent architecture for your support queue, ask them to name the second agent’s job in one sentence. If the answer is vague, you are looking at rung-three work wearing a more expensive costume.
Match the rung to the job, not to the demo
Nearly every bad outcome we get called in to fix is a mismatch, and the mismatches run in both directions.
Buying up. A rung-three agent deployed against a rung-one problem. We have reviewed proposals where the entire use case was collecting a name, an email, and a preferred appointment slot. A web form would have done it, in a week, for a hundredth of the price. The demo was impressive; the job never needed judgment.
The expensive mistake in this market is not picking the wrong vendor. It is buying the wrong rung: an agent for a job a web form could do, or a scripted bot for a job that needed hands.
Buying down. The opposite failure is more common and does more brand damage: a rung-one or rung-two bot deployed where customers expect action. The bot cheerfully explains the refund policy to someone who wants a refund issued. It understands the request perfectly and can do nothing about it, which customers experience as a company that built a machine specifically to not help them. A bot that raises the expectation of help and then cannot act scores worse with customers than no bot at all.
Four questions that reveal your rung
You can locate your problem on the ladder before you take a single vendor call. Write down honest answers to these:
- Does it need your data? If generic, pre-scripted answers genuinely cover the traffic, rung one is enough. The moment correct answers depend on your policies, catalog, or account records, you are at rung two minimum.
- Does it need to take actions? If the job is only done when a record changes somewhere, that is rung three, whatever the vendor calls it. If the job is done once the person has the right information, stop at rung two.
- What does a wrong answer cost? A mistaken shipping estimate is an apology. A mistaken dosage, quote, or refund is an incident. The cost of being wrong sets your budget for evals and guardrails more than the rung itself does.
- What is the volume? Twenty conversations a day does not justify rung-three engineering; a person with good templates wins on cost and quality. Two thousand a day changes the math completely.
Then make every vendor quote against those written answers. The 10x spread narrows fast once everyone is forced to price the same rung.
Where to start
- Pick one workflow, not a mission. Refund requests, order status, appointment booking. Something with a beginning, an end, and a number you can measure.
- Buy the lowest rung that fully solves it. Not the lowest rung you can afford to be disappointed by, and not the highest rung in the demo reel.
- Upgrade on evidence. Run rung two for a month and read the transcripts. The questions it could answer but not act on are the best requirements document for a rung-three agent you will ever get, and they are free.
- Budget with open eyes. We broke down the real numbers, including the ongoing costs nobody quotes, in how much AI development really costs.
- Get a second opinion on the rung. If you have a workflow in mind and want an honest read on whether it needs an agent or a form, that scoping conversation is exactly what our AI agent development engagements start with. Sometimes the answer is rung one, and we will say so.
Frequently asked questions
What is the difference between a chatbot and an AI agent?
A chatbot answers; an agent acts. Even an LLM chatbot with retrieval only reads — it can tell a customer what the refund policy says. An agent reads and writes: it looks the order up, checks eligibility, and issues the refund. The write capability is what adds permissions, evals and guardrails to the bill.
How much does each rung cost?
Scripted bot: a low one-time fee or small subscription. LLM chatbot with RAG: low five figures, three to six weeks. Tool-using agent: mid five figures upward, depending on how many systems it touches and what a mistake costs. A rung-three quote at rung-two prices means the evals and guardrails got deleted.
Does my business need an agent or a chatbot?
Four questions locate the rung: does it need your data; does it need to take actions; what does a wrong answer cost; and what is the volume? If the job is done once the person has the right information, stop at a chatbot. If the job is only done when a record changes somewhere, that is an agent.
Do I need a multi-agent system?
Probably not. Rung four is for genuinely multi-step operations problems — order-to-cash, claims, procurement. If a vendor proposes it for your support queue, ask them to name the second agent's job in one sentence.