The distinction nobody bothers to explain
Almost every AI enquiry we get opens with the same question: "how much does an AI agent cost?" The problem is that when I ask what the agent should do, half the answers describe a chatbot. Those are two different products with two very different price tags.
A chatbot takes a question and returns an answer. Even a very good one, grounded in your company documents and sounding perfectly human, ends at text. If the customer asked for a quote, someone still has to open the CRM and build it.
An AI agent takes a task, breaks it into steps, and performs actions in real systems. It reads the request, recognises it as a quote enquiry, pulls the customer record, checks whether they're an existing account, generates a quote against your price list, opens a CRM ticket, and schedules a follow-up in three days.
The practical difference: a chatbot talks, an agent does. If all you need is something that answers common questions, you don't need an agent and you shouldn't pay for one.
Where it genuinely pays off
1. Inbound request handling
A business taking 40–150 enquiries a day across WhatsApp, email and forms. Some are identical, some need a system lookup, and some simply get lost because nobody checked the inbox on Friday. An agent that reads, classifies, answers what it can and escalates what it can't is the clearest use case there is.
The metric that matters here isn't time saved per enquiry — it's how many enquiries fall through the cracks. A business losing 10% of its inbound loses far more than staff hours.
2. Document processing
Supplier invoices, purchase orders, onboarding forms. Someone types these into a system. It's slow, dull and error-prone — exactly what machines do better than people. The saving is easy to quantify: documents per month times average handling time.
3. Internal knowledge search
An organisation with hundreds of procedure documents, contracts and customer history that nobody can navigate. An agent that answers "what are the cancellation terms in customer X's contract?" with a citation. The value isn't one person's time — it's that institutional knowledge stops depending on one long-serving employee.
4. Cross-system workflows
Any process where someone copies data from system A to system B. If the process is well-defined and repetitive, you often don't need AI at all — plain rules-based automation is cheaper and more reliable.
Where it isn't worth it
- A process that runs five times a month. Even an annoying one. Build and maintenance costs won't come back.
- A process with no clear rules. If two experienced staff do it differently, you can't automate it — you first have to decide what correct looks like.
- A process where errors are expensive and nobody reviews output. Models make mistakes. One bad action can eat the entire saving.
- When your data is a mess. An agent reading a dirty database returns dirty answers, faster.
What it actually costs
Two separate costs that people routinely conflate: build cost, and monthly run cost.
Build
At FLOWDEV a single focused automation — an agent that takes inbound messages, classifies them and opens a CRM ticket — starts at $1,200 and takes 2–4 weeks. A full system with several agents, an admin dashboard and deep integrations starts at $18,000.
That range isn't arbitrary. It comes down to one question: what does the agent have to connect to? An agent working against a modern documented API is a two-week project. The identical agent against a legacy system with no API, requiring direct database access, is a two-month project.
Monthly run cost
This is the cost vendors tend not to mention. Every request the agent handles consumes model tokens, and that costs money. For most businesses we work with it lands somewhere between tens and a few hundred dollars a month, but it depends entirely on volume and how much context you send with each call.
Ask any vendor for a written estimate at your volume before signing. If they can't produce one, they haven't built enough of these.
The question to ask first
Before we build anything, we ask the client one question: how long does this process take today, in numbers?
Not "a lot". Not "all day". A number. How many requests a day, how many minutes each, how many people involved. Without it there's no way to know whether automating pays, and no way to prove afterwards that it worked.
Roughly a third of these conversations end with me telling the client not to do it. Either the volume is too low, or the process isn't defined enough, or an off-the-shelf tool solves it for a fraction of the price.
The architectural point that matters most: human handoff
Language models make mistakes. Not occasionally — at some rate, consistently, and that isn't going away. The engineering question isn't "how do we prevent errors" but "what happens when the model is wrong".
Every workflow we build includes a confidence threshold: when the model isn't sure, it doesn't guess — it escalates. For sensitive processes (money, legal commitments, medical information) the default inverts: a human approves before anything is sent.
If a vendor demos an AI agent and doesn't discuss confidence thresholds, logs and human handoff, they're showing you a demo, not a production system.
Bottom line
AI agents are a good tool for high-volume processes with clear rules and room for controlled error. They're a poor tool for rare, ambiguous or critical-without-oversight processes. Which one you have is mostly determined before a line of code — at the point where you measure the existing process.
Further reading: custom system vs off-the-shelf and the SaaS development guide. Want to check whether your process fits? See our AI agents and automation service, or just tell us what repeats every day and we'll give you an honest read.
Written by Evyatar Adler, CTO at FlowDev.
