Automation · 8 min read
What an AI chatbot actually costs to run
The build is the part everyone quotes and nobody is surprised by. The running cost has three parts, and the biggest one is not the model.
Updated 27 August 2026 · GoDesign build team
The short version
- AI chatbots are the largest topic in our job data with no pricing tool behind it: 78 of 28,614 briefs, and rising.
- Running cost has three parts — model usage, platform, and human escalation — and quotes routinely show one.
- At 2,000 conversations a month a mid-tier model costs around AED 180. The humans handling what it cannot finish cost around AED 2,112.
- Below a few hundred conversations a month a bot costs more to run than answering by hand. Build it for speed and out-of-hours cover, not for the saving.
Chatbot quotes tend to have one number for the build and one for 'AI usage'. The build is fine — it is ordinary project work and it behaves like ordinary project work. The running cost is where these projects go wrong, because it has three separate components that get collapsed into one line, and the one everybody focuses on is the smallest.
The three running costs
Model usage is billed per million tokens in and out. Tokens in are your prompt plus whatever context the bot retrieves before answering; tokens out are the reply, and they are priced several times higher on every provider. A support conversation is commonly 3,000 to 15,000 tokens in and 500 to 1,500 out, depending on how much context you feed it.
Platform is what you pay if you buy a bot product rather than building against the API. Visible, predictable, and the easiest of the three to compare.
Human escalation is everything the bot cannot finish, and it never reaches zero. This is the line that decides the project, and it is the one almost no quote contains.
| Line | At 2,000 conversations/month | Share |
|---|---|---|
| Model usage, mid-tier | AED 180 | 6% |
| Care and prompt maintenance | AED 660 | 22% |
| Human escalation, ~32% of contacts | AED 2,112 | 72% |
| Monthly total | AED 2,952 | 100% |
Why the model tier matters less than you think
There is roughly an order of magnitude between a small model and a frontier one. That sounds decisive until you notice the model is 6% of the total: taking the frontier tier down to a small one moves the monthly bill by a few percent. For answering questions out of your own documents, the small model is frequently indistinguishable anyway, and paying frontier prices to read a FAQ is the most common way an AI budget disappears without anyone noticing.
Where the tier does matter is when the bot acts rather than answers — writing to a CRM, changing an order, booking a slot. Getting that wrong has a cost that is not measured in tokens.
When a bot does not pay for itself
Below roughly a few hundred conversations a month, a chatbot costs more to run than simply answering by hand. That is not a reason to abandon the idea — it is a reason to be honest about why you are doing it. At low volume you are buying an instant first reply and cover outside working hours, both of which are worth real money in a market where a two-hour reply reads as being ignored. You are not buying a saving, and a vendor telling you otherwise is selling.
Our estimator shows the same volume handled entirely by people alongside the bot's total, and states plainly when the bot is the more expensive option. It also gives a payback period on the build: beyond about eighteen months, the volume is not there yet and the sensible move is a narrower bot on a smaller model.
What to do before you commission anything
- Write down the ten questions customers actually ask most. A bot that answers those ten well deflects most of your volume, and the scope you need is far smaller than the one you were about to buy.
- Decide what it must never do. Guardrails and an evaluation set are the first thing cut and the hardest thing to regret, because a bot that confidently says something wrong costs more than the whole project saved.
- Get your real contact volume from wherever it already exists. The economics change completely between 200 and 2,000 conversations a month.
- Price escalation explicitly. If the proposal in front of you has no escalation line, it is describing a system that does not exist yet.
Questions people ask
How much does an AI chatbot cost per month?
For a small business at around 2,000 conversations a month, roughly AED 2,900 all in — of which the model is about AED 180 and the humans handling escalations about AED 2,100. The spread across the market is enormous because most quotes count only some of those lines.
What deflection rate is realistic?
Fifty to sixty percent for a bot answering from your own content, higher for one that can act inside your systems, and lower than any vendor demo suggests in the first month. Treat published figures as a target you reach after tuning rather than a launch number.
Should I use a chatbot platform or build on the API?
A platform if you want it live in weeks and are happy with what it can do; the API if the bot needs to act inside systems you already run. The platform fee is visible and the API is cheaper per conversation, so the crossover depends almost entirely on volume.
Why do model prices keep changing?
Providers cut them frequently and without notice, and the tiers differ by an order of magnitude. Any tool that hardcodes a model name is wrong within a quarter, which is why ours takes the price per million tokens as an editable input with tier anchors.
What is not included in a typical chatbot quote?
Per-resolution platform charges where they exist, the content work to make a knowledge base worth retrieving from, and your team's time correcting the bot in its first month. The last one is real and always underestimated.
Built by GoDesign FZE, who build WhatsApp and CRM automation for UAE businesses.
Get in touch
Want someone to just do this part?
The tool is free and stays free. If you would rather hand the work over, tell us what you are dealing with and you get a straight answer on scope, cost and timeline.
- You own everythingRepository, hosting and domain credentials sit in your name from day one, not handed over at the end.
- A written timelineMilestone dates are agreed in writing before work starts, so you always know what ships next.
- One business dayEvery enquiry gets a reply from the person who would scope the work, not a sales sequence.
Media City, Dubai, UAE · DHA Phase 2, Islamabad, Pakistan
Would rather type than fill in a form?
Message on WhatsApp