How much does AI customer service really cost?
Custom builds, per-resolution help-desk AI and flat-price assistants compared: what each really costs and how long each takes to implement, as of November 2026.
ComparingBy Ron Pinkas8 min read8 sections
You asked for two quotes. One says “$0.99 per resolution.” The other says “$40,000 and twelve weeks.” Both are for an AI that answers your customers, and you have no idea how to compare them.
That is the point of this guide. We build one of these assistants ourselves, so we have a side; we will show our numbers next to everyone else’s, with the date we read them, and you can check every one. There are three ways to get an AI assistant today, and each is priced in a completely different way.
Path 1: Have one built for you
A development firm builds an assistant around your systems. The firms that publish their prices quote $40,000 to $120,000 to build a customer-support agent and $2,500 to $5,000 a month to run it, with 8 to 20 weeks for a typical agent, simple to complex (SoftTeco, 2026). Others quote $25,000 to $50,000 for a proof of concept and over $100,000 for production (Neoteric). These are the firms’ own ranges, not independent research; no analyst we found publishes one.
Who it suits: a large company with unusual systems, its own technical team, and a budget measured in quarters.
Path 2: Add AI to a help-desk platform, and pay per answer
Most of the big customer-service platforms now sell an AI agent billed per resolved conversation. As of November 8, 2026:
| Platform | The AI | What you need first |
|---|---|---|
| Fin (formerly Intercom, now owned by Salesforce) | $0.99 per “outcome”: a resolved conversation, or a hand-off to your team that you have set up as a workflow | Works with your own help desk (a monthly minimum applies, for example 50 outcomes) or Fin’s help-desk seats from $29 a month billed annually |
| Zendesk | 5 to 10 resolutions per agent included, then $1.50 each prepaid or $2.00 pay-as-you-go | A Zendesk plan, from $19 per agent a month billed annually |
| HubSpot (Breeze) | $0.50 per resolution; $0.50 a minute on the phone (beta) | Service Hub Professional, from $90 per seat a month billed annually, plus a $1,500 onboarding fee |
| Freshdesk (Freddy) | 500 sessions free, once per account, then $49 per 100 | Freshdesk seats, from $19 per agent a month billed annually |
The vendors say the AI itself goes live “in minutes” if you already run their platform. Getting there is another matter: many of these platforms start with “Book a demo” or “Talk to sales,” onboarding runs through a sales rep, and mid-size companies typically hire an implementation partner. One such agency quotes $3,750 to $17,500 and 2 to 6 weeks to set up a single help-desk platform’s AI. The other catch is the meter. At 500 resolved conversations a month, $0.99 each is $495, and the bill rises in exactly the months you are busiest. Read the overage terms too: on Zendesk, resolutions beyond your allowance cost more than prepaid ones, unless you choose to pause the AI when you reach the limit.
You should not need an accountant to read a price list
Look at the units in that table: outcomes, resolutions, credits, sessions. Each vendor counts something different, so their prices cannot be compared side by side. Gorgias, a help desk popular with online stores, shows how far this goes. As of November 8, 2026, its pricing page lists:
- four plans, each with its own monthly ticket allowance (50 to 5,000) and its own price per extra ticket ($0.36 or $0.40);
- AI billed separately, per “automated interaction,” at $0.90 or $1.00 depending on your contract, each plan including a different number of them (30 to 530), and $1.50 for each one beyond that;
- a rule that if one of your agents replies within 72 hours of an AI answer, it is billed as a ticket instead;
- phone and text as add-ons whose price “scales with” your volume, with no figures given.
Before you sign with anyone, ask them to price your last three months of real conversations, in dollars.
Who it suits: a company already paying for one of these help desks, with a support team that works in it every day.
The part that takes months: your back-end systems
Answering questions is the easy half. The valuable half is doing things: taking a payment, looking up an order, changing an account. On most platforms, connecting the assistant to your billing, order or account systems means months of work and an expensive contract, if it is offered at all. And there is a second question most buyers never ask: who decides what happens to the money? For a payment, “usually right” is not good enough: the steps that move money should run as fixed code, with no AI judgment in them. The largest platforms have only recently begun to offer this. Salesforce made its version, which runs such steps “as code with no LLM involvement,” generally available in February 2026; Fin’s Procedures add “deterministic controls” such as branching and code to an AI-led conversation. Both are built and configured inside their own platforms.
Path 3: A ready-made assistant at a flat price
This is what we sell, so judge it hardest. An instantAIguru Guru is built from your own website, by you: no demo, no sales call, no implementation partner. Your website, email and phone are answered within minutes of signing up, plus the time it takes to read your site. Text messaging and WhatsApp, which not every platform includes, take a few days for carrier and Meta approval. It is priced per channel by the size of your site, with no charge per conversation, per resolution or per seat: from $95 a month per channel on Starter for up to 100 pages and 1,000 chats a month across all channels, up to $695 on Scale for up to 1,000 pages and 10,000 chats, 25 percent less paid annually. Web, email and phone together on Starter are $285 a month. For the phone line you also pay Twilio directly: about $1.15 a month for a US number, plus calls. Every plan starts with a 7-day free trial, no card, and a 90-day money-back guarantee.
For payments, account changes and other transactions, our Guru takes the AI out of the decision entirely. The AI understands what the customer wants; the transaction itself runs as a fixed, scripted business flow that does exactly the same thing every time. This is not new for us: it has been processing hundreds of financial transactions a day at our largest customer, Curacao, for over a year, and more than 200,000 of these flows, including over 100,000 phone calls, have run exactly as scripted, without a single error.
Connecting your own systems is self-serve, too, in our Flow Designer. We connected dozens of Curacao’s services in days; with the AI copilot now built into the designer, we expect most integrations to take hours.
Who it suits: any business that wants its website, email and phone answered today, without a sales process, a help-desk project or a meter; and any business that needs its assistant to handle real transactions safely.
What about the phone alone?
Phone-AI platforms sell by the minute: from $0.07 to $0.31 a minute (Retell AI), $0.12 to $0.14 plus telephony (Bland), or a $0.05 platform fee plus voice and carrier costs (Vapi), as of November 8, 2026. Those are toolkits: you or a developer design the call flows. A human answering service costs far more: Ruby’s plans work out to about $3.45 to $5.00 a minute.
The cost you are paying now
Doing nothing is not free. The typical customer service representative in the U.S. earns $21.53 an hour (Bureau of Labor Statistics, May 2025), and benefits make up about 30 percent of a private employer’s total compensation cost (BLS, June 2026), so a person answering your phone costs roughly $30 an hour, for the hours they work. The questions that arrive at 9:40 at night still go unanswered.
The costs nobody puts in the quote
- The meter. Gartner predicts that by 2030 the cost per resolution for generative AI will exceed $3, “higher than many B2C offshore human agents” (January 2026). Per-answer pricing that looks low today may not stay low.
- The prerequisites. Seats on a help desk, onboarding fees and monthly minimums often cost more than the AI itself.
- The people. 87 percent of customers say access to a human is essential when a company uses AI, and only 27 percent would try a chatbot again after a bad experience (Gartner survey of 3,566 customers, released August and September 2026). Whatever you buy must hand off to a person well.
- Keeping it current. An assistant is only as good as what it knows. Ask every vendor how its answers stay up to date when your website, prices or hours change, and who does that work.
Six questions to ask any vendor, including us
- What will I pay in my busiest month, not my average one?
- What do I have to buy first, and is there an onboarding fee?
- How long until it answers my real customers, and who does the work?
- What happens when it does not know the answer?
- How does it hand a customer to a person, and on which channels?
- Can it take a payment or change an account, and does the AI decide what happens, or a fixed business process?
Prefer a hand?
You do not have to do any of this alone. If you would rather have someone walk you through it, or simply do it for you, contact our team. We are glad to help with any part of your onboarding, from preparing your content to connecting your phone line, and there is no such thing as a question too small.
My view: if you need a CPA to understand a price list, that is the first thing it is telling you. The most cost-effective assistant is the one your customers actually use and that you never have to think about in December. Ask for the busiest-month number. If a vendor cannot give you one, you have your answer.
See it on your own content
Paste your website URL and your AI customer service agent goes live in minutes. Free to start, no credit card.
Prefer to look around first? See pricing or watch a demo.