The short version
- Rent it: a managed assistant costs tens of dollars a month and is working this week. Where most 10–50 person companies should start. It is also my own product, so weigh the recommendation accordingly.
- Commission it: a focused custom build is 4K–17K (general market range) over 3–8 weeks, once the assistant has to work your way, in your systems.
- Build a platform: 25K–128K, 8–20 weeks, when several agents, several teams and your own data layer are the actual requirement.
- Running costs are the lie: a 12K assistant costs roughly 1,800–3,600 a year to keep healthy, not the six figures the enterprise pricing pages imply.
If you search for what an AI assistant costs, page one hands you a single number band and lets you assume it applies to you. It usually does not. I build these systems for small and mid-sized companies, and the honest answer is that a 12-person firm and a 1,200-person firm are buying two different products that share a name.
So here are all three prices, in the order I’d actually give them on a discovery call. Disclosure before you read another line: I have a commercial interest in all three. The rental in option one is my own product, and it is the one I earn least from. That is exactly why it goes first.
What the top-ranking pricing pages actually quote
These are the numbers a small-business owner meets first. I checked them on 26 July 2026:
| Source | Build cost quoted | Ongoing cost quoted |
|---|---|---|
| Riseup Labs | $10K–450K+ | 15–30% of build per year; year-one total “~$160K (lean MVP)” |
| ProductCrafters | $20K–35K+ simplest agent, $100K–200K+ domain-specific | 15–25% of build per year ($10K–50K+) |
| Softermii | “~$10,000 to $250,000”; enterprise multi-agent “$100K–$500K+” | Maintenance and ops $400–$6,000/month |
None of that is dishonest. It is accurate pricing for enterprise engagements: dedicated model work, multi-department rollout, compliance review, a delivery team on the account. The problem is the audience. A ten-person company reads “lean MVP, $160K in year one” and quietly decides the whole category is out of reach.
Notice too that every number on that list is a build price. Not one of those pages offers a tier where you do not commission anything. That is the option most callers should take, so it goes first.
The three real options, priced
One note on currency before the numbers. The vendor figures above are US dollars, as those vendors publish them. My own build bands are written currency-neutral, the way they appear everywhere else on this site, because I quote in euros, pounds or dollars depending on where the client sits. Read them as the size of the number, not an exchange rate.
| Option | Cost | Live in | Right when |
|---|---|---|---|
| 1. Rent — a managed assistant, run for you (Nora, my own product) | From $59/month, 2-week free trial | This week | You want an assistant on WhatsApp, email and phone that answers, chases and summarises, using the tools it already connects to. |
| 2. Commission — a focused custom build | 4K–17K (up to 42K across many systems) | 3–8 weeks | The assistant has to write into your systems, follow your rules, and behave in a way no product supports. |
| 3. Build — your own platform | 25K–128K | 8–20 weeks | Several agents, several teams, your own data layer, admin controls and an audit trail you own. |
Row one is a subscription, and its terms are public: a two-week free trial, then from $59 a month for unlimited assistants across phone, email, WhatsApp and voice notes (checked 26 July 2026). It is my product, so read the recommendation with that in mind. Rows two and three are my project work. You can price row two in about thirty seconds with the custom AI cost estimator, and design row three piece by piece — channels, agents, integrations — in the interactive system builder. Both quote in euros, so expect the figures there to sit slightly differently against the neutral bands above.
One thing to keep straight if you click through to that estimator. It also quotes a monthly managed plan, from €249, and that is not row one. It is ongoing hosting, updates and support wrapped around a custom build you have already commissioned. Row one is the opposite: a subscription with nothing built, nothing to commission, and no project to support.
The gap between row one and row two is the whole point of this article. Renting the entry plan costs about $710 across a full year. The cheapest custom build starts at roughly 4K before a line of running cost. Neither number is wrong, and none of the pages I checked shows both on the same screen.
For a sense of where real projects land inside those bands, the anonymised ranges published on my cost page run from a WhatsApp support bot for a retail chain (two systems, four weeks) up to a multi-department operations platform (twelve-plus systems, on-premise, sixteen weeks). Same category, two orders of scope apart.
Estimate your custom AI project in 30 seconds
Three questions, an instant cost range and timeline based on real shipped projects. After 30 minutes on a discovery call you have a written fixed-price quote.
Or build your own AI system piece by piece and send the design in for a written quote →
Why I tell most callers to start at option one
Option one earns me tens of dollars a month. Option two earns me thousands. I still send most callers to option one, and you should weigh that against the fact that I own it. When someone says “we want an AI assistant”, what they describe on the call is usually: answer questions from our documents, chase leads that went quiet, summarise the inbox, take notes, remind me about things, and let me do it from my phone. That is a configured product. Paying to have it built from scratch is paying to rebuild something you can rent for the price of a phone contract.
Two more reasons the rental comes first.
It writes your spec for free. Sixty days of real usage tells you which handful of things the rented assistant cannot do for your business. That list is the scope of the custom build, written in the only language that matters: daily annoyance. Custom projects that start this way get scoped faster and land better, because nobody is guessing. That is the same hybrid pattern I argue for in custom AI vs off-the-shelf software.
The alternative failure is expensive. The worst outcome in this category is not overspending. It is spending 12K to automate the workflow you described in a meeting, then discovering it is not the workflow your team actually runs. Rent first and that discovery costs a subscription.
What running one actually costs
The maintenance percentages quoted above (15–30% of build cost per year) are fine as ratios. The damage is done by applying them to enterprise build numbers and showing the result to a small business. Run the same ratio on a small-business build and it looks like this:
| Line item | Rented assistant | 12K custom build |
|---|---|---|
| Subscription | $59+/month | — |
| Model API usage | Included | 50–500/month |
| Hosting | Included | 10–300/month |
| Maintenance and changes | Included | 1,800–3,600/year |
| Year one, all in | ~$710 at the entry price | ~14,500–25,000 |
The rental column is the published entry price; higher tiers cost more and I have not put a number on them here. The model API band is the one I quote on my own cost page. The hosting and maintenance figures are what I see in practice on the systems I run rather than a published rate card: a support assistant handling a few hundred conversations a month sits near the bottom of the API band, a document-processing automation chewing through long PDFs sits near the top. Nothing about a small-business assistant requires a maintenance team. It requires one person who is accountable when it breaks, and a small budget for the changes real usage exposes in month two.
One caveat I will not skip: “maintenance and changes” is not optional. Models change, APIs change, your business changes. A build with nobody accountable for it degrades quietly and gets blamed on AI. Budget the line or buy a managed plan that covers it.
The four questions that route you
This is the routing test I run on discovery calls, in this order. Stop at the first honest “no”.
1. Can you name the systems it must write into?
Not read from — write into. Reading is cheap. Writing is where cost lives: permissions, duplicate protection, an audit trail, and a plan for when the assistant gets it wrong at 2am. If you can name three systems and what changes in each, you have a custom build. If you cannot, you have a rental.
2. Does it need judgment nobody has written down?
“Prioritise this client because of what happened in March”, “never quote this material below that margin”, “escalate anything from a hospital”. Encoding your best operator’s judgment is exactly what custom is for, and no product ships with your rules. If the judgment is generic, a configured assistant already handles it.
3. Is the workflow the reason customers choose you?
If the answer is yes, the workflow deserves software shaped around it. If the workflow is admin — scheduling, chasing, filing, summarising — it is the same problem as at a million other companies, and someone has already built it better than a bespoke project would.
4. Have you run a rented assistant for 60 days?
If not, do that before commissioning anything. It costs less than one day of development, and it converts a wish list into a specification. I have talked callers out of a build at this question. The ones who come back come back with a smaller, better-defined project, and those are the ones that go well.
What actually moves the price of a custom assistant
When a build is the right answer, these are the drivers that decide whether you are at 5K or 40K. None of them is “the AI”.
- Write access, not read access. An assistant that answers from your documents is a short project. One that creates the invoice, moves the deal stage and emails the client is a longer one, because every write needs guardrails, logging and a way to undo it.
- Channels. A web widget is the cheap one. WhatsApp, phone, email and calendar each bring their own authentication, delivery rules and threading behaviour. Three channels is not three times one channel, but it is not one either.
- Memory and handoff. Remembering a customer across conversations, and handing a live thread to a human with the context intact, is where support assistants either earn their keep or annoy people.
- The condition of your data. The single most underestimated line. Clean, exportable records in a modern system are easy. A decade of PDFs, a spreadsheet with merged cells and a 2014 database with no API is a data project wearing an AI project’s clothes.
- Compliance and hosting. If patient or financial data cannot leave your infrastructure, hosting choices and audit requirements add real engineering, and they are the reason some businesses cannot use option one at all.
A serious quote prices these individually. If a vendor hands you one number with no breakdown, you are comparing guesses. There is a fuller breakdown of the drivers in the custom AI development cost guide, and agent-specific pricing in the AI agent development cost breakdown.
A worked example
A finance team at a power-generation company was reconciling bank statements against invoices by hand, every day. No product covered it, because the mapping between their ERP and their bank feeds was specific to them, and the assistant had to write results back into the ERP. That is questions one and two answered “yes” before anyone talked about price.
It shipped in three weeks and replaced hours of manual entry a day. It was a focused custom build, not a platform, and not a subscription — the shape of project that sits in the middle row of the table above. The small-business page has more of these, and the point is always the same: one painful, repetitive workflow, automated properly, paying for itself in months.
Compare that with the caller who wants an assistant to answer website questions and chase quiet leads. Same phrase, different product. That one should be renting by Friday.
Frequently asked questions
How much does it cost to develop an AI assistant?
In 2026 there are three honest prices, not one. Renting a managed assistant starts at around $59 a month and is live the same week. Commissioning a focused custom build (one workflow, a few integrations) runs 4,000-17,000 as a general market range and takes 3-8 weeks. A full platform with multiple users, your own data layer and several connected agents runs 25,000-128,000 over 8-20 weeks. Build bands are written currency-neutral because I quote in euros, pounds or dollars depending on where the client sits. The $20,000-200,000 quotes you see on most pricing pages describe the third shape only.
How much does custom AI assistant development cost for a small business?
For a 10-50 person company, the realistic band is 4,000-17,000 for an assistant that answers questions and handles one workflow, and 8,000-42,000 once it has to read and write across several systems. Below that, a managed subscription is usually the better buy. Above that, you are buying a platform, not an assistant, and the scope should be written down before anyone quotes it.
What does it cost to run an AI assistant each month?
For a small business: model API usage typically 50-500 a month depending on volume, hosting from single figures a month on a small server to a few hundred on managed infrastructure, plus a change budget for the tweaks real usage exposes. A 12,000 build usually costs 1,800-3,600 a year to keep healthy, which is the standard 15-30% maintenance ratio applied to a small-business build instead of an enterprise one. The six-figure annual maintenance numbers on vendor pricing pages are that same ratio applied to enterprise build costs.
Is it cheaper to build or rent an AI assistant?
Renting is cheaper until the assistant has to work your way. The entry plan on the managed assistant I run is $59 a month, about $710 across a full year, with no build cost and no delivery risk. A custom build wins when the assistant must write into systems no product supports, follow judgment rules only your team knows, or keep data somewhere specific. Most companies should rent for 60-90 days first, because the rental writes the spec for the build. Disclosure: I sell both the rental and the build, and the rental is the one I earn least from.
Why do agencies quote $20,000-200,000 for an AI assistant?
Because they are pricing enterprise engagements: dedicated model work, multi-department rollout, compliance review, a delivery team. Those are real costs for that shape of project. The problem is that a 10-person company reading the same page assumes the number applies to them. It does not. Ask any vendor quoting six figures which line items disappear if you cut the scope to one workflow and one integration.