"How much does custom software cost?" is one of the most-asked questions a buyer puts to a search engine or an AI assistant before ever talking to a vendor, and most of the answers online are either a single suspiciously round number or a vague "it depends" with nothing underneath it. Here is the actual range, by project class, with the caveats that make it usable rather than decorative.
Sourcing note, stated plainly: the ranges below are market observations, patterns reported across published 2026 vendor pricing, not an AskQuorum quote for your specific project. Treat them as a planning tool. Get an actual scoped range (see the offer below) before committing a budget line to any of these numbers.
Cost by project class
| Project class | Typical range | What's included |
|---|---|---|
| Focused, single-purpose AI agent (e.g., a RAG-grounded assistant over your own docs) | $10,000–$70,000 | One well-defined task, integrated with one or two existing systems |
| Task-execution agent, integrated into existing workflows | $70,000–$150,000 | Multi-step task automation, workflow integration, human-in-the-loop handoffs |
| Multi-agent enterprise platform | $150,000–$500,000+ | Multiple coordinated agents, custom orchestration, enterprise integration and governance |
| Ongoing running cost, low-volume | $500–$3,200/month | Model/API usage, basic monitoring, low traffic |
| Ongoing running cost, production-scale | $3,200–$13,000+/month | Model/API usage at real traffic volume, monitoring, incident response |
These bands describe build cost for a defined project, plus the separate, recurring running cost that continues after launch, the same distinction our 2027 budgeting worksheet argues every planning exercise should keep separate rather than blending into one number.
Rate bands by provider type
The same feature set quoted by different provider types will land at genuinely different day rates, for reasons that are not always about quality:
- Boutique specialist firms typically carry the highest day rates but the lowest coordination overhead, with fewer people between you and the engineer actually writing code.
- Large agencies often carry lower nominal day rates but higher total project cost once account management, process overhead and larger team sizes are priced in.
- Offshore and nearshore teams can offer the lowest day rates, with the trade-off usually being timezone coordination cost and, in some cases, less direct access to the engineers actually doing the work, worth checking against the "named engineers" question in our vendor-vetting checklist.
None of these is categorically the right answer. The right answer depends on how much you value direct access to named engineers versus lower nominal rates, which is a fair trade-off to make consciously rather than by accident.
Comparing those quotes only works if every vendor is pricing the same scope. Where bids differ wildly on what looks like one project, the usual cause is that each vendor filled the gaps in the brief differently, and priced its own assumptions. Handing all of them one written scope is what makes the numbers comparable: our software development RFP template is the version we use, ungated.
The maintenance line almost every estimate omits
A quoted build cost answers "what does this cost to build." It rarely answers "what does this cost to keep running," and the two are not close. The widely-used industry planning ratio is roughly 15–20% of build cost, per year, for ongoing maintenance, support and small enhancements, before you add the AI-specific running costs in the table above. A $150,000 build with no maintenance line budgeted is a budget for year one only.
What AI-augmented delivery actually changes about the number
It is tempting to assume "AI writes the code now" means costs collapse. The honest picture is narrower than that.
What it does change: the time spent on boilerplate, first-draft implementations, and well-specified, repetitive code can compress meaningfully when a delivery team uses AI tooling well. For work that is genuinely well-specified up front, that compression can show up in the build quote.
What it does not change: discovery, architecture decisions, the judgment calls about what to build and why, code review, and the ongoing maintenance burden are still fundamentally human-judgment-heavy work. A vendor claiming AI collapses the entire cost structure is either overselling the tooling or underselling the parts of delivery that were never really about typing speed.
A worked example
A mid-sized company wants a task-execution agent that reads incoming support tickets, classifies them, drafts a response, and hands off to a human for anything outside a defined confidence threshold, a real and common request in this exact shape.
- Build: lands in the $70,000–$150,000 task-execution band, depending on how many existing systems it needs to integrate with and how much human-review workflow needs building alongside it.
- Running cost: likely in the $3,200–$13,000/month range once it's handling real ticket volume, model costs plus monitoring.
- Year-one maintenance: roughly 15–20% of the build number, for refinement, edge cases found in production, and keeping the model/prompt layer current.
That is three numbers, not one, which is the point of separating them rather than quoting a single blended figure that hides where the ongoing commitment actually sits.
Get an actual range, not an estimate off a table
Every number above is a market range, useful for planning and wrong for your specific project the moment you need a real one. Send a one-paragraph description of what you're building and we'll come back with a scoped range within two business days, NDA-first if you'd rather not describe it in the open. See how engagements are structured or what we actually build, then get in touch with the one-paragraph version.
Common questions
It depends heavily on project class. A focused, single-purpose AI agent typically runs $10,000-$70,000. A task-execution agent integrated into existing workflows runs $70,000-$150,000. A multi-agent enterprise platform runs $150,000 and up, often well past $500,000. These are market ranges observed across published 2026 vendor pricing, not a quote for any specific project.
It changes where the cost sits more than it changes the total. AI-augmented delivery tends to compress the time spent on boilerplate and first-draft code, which can lower build cost for well-specified work. It does not reduce the cost of discovery, architecture decisions, or the ongoing maintenance line. Those stay human-judgment-heavy regardless of how the code was drafted.
Published market ranges put ongoing running cost at roughly $500-$15,000 a month depending on scale, with actively-used production agents commonly landing in the $3,200-$13,000 a month range once model costs, hosting and monitoring are included. Budget this as a recurring operating cost, not folded into the one-time build number.
Mostly because "custom software" describes wildly different scopes under one phrase, and because rate bands differ by provider type: a boutique specialist, a large agency, and an offshore team can quote the same feature set at very different day rates for reasons that have nothing to do with quality.