AI makes writing code much cheaper and faster, but writing code was never the whole cost of building software. In 2026, the budget for an AI-assisted build shifts toward the parts AI doesn't do well: deciding what to build, writing clear specs, reviewing what the AI produces, integrations, data migration, and the running cost of any AI features inside the product. If a quote assumes AI makes everything cheaper, it's probably missing those.
We build our own products and client platforms this way. Here's where we see the cost actually go.
Where AI genuinely cuts cost and time
Let's start with the good news, because it's real.
Implementation of well-specified work. Give an AI coding agent a clear ticket (what's wrong, what done looks like, which files, how to check it) and it builds the thing quickly and usually correctly. This is the biggest saving, and it's why a small studio can ship a lot of changes in a week.
Small changes stop being expensive. When building a small improvement takes minutes rather than hours, you stop batching them up and stop arguing about whether they're worth it. Combined with fast deploys (ours take about nine minutes), small, checked changes go out the moment they're ready.
Tests and boilerplate. The tedious, necessary work of writing tests, wiring forms, adding validation and repeating a pattern across many screens is where AI is at its most reliable.
Quality sweeps. Work that used to get deferred forever because it was spread across hundreds of files, like accessibility, becomes a week's focused effort. We did accessibility across three codebases in one week, with automated checks so it doesn't slip back.
Drafting. First drafts of specs, help articles, follow-up tickets and copy. A person still edits and decides, but starting from a draft is faster than starting from nothing.
Where AI doesn't cut cost
This is the part that budgets tend to miss.
Product decisions
AI can implement any decision you give it. It can't tell you which decision your customers need.
When we added tax to project costs in Handl, the work that mattered was deciding that agencies who reclaim GST or VAT should see margin on the net figure, that it should be a setting, and that it should ship switched off. When we killed Mortar's pricing tiers in favour of one plan, the code was trivial. The decision was the work. None of that got cheaper.
Discovery and scope
Understanding the users, the market and the assumptions still takes conversations, research and time. If anything, cheap code makes this more important: it's now very easy to build the wrong thing quickly. Our post on how to scope a SaaS MVP covers why getting this right is most of the battle.
Review
Every change still needs reviewing before it ships, and anything touching money, logins or customer data needs a deeper, adversarial review. AI-written code arrives faster, which means there's more of it to review. Review time doesn't shrink in proportion to build time. Plan for it. We explain how we tier it in reviewing AI-written code that handles money.
Integrations
Connecting to someone else's system costs what it costs. You're working to their documentation, their sandbox, their quirks and, sometimes, their approval process. When we built Xero and QuickBooks sync for Mortar, the code came together quickly, and then it went through the app store review process, which runs on someone else's timeline.
Every external system in your build (payments, accounting, CRMs, industry data feeds, single sign-on) is a line in the budget that AI only partly reduces.
Data and migration
Moving content and data from an old system to a new one is detail work: mapping every old address to a new one so bookmarks and search rankings carry over, checking every link and asset after launch, and deciding what to do with records that don't fit. On our association projects, the migration (full URL maps, redirects, re-checking after launch) is a real chunk of the effort, whatever wrote the code.
Reuse beats AI
The biggest time saving we've seen on a client project didn't come from AI at all. Columbus REALTORS® went live in twelve weeks because the association-specific parts already existed in TaliCMS. Nobody had to invent a member directory, an events system or a document library on the client's budget. The work was configuration, design, content migration and the data feed. If an existing product already does most of what you need, that's usually a bigger saving than anything an AI agent can offer.
The hidden costs
These are the costs that don't show up in "AI builds it faster" estimates.
Rework from vague specs
An AI agent given a vague request will confidently build something. If it's the wrong something, you pay twice: once to build it, once to unpick it and build the right thing. The fix is writing proper specs up front, which takes people's time. Budget for the specs, or budget for the rework. You'll pay one of them. We go into how we write them in spec-driven development with AI agents.
Token and agent spend
AI agents cost money to run, and the cost scales with how you use them, not just how much you build. The expensive habits are easy to fall into:
- Running many agents in parallel on files that overlap, so they collide and need rework and re-testing.
- Re-running the full test suite locally on every small fix instead of running the relevant tests and letting CI run the rest.
- Agents that keep running after their useful output is done: polling, narrating, handing work to more agents.
- Using the most expensive review everywhere instead of where the risk is.
We learned this by doing it. Running more builders at once looked productive and mostly produced conflicts. We now run at most three at a time, merge in a fixed order, use scoped tests while building with CI as the authority, and stop agents the moment their work is done. Spend discipline is part of the engineering process, not an afterthought.
Running costs of AI features
If your product has AI features inside it, every use costs something, forever. That's a per-customer running cost, not a one-off build cost, and it affects your pricing.
When we prepared First Shift, Arbeo's home care hiring product, for launch, we cut the AI cost per application significantly: a smaller model where it's good enough, and rolling summaries instead of re-reading the whole conversation every turn. Unit economics matter when you're charging small businesses small amounts. Budget for running costs, and budget for the work to bring them down.
Waiting time
Some time isn't spent building at all: waiting for a third party's approval, for a client to provide content, for a domain or account to be set up. AI doesn't touch any of it.
How to budget an AI-assisted build
If you're planning a budget, here's how we'd think about it.
| Area | Does AI reduce it? | What to plan for |
|---|---|---|
| Discovery and scoping | Barely | Time with users and decision-makers |
| Product and design decisions | No | Senior judgement, written down |
| Writing specs | Somewhat (drafting) | People's time to make them precise |
| Building well-specified features | A lot | Less than you'd have budgeted a few years ago |
| Tests | A lot | Still needs CI and maintenance |
| Review | No, and there's more to review | Tiered review, deeper on money and data |
| Integrations | Partly | Third-party docs, sandboxes, approvals |
| Data migration | Partly | Mapping, checking, edge cases |
| AI running costs | n/a | Ongoing, per use |
| Launch and after | Partly | Docs, monitoring, iteration |
A few practical rules:
- Ask what's being reused. An existing platform or product that fits most of your need usually saves more than AI does.
- Ask how review works. A cheap quote with no clear review process is cheap because something is missing.
- Ask about running costs for any AI feature, not just build costs.
- Spend early on specs. Clear requirements are the cheapest place to save money in an AI-assisted build.
For the general cost drivers of a SaaS build, see how much it costs to build a SaaS product and what it costs to hire a development studio. The principles we build by are in how we build.
FAQ
Does AI make software cheaper to build?
It makes writing well-specified code much cheaper and faster. It doesn't make product decisions, discovery, review, integrations or data migration much cheaper, and those are a large part of most builds.
What are the hidden costs of building with AI?
Rework from vague specs, review time for a larger volume of code, the cost of running AI agents inefficiently, and the ongoing running cost of any AI features inside the product.
How should I budget for an AI-assisted software project?
Budget properly for discovery, specs and review, expect implementation to cost less than it used to, and plan separately for integrations, migration and AI running costs. Ask any studio what they're reusing and how they review.
Can AI replace a development team?
No. AI writes a lot of code, but people still decide what to build, write the specs, review the work, own decisions about customers and money, and choose what ships.
If you're budgeting a build and want a straight view on what it involves, talk to us.
Consulting Inquiries
