Some version of this question comes up on most of my sales calls now. "My ops lead built a lead bot over a weekend with Claude Code. It cost about two dollars. Why is your quote five figures?"
It is a fair question, and the honest answer is not "because we write better code." Increasingly, we don't. A good coding agent writes a clean integration faster than any person on my team. If what you are buying from a studio is typing, you should stop buying it.
So I want to answer this properly, including the parts that argue against hiring us. Some of you should build it yourself, and I will tell you how to do that well.
Why pay an AI agency when Claude Code can build it?
Start with what is true. Code got cheap, and it keeps getting cheaper. Nick Saraev, who runs an automation business and talks about this a lot, estimates that the same level of AI capability gets something like 30 to 40 times cheaper every year. That is his estimate, not a law, but the direction is not in doubt.
He makes a sharper point too: the "custom implementation partner" is losing value, because soon everyone will have one a few clicks away. I agree with about half of that. The half I agree with: if a studio's value was turning a spec into working code, that value is collapsing. The half I don't: the spec was never the hard part, and neither was the code.
Look at where the effort goes on a project that is still running six months later.
Model usage to generate the code for a small automation.
Share of a typical small project that is not writing code.
The tip is real, and it is cheap. None of what sits below it is hard to describe. All of it takes time, taste and a few scars. That is the part you are paying a studio for, so let's take it layer by layer.
Judgment: what to automate, what to delete
The most valuable thing I say in a first session is often "don't automate that." A physio clinic in Vancouver wants an AI to chase patients who haven't finished their intake form. Fair enough, until you look at the form: forty fields, a dozen of which nobody at the clinic ever reads. Cut the form in half and most of the chasing disappears. No model required.
A coding agent will build whatever you describe, beautifully. It will not tell you that you described the wrong thing.
Saraev makes another point I agree with completely: build time has nothing to do with value. The automations that earn the most are often the simplest. A twenty-minute flow that replies to every new lead within a minute can be worth more than a three-month agent project. Knowing which one you need is the job, and it is the whole subject of the test we use before building any agent.
Judgment also means deciding where a person stays in charge. Every action the AI could take gets sorted by how bad a mistake would be and whether it can be undone: some run on their own, some need a quick yes, some go to the owner. I wrote about those three lanes and the approval screen separately. A coding agent will happily build a system that sends refunds on its own. It will not stop and ask whether it should.
The interface is where AI projects live or die
An automation without an interface is a pile of scripts that one person understands. It works until that person is on holiday, and then nobody trusts it.
Before LoopSuit I led product design on TD Bank's mobile app. The lesson from that job that transfers best to AI is that people trust systems they can see into. They want to know it is working, what it cost, and what it needs from them, in the thirty seconds they have between other things.
Alex Hormozi has a line about this, aimed at agencies: don't say you're an AI design firm, just be a design firm. Customers buy the outcome, not the machinery. A clinic owner does not want "an AI system." She wants a front desk that never drops a lead, and a way to know it's working without logging into five tools. So we design the screen she will actually open.
Her system, on one screen
This is what an owned, designed system feels like from the owner's side. Not logs. A weekly summary and one decision.
- Monday morning: runs, failures and cost for the week, on one screen.
- Twelve failures were handled without her. She only sees the count.
- One decision is waiting: a model she relies on is being retired.
- Her own eval set shows the replacement matches her approved work.
- She confirms. The switch is logged and reversible for 30 days.
This changes the model in your own AI account. The old one stays selectable for 30 days, so rolling back is one tap.
That screen took longer to design than the automation took to build. It is also why the owner can approve a model change in under a minute instead of emailing a developer and hoping. The interface is not decoration on top of the AI. It is how the AI becomes something a business can run.
Who owns it when the studio leaves?
There are two ways this goes wrong, and they mirror each other.
The DIY version has a name now. Sonny Sangha describes "shadow AI" well: someone vibe-codes a shiny internal dashboard, and within a week it holds customer data, has an API key hard-coded into it, has no sign-on or permissions or audit trail, and IT doesn't know it exists. It works. It is also a liability nobody signed up for.
The agency version is quieter. Everything works, but it runs on the agency's accounts, with the agency's keys, in the agency's repository. The day you want to leave, you find out you were renting. The fix for both is the same checklist, and it should be in the contract, not a nice-to-have.
- Domain, hosting and AI accountsIn your company's name, billed to your cardStudioYou
- API keys and credentialsIn your password vault, never in the codeStudioYou
- Code repositoryTransferred to your organizationStudioYou
- RunbookWhat breaks, how you'll know, what to doStudioYou
- Eval setReal cases you can rerun before any changeStudioYou
- Model switch-out planWhere to move if prices or rules changeStudioYou
Ownership has a second half that gets less attention: designing so there is little to maintain. Saraev says he caps real maintenance at around three to four hours a month per client, and if something needs more than that, it is probably the architecture, not bad luck. I use the same test.
The classic example is a lead form with a free-text "budget" field. People type "around 5k?", "5-8L" and "depends", and the automation that reads it breaks every week. You can pay someone to fix it every week, or change the field to a dropdown and never think about it again. Validated fields beat free text. Retries beat manual reruns. Alerts that reach a named person beat logs nobody reads. If your system needs more than a few hours of support a month, the design is wrong, whoever built it.
When building it yourself is the right call
Here is the part most agencies leave out. If your process is simple, the stakes are low, and you have a technical person with a few days to spare, build it yourself. Claude Code, Codex and tools like n8n are genuinely good enough. We tell people this on calls, and some of them never pay us anything. That is fine.
Two questions sort most projects: how messy is the process, and how bad is it if the system gets something wrong?
- Weekly sales summary posted to SlackBuild it yourself
- Call notes turned into draft posts for reviewBuild it yourself
- Website lead to CRM, with an instant WhatsApp replyBuild it, get it reviewed
- Stock sync across three sales channelsBuy a tool, or build slowly
- Invoice matching into your accounting systemBring in a studio
- Refund approvals across four clinic locationsBring in a studio
If you are in the bottom-left, here is how to do it well. This is the same checklist we work from.
- Write the workflow down before you prompt. The trigger, the steps, who decides, and what "done" looks like. Half the bugs are really unclear decisions.
- Create every account in the company's name. Never on someone's personal email.
- Keep keys out of the code. Environment variables or a password vault. Ask the coding agent to check for hard-coded secrets before you ship.
- Validate inputs at the source. Dropdowns, number fields, required fields. Fix the form, not the parser.
- Save 20 to 50 real examples with the right answers. That is your eval set. Rerun it before you change a prompt or a model.
- Put a person in front of anything that sends money or messages customers. At least until the edit history says otherwise.
- Log every run and alert a named person on failure. Silent failures are the expensive ones.
- Set a spending cap on your AI account and know your cost per task.
- Write a one-page runbook while it's fresh. Future you is a stranger.
Do those nine things and you have done most of what a good studio does on a small project. If the list looks like more than you can own, that is useful information too.
What an AI studio should cost, and what you keep
Whoever you hire, judge the price by what you own at the end, not by hours or lines of code. Here is our ladder, with that column included. You can stop at any rung.
| Step | What it buys | What you keep |
|---|---|---|
| AI Clarity Session $1,500 | One workflow, one working session: automate it, delete it or leave it alone, and whether to build it yourself | A written decision and a lane map for that workflow |
| AI & Product Blueprint from $4,000 | The lanes, the screens, the data and the build plan, before any code | A spec any competent builder can work from, including your own team |
| Custom Build from $12,000 | The automation and the interface, tested on your real cases | Code, accounts, keys, eval set and runbook, all in your name |
| AI Ops from $2,500/mo | Monitoring, model changes, eval reruns and small improvements | A monthly report. Cancel any time and keep everything. |
| Embedded AI Partner from $6,000/mo | A senior AI and product design partner inside your team, working a roadmap | A team that gets better at doing this without us |
Two things about that table are deliberate. The Blueprint is written so that anyone can build from it, including a cheaper shop or your own developer with a coding agent. And AI Ops is optional by design: if a system we built needs us every month just to keep running, we built it wrong.
We are a Canadian studio in Vancouver with a second base in New Delhi, so a clinic group in Burnaby, a founder in Austin and a real-estate team in Gurgaon get the same process, in overlapping working hours.
How we build this at LoopSuit
The order never changes. We start with judgment: one workflow, sorted into what runs on its own, what needs a yes and what goes to the owner, and a hard look at what should simply be deleted. Then we design the screens people will actually open, before any prompt is written. Then we build, using the same coding agents you could use, inside accounts that belong to you from day one. We test on your real cases, set up alerts that reach a person, and hand over the keys, the eval set and a runbook your team can read.
The code is the cheapest part of that, and we are happy to say so. If a two-dollar weekend build does the job, use it. Pay a studio for everything under the waterline, and if you are not sure which side of the line you are on, that is exactly what a Clarity Session is for.
Questions people ask us
Is it cheaper to build AI automation yourself?
The code, yes. Tools like Claude Code and Codex can write a working automation for a few dollars of usage. The real costs are your time on discovery, edge cases, testing and upkeep. For a simple, low-stakes workflow with a technical person in-house, building it yourself usually wins. For customer-facing or money-moving work, the hidden costs usually outweigh a studio's fee.
How much does custom AI development cost?
It ranges widely. At LoopSuit a Clarity Session is $1,500, a Blueprint starts at $4,000, a Custom Build starts at $12,000, and ongoing AI Ops starts at $2,500 a month. Whoever you hire, ask what the price includes beyond code: discovery, interface design, evals, documentation and handover.
What should I own at the end of an AI agency project?
Everything: the domain and accounts in your company's name, API keys in your own vault, the code repository, a runbook, the eval set used to test changes, and a written plan for switching models. If the agency still holds any of these, you are renting the system, not owning it.
How much maintenance should an AI automation need?
Very little, if it is designed well. A few hours a month is a reasonable ceiling for a typical small-business system. If it needs more, the design is usually the problem: free-text fields where validated ones belong, no retries on flaky services, or failures that nobody gets alerted about.
Do you work with businesses outside Vancouver?
Yes. LoopSuit is a Canadian studio based in Vancouver with a second base in New Delhi. We work with operators across Canada, the US and India, remotely and across overlapping time zones.