Build vs buy AI: keep building it yourself, hire a developer, buy no-code, or hire an agency

Why nobody has written this down

The chief executive of a 19-person accounting firm asked this question in public recently. He was already shipping internal tools with an AI coding assistant, and he wanted to know whether to hire a developer, engage an agency, or carry on himself. The replies were “why don't you ask the AI”, “lol”, and a suggestion that he should not be running a company if he had fallen for the hype.

He was the most qualified person in that thread to ask it. He had a budget, a working prototype, and a real decision in front of him, and what he got back was mockery. So here is the artefact nobody built. The answer is a routing problem with four destinations, and three of them do not involve paying us.

The four paths, honestly compared

We lose four rows in this table, including the two most people care about. That is what the table is for.

Where we lose: cheapest over 18 months (build it yourself, then no-code), fastest to a working version (no-code, in days), and ownership at month 12 (an in-house engineer beats us outright, because they are still there). We win on time-to-scoped-build and on the ceiling for anything with real access rules. Four rows to three.

What each path costs over 18 months

Invoices are the easy part. The numbers below include salary and the time that never appears on a purchase order, because that is the money people forget when they build the case internally.

Salary figures come from the Stack Overflow 2025 Developer Survey; licence prices are n8n's published plans as of 24 August 2026; the $8,000 figure is our own published fixed-fee starting point. Nothing here is a quote and nothing here is estimated on your behalf.

The ten questions, in full

The whole thing as a checklist, so you can run it in a meeting without the tool. Each option pushes toward one or more of the four paths; the numbers in brackets show how hard, in the order build-it-yourself / no-code / hire / agency.

1. 1. Does an engineer on your team have genuine spare capacity?

   Genuine means hours that already exist, not hours you hope to find.

   • [5/1/0/-2]Yes — someone has about a day a week free and wants this work
   • [-1/2/3/2]We have one technical person and they are the only reason production stays up
   • [0/2/2/3]Technical people, but zero spare hours
   • [-2/3/2/3]No engineer at all

2. 2. What have you already built with AI coding tools?

   Be strict. Something people open every Monday counts. A demo does not.

   • [5/0/1/0]Something we built is in daily use and it holds up
   • [-1/2/2/3]We built something, it broke, and nobody fixed it
   • [2/2/1/1]Demos and prototypes, nothing anyone depends on
   • [1/3/0/1]Nothing yet

3. 3. How many systems does this have to read from or write to?

   • [3/5/-1/-2]One or two, and they are everyday SaaS with documented APIs
   • [2/2/1/1]Three or four, all with usable APIs
   • [-1/-2/3/4]Five or more, or several without clean APIs
   • [-1/-3/2/4]One of them has no API at all — a portal, a desktop app, or PDFs by email

4. 4. Does the answer change depending on who is asking?

   This is the question that decides whether no-code can carry it.

   • [2/3/0/-1]No. Everyone who can reach it should see all of it
   • [1/1/1/2]A couple of role rules — managers see more than staff
   • [-1/-3/2/4]Real walls — per client, per matter, per region
   • [-2/-4/3/4]Regulated — someone outside the company audits who saw what

5. 5. If it quietly produced wrong output for a week, what would happen?

   • [3/3/-1/-2]Somebody would be annoyed
   • [2/2/0/0]A few hours of rework
   • [-1/-1/2/3]Money moves, or a client notices
   • [-2/-2/3/3]A filing, a contract, or a regulator is involved

6. 6. Is this one workflow, or the first of many?

   • [3/3/-3/0]One workflow. Build it and we are done
   • [1/2/0/3]This one, then probably two or three more
   • [-1/-2/4/3]Several teams will build on top of it
   • [-2/-2/3/3]Whole-company knowledge, everyone touching it

7. 7. Do you know what is missing and what needs doing at regular intervals?

   This is the gate. A paid build needs a repeatable problem, not a promising one.

   • [1/2/2/4]Yes — written down, same steps, known cadence
   • [2/2/0/0]We know the area. Not the cadence
   • [3/3/-3/-6]Still exploring what AI could do here

8. 8. Is there one area, one number, and an appetite for training?

   All three together is the condition under which our own engagements go best.

   • [1/1/2/4]One area, one named number, and we accept our team will need training
   • [2/2/1/0]One area and a number, but we expect the tool to just work
   • [2/2/0/-1]One area, no number
   • [0/1/1/-2]Several areas at once

9. 9. Is the work written down — clear KPIs and clear SOPs for the role?

   This is the strongest single signal we know that an agent will work. It matters more than the size of your team.

   • [0/1/0/4]Yes — the KPIs are defined and the steps are written down, and the work is mostly digital
   • [1/2/1/2]The KPIs are clear, but the steps live in someone's head
   • [1/2/1/2]The steps are written down, but nobody has agreed what good looks like
   • [3/2/1/-3]Neither, or the role is mostly physical or relationship-driven

10. 10. It is month 12 and a vendor changes an API. Who fixes it?

    • [4/3/1/-1]A named person who has the time
    • [-1/1/3/3]A named person who is already at capacity
    • [-3/-1/3/3]Honestly, nobody
    • [-2/0/0/4]Whoever we paid to build it

11. 11. When does this need to be working?

    • [3/2/2/0]No hard date
    • [1/3/-1/2]This quarter
    • [-1/3/-4/3]Six weeks — committed to a board or a client
    • [-1/4/-5/2]It was due last month

How the recommendation is decided

Every option adds or subtracts points from each of the four paths, and the highest total normally wins. Five rules sit on top of the points, and two of them exist purely to stop this tool recommending us.

1. Spare capacity plus a small surface means build it yourself. An engineer with a genuine free day a week, four systems or fewer, nothing regulated and no legal consequence — that combination is decided before the other answers get a vote.
2. An everyday stack with no access walls means buy no-code. One or two documented-API SaaS systems, everyone sees the same output, and a bad week costs rework rather than money. Custom code there is flexibility you will not use.
3. Three walls take no-code off the table entirely. Real per-client or per-matter access rules, a regulator who can ask who saw what, or a system with no API at all. Any one of those and no-code is removed from consideration rather than just penalised, because a pile of small advantages elsewhere should not be able to out-vote a hard limit.
4. Still exploring means no paid build. If you say you are still working out what AI could do here, the recommendation is demoted to build-it- yourself or no-code even when the points say otherwise. This is the founder's own rule, used verbatim on sales calls.
5. The agency gate. An agency can only be recommended when the process repeats on a known cadence and there is one area with one named number and the team accepts it will need training. Two out of three demotes the answer to the next-highest path.

Rule 5 is the reason this tool sends a lot of people away. Those three conditions together are the ones under which our own engagements go best — one area to double down on, usually sales and new prospects, a specific number to move, and an accepted training burden. Without them a build still gets delivered, and nobody can tell you afterwards whether it worked.

The strongest signal that an agent will work: written KPIs and written SOPs

If a role has clearly determined KPIs and clearly determined SOPs, and the work is mostly digital, it is very likely an agent can do that work better — or at least take a real part of it off the person doing it now. That is the single best predictor we have, and it matters more than headcount, budget or which tools you already use.

The reason is unglamorous. Writing the KPIs forces you to say what good looks like, and writing the SOPs forces you to say what actually happens step by step. Those two documents are most of the specification for an agent. Companies that have them get a build that works, because the hard thinking was already done. Companies that skip them end up paying someone to guess, and then judging the guess on impressions.

So if you have both, you are further along than you think, and the honest answer is usually yes, build the agent. If you have neither, the cheapest next step is not a vendor — it is a week of writing down what the role actually does and what number it is supposed to move.

When we tell people not to hire us

If the team is small and new to this, they should wire the automations themselves with a no-code tool, rather than commission a build. Only once the problems are repeatable — once you know what is missing from your workflows and what needs doing at regular intervals — is a paid build the right call. The founder's phrasing on a recent call was blunt: “For those people, we mentioned: do not hire us.”

This is not modesty. A build commissioned during the exploring phase produces an expensive, well-engineered answer to a question that changes two months later, and then it is nobody's job to change it. The same money spent after the process repeats buys something people actually use.

Anthropic's own engineering guidance lands in the same place, from the technical side rather than the commercial one: find the simplest solution possible, and only increase complexity when needed — which, in their words, “might mean not building agentic systems at all”. The same instinct applies to who builds it.

When hiring an agency is the right call

Three conditions, all at once. One area you want to double down on — sales and new prospects is the most common first pick. One named number to move, with a value you could read today. And an acceptance that your team will need training, because the system being correct and the team knowing what to hand it are two different projects.

Add a fourth practical condition of your own: ask what happens in month 13. The month-12 ownership row is where this path is weakest, and it is the row nobody raises during a sale. Get the maintenance arrangement written down with a response time and a price, or get the training written into the scope so somebody inside the company can open the code.

Our own shape, so you can compare it against anyone else you are vetting: a fixed fee from about $8,000 for a scoped build, first deployment in two to four weeks, and the detail on how we run an AI agent build.

Should you hire an AI developer in-house?

Hire when this is a capability rather than a project. Several teams building on the same foundation, work that keeps arriving, and no deadline inside six months — that is a job. One workflow with a six-week deadline is not, and no amount of urgency makes the arithmetic work: search, notice period and ramp-up eat the entire window.

Two things go wrong with these hires and both are avoidable. The first is casting: people write the job around the interesting build and then hand over a maintenance role on four small tools, and the hire leaves in month 14. Write the job around the maintenance. The second is the single point of failure — one person holding the only mental model of the system. Whatever you build, make somebody else able to read it.

On cost: the Stack Overflow 2025 Developer Survey median for a United States AI/ML engineer is $189,500, and for a back-end developer $175,000. That is base pay, before benefits, payroll taxes, equipment and the recruiting time. Over 18 months the AI/ML figure is roughly $284,000 in salary alone, which buys a great deal of anything else on this page.

Is no-code enough, or will you outgrow it?

n8n, Make and Zapier carry more than people expect, and the honest limits are specific. No-code holds while the surface is everyday SaaS with documented APIs, everyone sees the same output, and failure means rework. It stops at three walls: per- user or per-client permissions, a system with no API, and error handling more complicated than retry three times.

Hitting a wall is not evidence you chose wrong. It is the cheapest possible way to discover the real requirement, and the flow you built is the specification for whatever replaces it. For the difference between a fixed automation and something that has to make a judgement call, read AI agents vs workflows vs RPA. If you are not yet sure the job needs an agent at all, start with the do-you-need-an-agent check — a lot of work routes to a plain scheduled script, and that is a win.

What to do before you talk to anybody

Two weeks of preparation changes what you get quoted, whichever path wins. Log what actually happens in the process for ten working days: the steps, the order, how often, and what the finished output looks like. That log is the cheapest scoping document in existence, and it routinely changes what people ask for.

Then name the number and name the owner. If you want to know whether the rest of the conditions are in place before you commit money, the AI readiness assessment covers the ones this tool does not: where your answers live, who approves what, and whether the knowledge is in a state anything can retrieve. On that last point, not more than a quarter of the clients who come to us arrive with their knowledge in a usable state, so budget for the cleanup work rather than being surprised by it.

The rest of our free tools are in the tools library.

Frequently asked questions

Sources

• Stack Overflow, 2025 Developer Survey — Work — median United States salaries used throughout: AI/ML engineer $189,500, back-end developer $175,000. Self-reported survey data, not a government wage series.
• Anthropic, Building effective agents — the guidance to find the simplest solution possible and only increase complexity when needed, which “might mean not building agentic systems at all”.
• n8n, published pricing — Starter €20 a month and Pro €50 a month, checked 24 August 2026. Prices change; check before you budget.
• Everything marked “from our deployments” or attributed to a client is our own first-hand experience across Cognio Labs builds, anonymised. It is observation, not research, and we have not rounded it in our favour.