On 6 October 2025, OpenAI unveiled Agent Builder at DevDay. A visual canvas where you drag boxes together and get an agent out the other end. Eight months later, on 3 June 2026, it was deprecated. It shuts down for good on 30 November 2026.
Nobody did anything wrong there. OpenAI decided the visual builder was not the path and pointed users at the Agents SDK instead. But if you built your order processing on that canvas in January, you spent the summer migrating instead of shipping. That is the part of the "how do I build an AI agent" question that almost nobody asks about in advance, and it is usually the part that decides whether the thing is still running in two years.
There are four honest routes to a working agent. They differ less in what they can do than in where they break. Here is what each one actually costs, where its ceiling sits, and which kind of company each one fits.
The Decision You Are Actually Making#
Most comparisons start with features. That is the wrong end. Every platform on this list can call an API, read a document and write to a CRM. Feature lists converged eighteen months ago.
What has not converged is the answer to four questions:
What does the agent touch? An agent that drafts a reply is a different risk class from one that issues a refund. The first can be wrong and cost you nothing. The second can be wrong and cost you money, twice, because you also have to unwind it.
Who is liable when it is wrong? Not philosophically. Operationally. Which human finds out, how fast, and what do they do about it. If the answer is "nobody, until a customer complains", the agent is not ready regardless of platform.
How often does the process change? A process that changes twice a year suits a rigid platform fine. A process that changes twice a month punishes anything you cannot edit yourself.
Is this your competitive edge or your plumbing? Nobody wins a market with better invoice routing. Buy the plumbing. Build the edge.
Answer those four and the route picks itself. Skip them and you end up in the Gartner statistic: the analyst firm expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, and the three reasons it gives are escalating costs, unclear business value and inadequate risk controls. None of those are technology problems. All three are decision problems that got deferred.
Route One: No-Code Builders#
Chatbase, Lindy, Relevance AI, Gumloop and a dozen others. You upload documents, paste URLs, connect a few tools through a visual editor and have something answering questions the same afternoon. Across the platforms with a published flat entry plan, the median first paid tier sits around 24 dollars per month.
This route is genuinely good at one thing: proving the use case is real before anyone spends real money. I have watched a company argue for three weeks about whether an agent could handle their intake questions. A Saturday afternoon on a no-code builder ended the argument, in the direction nobody expected. Two of the four question types were trivial. The other two needed a human, permanently.
The ceiling arrives faster than the sales page suggests, and it arrives in a specific place: anything that requires a decision the platform did not anticipate. You can usually connect tools. You can rarely control what happens between them. When the agent needs to check one system, decide based on what it found, then take a different path through a second system, most no-code builders either cannot express it or express it so awkwardly that the workflow becomes unmaintainable.
The second ceiling is data. These are hosted services. Your customer records pass through infrastructure you do not control, under terms you did not negotiate. For a public FAQ that is fine. For anything involving personal data of EU customers, read the terms before you upload, not after.
Fits: a company testing whether an agent helps at all, or running something small and public-facing that touches no sensitive data.
Route Two: Enterprise Suites#
Salesforce Agentforce and Microsoft Copilot Studio are the two that matter, and their pricing tells you almost everything about who they are for.
Agentforce charges 2 dollars per conversation. Or 20 Flex Credits, about 10 cents, per action. Or 125 to 150 dollars per user per month for a named licence. Copilot Studio charges in Copilot Credits: 200 dollars for a prepaid pack of 25,000, which works out around eight tenths of a cent each, or a penny each pay as you go.
At moderate interaction depth, a Copilot Studio conversation lands somewhere near 18 cents while an Agentforce conversation stays at 2 dollars flat. That looks like an eleven-fold gap, and every Microsoft sales deck will show you exactly that comparison.
The comparison is also useless on its own, because both numbers assume you already own the thing underneath. Agentforce in practice needs Service Cloud plus a Data Cloud subscription that lists around 60,000 dollars a year. Copilot Studio assumes you already pay for the Microsoft 365 and Power Platform tenant. If you have that tenant, Copilot Studio is close to free to start. If you do not, the entry price is the whole Microsoft stack.
The real trade in this route is not price. It is that you get governance for free. Audit logs, permission models, data residency, retention policy, the whole compliance apparatus that takes months to build yourself, arrives on day one and satisfies your legal department without a conversation. For a regulated company that is worth more than the licence fee.
The cost is the ceiling, and it is a hard one. Both platforms are excellent inside their own ecosystem and awkward outside it. The moment your agent needs to reach a machine on your own premises, a niche piece of industry software or a database that predates the cloud, you are writing custom connectors anyway, and the platform stops paying for itself.
Fits: companies that already live inside Salesforce or Microsoft 365, where the agent's work stays inside that world, and where compliance sign-off is the slow part of any project.
Route Three: Workflow Platforms With Agent Steps#
n8n is the clearest example, and it is the route most small and mid-sized companies I talk to end up on without planning to. The Community Edition is free and self-hosted, so you pay for a server and for the model calls, nothing else. Cloud Starter runs 20 euros a month for 2,500 executions, Pro 50 euros for 10,000.
The mental model is different from the other two routes and that is the point. You draw the process as a flow with explicit steps, and at the points where judgement is needed you put a model in. Everything else stays deterministic. Read the mailbox, always. Extract the fields, always. Decide which of three departments this belongs to, model. Write to the ticket system, always.
That structure quietly solves the biggest reliability problem in this whole field. Chain enough model decisions together and the maths turns against you. If each step is right 95 percent of the time, which sounds fine, twenty steps in sequence come out right about 36 percent of the time. Reliability is multiplicative and it collapses faster than intuition suggests. A flow with three model decisions and seventeen deterministic steps is a fundamentally different machine from one with twenty model decisions, even though both look like "an agent" on a slide.
Self-hosting matters more in Europe than the international comparisons acknowledge. In the Bitkom survey of 604 German companies conducted in early 2026, uncertainty about data protection was the single most-named obstacle to using AI at all, cited by 41 percent. Running the orchestration on your own server does not make you compliant by itself, but it removes the question that stalls the project for six weeks.
The cost is that low-code is not no-code. Expressions, error handling, retries and custom nodes all reward someone technical. A workflow platform in the hands of someone who has never debugged anything becomes a pile of half-finished flows that nobody trusts.
Fits: companies with a defined process, some technical capacity in-house or on retainer, and a reason to keep data on their own infrastructure.
Route Four: Building It Properly#
At the far end you assemble the agent yourself. A model chosen for the task, a loop around it, tools defined as real interfaces, memory that survives restarts, guardrails that sit between the decision and the action, and an evaluation set that tells you whether last week's change made things better or worse.
Industry estimates for a custom workflow agent cluster between 25,000 and 100,000 dollars, with multi-agent systems running well past that. Treat those as orders of magnitude rather than quotes, because the range depends almost entirely on how many systems the thing has to touch.
What you get for it is the only route with no ceiling and no landlord. Nobody deprecates your architecture in eight months. Nobody reprices your per-conversation rate. When the process changes, you change the agent.
The connective tissue that made this route cheaper than it was two years ago is the Model Context Protocol. Anthropic released it as an open standard in November 2024 and handed it to the Agentic AI Foundation under the Linux Foundation in December 2025. OpenAI, Google, Microsoft, IBM and Amazon all support it. SDK downloads went from 100,000 at launch to 97 million a month by March 2026. Practically, that means the connector you write for your ERP works with whichever model you point at it, and swapping the model later is a configuration change rather than a rewrite. That single property is why building custom no longer means betting your architecture on one vendor.
The cost is that you now own an operational system. It needs monitoring, it needs someone who understands it, and it will surface edge cases in month four that nobody imagined in month one.
Fits: a process that is genuinely yours, touches systems no platform knows about, handles data that cannot leave your control, or is close enough to your product that it should not be rented.
What the Four Routes Look Like Side by Side#
| Entry cost | Ceiling | Breaks when | Your data lives | |
|---|---|---|---|---|
| No-code builder | ~24 USD/month | Low, hits fast | The decision gets conditional | On their servers |
| Enterprise suite | Stack you already own | Medium, ecosystem-bound | You step outside the ecosystem | Their cloud, contractually |
| Workflow platform | 0 to 50 EUR/month | High with technical help | Nobody maintains it | Wherever you host it |
| Custom build | Five figures upwards | None | Nobody owns it internally | Wherever you decide |
The row that matters is the third one. Every route works on day one. They differ in how they fail in month nine, and the failure is almost never technical.
Why Most of These Projects Die#
Two numbers frame this honestly. MIT's Project NANDA found in August 2025 that 95 percent of generative AI pilots showed no measurable contribution to profit and loss. Gartner puts more than 40 percent of agentic projects on the cancellation list by the end of 2027.
Read together, and against what I see in practice, the pattern is not that agents do not work. It is that companies pick a process that was never the bottleneck. Automating something that took a person twenty minutes a week produces a working agent and no measurable result. The 20 minutes go somewhere else, nobody notices, and when the budget review comes the project has nothing to point at.
The counter-move is unglamorous. Pick the process where somebody is visibly drowning. Measure it before you touch it. Automate the boring middle and leave the judgement calls where they are. Then the number in the review is a real number.
There is a German data point worth holding next to that. The Salesforce and Deutscher Mittelstands-Bund index published in March 2026 found 16.6 percent of German mid-sized companies already using AI agents, up from 8.7 percent a year earlier. A doubling. But only 21 percent of companies in the Bitkom survey had an AI strategy at all. That gap between deployment and strategy is exactly where the cancelled projects come from.
How We Approach It#
We build agents as part of our AI systems work, and the process is deliberately boring at the front.
First we look at the actual task with the person who does it today. Not the process diagram. The task, including the exceptions they handle without thinking about them, because those exceptions are what breaks demo agents in week two.
Then we scope hard. One task, clear boundaries, defined inputs and outputs. An agent that tries to do everything becomes something nobody can debug and nobody trusts.
Then we decide the route from the four above, based on those four questions, not on what we would enjoy building. Plenty of jobs are a workflow platform with three model steps and no more, and saying so is part of the work. When the job genuinely calls for a custom system, we build on open standards, keep the model swappable, put a human gate in front of anything irreversible, and host it where the customer wants it, including on their own hardware.
Then we test with real cases from the customer's own history, not invented ones. Real cases are messier and they find the holes.
And then it runs, with somebody responsible for it.
Our own operation runs on this. The agents that research, monitor, prepare and report for us are the same class of system we hand over, which is a reasonable standard to hold a supplier to: ask whether they use the thing they are selling you, every day, on their own business. If not, ask why.
The Part That Will Not Change#
Models will keep improving and the gap between the four routes will keep narrowing on capability. It will not narrow on ownership. In five years the question "who can switch this off, reprice it or discontinue it" will matter more than it does today, not less, because more of the daily work will be running through it.
Agent Builder ran for eight months. That is not an argument against platforms. It is an argument for knowing, before you start, what happens to your work when a platform decides something different. If you want to talk through which of the four routes fits a specific process in your company, that conversation is what our agent work starts with.
