AI at the center or AI as the glue: two ways to build a harness

For most tasks, the model no longer makes the difference

For most tasks, what a company expects from AI comes down to precise, repeated operations: linking two pieces of information, summarizing an exchange, correcting a text, sorting signal from noise in an inbox. Almost every recent model does these well; picking one over another changes almost nothing.

The real choice lies elsewhere. You can aim wide: real, reliable time savings for most teams, on tasks that come back every day. Or you can aim high: hand the work of your sharpest experts to more autonomous agents, and accept that security goes from near-certain to uncertain. The first choice pays more across a whole company, without giving anything up on security. It is the one PipeDuck serves.

It still takes an architecture that makes these tasks reliable everywhere. That is where harnesses, everything around the model, split into two families.

Two ways to build a harness

In the first family, AI sits at the center. You give it a goal and access; it explores, decides on its calls, reads what comes back and starts again until it reaches its goal. Coding agents work this way. It is the right architecture when the task is open-ended and a specialist reviews the result, provided the specialist runs the agent on a machine with tightly restricted access. In many companies, developers no longer have write access to production databases, and sometimes not even read access, to prevent leaks; the agent, working with their rights, inherits those limits.

The second family, PipeDuck’s, reverses the relationship. The process is drawn, and each step takes the form that suits it: a script, cheap and deterministic, so reproducible and reliable; an AI step for the complex cases a script handles badly, or for low-stakes decisions; a person for the big decisions, and to carry the responsibility for them. At design time, AI helps turn your ideas into steps and connect them: that work is paid for once, and a person reviews it. At run time, it only works in the steps that belong to it, with the model and tools chosen for each. Code calls AI more than AI calls code.

The daily Excel report, in both architectures

Every morning, an export of orders lands in an Excel file, and someone has to pull out revenue by region, the best-selling products and the anomalies to deal with.

In the first architecture, an AI receives the file every day, reads all of it and calculates. We ran the test on an export of 1,000 orders containing eight randomly placed anomalies, with Claude Opus 5 set to its maximum reasoning effort and no calculation tool. The expected report contains 23 results: total revenue and revenue for each of the six regions, the five best-selling products with their amounts, the number of orders for each of the four statuses, the average basket, and the list of anomalies, which must find all eight without inventing any. All 23 are correct. The work took 16 and a half minutes and nearly 116,000 tokens, or about $3.24 at the public API price¹. The model got it right because it redid every calculation along two independent paths, which is also what made it expensive.

In the second architecture, AI is used once. PipeDuck’s build assistant helps write the report’s rules, a script applies them, and a person reviews that script before it runs. After that, every morning, the same report comes out in 6.5 milliseconds, with no tokens, and without any model reading customer data.

Over 250 working days, the first architecture costs about $810 for a single report, and every morning keeps a chance of error. Our test was right once; it says nothing about the other 249 days. And an AI that succeeds on 99% of its runs gets through a whole year without an error only about one time in twelve. Verified rules return the same result every day.

Compared with Claude Code, n8n and Zapier

Claude Code, which we use ourselves and with which we ran this test, is an excellent harness, for code: an agentic tool that reads a codebase, edits files and runs commands, within limits set by permissions and by an operating-system sandbox³. AI is at the center there, which suits an open-ended task reviewed by a developer. A business process that comes back every day, run by people who do not program, calls for the opposite architecture.

n8n and Zapier are the closest competitors, and they do a lot well: AI steps where you choose the model, human approvals, and credential sharing that does not reveal the secret^{4, 5}. Our differences lie elsewhere. In PipeDuck, every version of a script is reviewed before it can run: an AI-written summary of the changes and security notes, then a person’s approval; in n8n, that review goes through a pull request at the Git provider, outside the tool, on the Business and Enterprise plans⁴. And PipeDuck is hosted in France or installed on your premises, with your own model if you want, whereas Zapier only exists as a cloud service⁵. Above all, the paths differ: n8n and Zapier were born to connect applications and added AI to that; we started from a company’s rules and from reviewed code, and placed AI between the two.