Why We Stopped Letting the AI “Just Set Up the CRM”, and Built Tools Instead

Published 12 July 2026 · Platform capability · By Theodora Christou

The situation: standing up a CRM is death by a thousand settings

Every time we provision a new CRM/ERP for a client, the same long checklist comes out: create API users, configure dictionaries and custom fields, set up services and tax rules, install document templates, seed demo data, wire up the close-won automation. None of it is hard. All of it is fiddly, order-dependent, and unforgiving, miss one step and the instance behaves subtly wrong weeks later. Doing it by hand is slow and, worse, slightly different every time.

The obvious approach that quietly falls apart

The tempting shortcut in 2026 is to point an AI agent at the system and say “set it up.” For a demo it’s magic. In production it’s a liability, because AI is probabilistic: if every step is 90% reliable, five steps in a row leave you around a coin-flip of getting the whole thing right. Provisioning is *many* steps. An agent free-handing them will get most instances mostly right and occasionally ship one that’s wrong in a way nobody notices until it costs someone. “Let the AI do it all” doesn’t scale, it just moves the errors somewhere harder to find.

What changed: reasoning and execution, split on purpose

We built the provisioning work as a toolkit on a simple principle: let the AI decide what to do, and let deterministic code do it. The setup steps live as plain Python tools, create the users, provision the fields, install the templates, each one consistent, testable, and fast. The AI’s job isn’t to perform those steps; it’s to read the situation, pick the right tools, run them in the right order, and recover when something fails.

Alongside the tools sit written procedures, plain-language SOPs describing each job and how to handle the edge cases, so the knowledge of *how we do this* is captured once and reused, not re-derived on the fly every time. When something breaks, we fix the tool and update the procedure, so the whole system gets more reliable with every run instead of depending on the model getting lucky.

The result: a client CRM goes from empty to configured in minutes, set up the same correct way every single time.

The takeaway

The mistake everyone is making right now is asking AI to *execute* multi-step work directly, then being surprised when reliability collapses. The durable pattern is the opposite: put the probabilistic model where judgement lives, deciding, sequencing, recovering, and put deterministic code where correctness lives. That split is why our provisioning is fast *and* trustworthy, and it’s how we build every automation we ship.

If you’re relying on an AI to run a repeatable process end-to-end and it works “most of the time,” most-of-the-time is the problem. Book a free discovery call and we’ll show you the reliable way to structure it.

Related service: CRM Integration