Most AI documentation treats the model call as the destination. Retool treats it as one node in a workflow that also queries a database, checks a permission, updates a record and notifies someone, and that framing is both more accurate and more useful. Internal tools are where a lot of real AI value sits, meaning summarising support tickets, classifying inbound requests, drafting responses that a human then edits, and the material is oriented around exactly that kind of work. The data connection material is the practical strength.
A model that can see your actual records is far more useful than one answering from general knowledge, and getting that right involves query design, context construction and thinking about what the model should and should not be able to see. The documentation covers this properly rather than hand waving at it, which reflects the fact that Retool's whole business is connecting to company data. Permissions get real attention and that deserves credit. When a model can read customer records, who can run the tool and what it can access are security questions rather than configuration details, and the material treats them accordingly.
That is a maturity you do not often find in AI feature documentation. For existing users, the learning curve is short because AI actions slot into a mental model they already have. That is the whole argument for adding AI to a platform people already use, and it holds up here. The gap is human review.
A tool that drafts a customer response or classifies a ticket is producing output that affects someone, and the design question of where a person checks it, how you make that check easy and what happens when the model is confidently wrong is the difference between a useful tool and an embarrassing one. The material covers building the tool thoroughly and reviewing its output barely at all. Anyone deploying this into a business process should design the review step first and will not be prompted to. Agent material is the weakest section.
It is newer, more speculative and reads like the company is still working out what it believes, with capability described more confidently than the reliability supports. The workflow and action material is the mature part and I would build on that. Pricing bites at rollout. Per user costs are reasonable for a small team and become significant when a tool goes to fifty people across an organisation, which is exactly what happens when an internal tool works.
Model the cost at the scale of success rather than at the scale of the pilot. Platform binding is total, as it always is with low code. What you learn applies here and nowhere else, which is an acceptable trade if you are committed and a real cost if you are still choosing. Three point seven.
Grounded, security aware documentation that puts AI in its proper place inside a working system, marked down for ignoring the human review problem and for agent content that outpaces the evidence.