AI Chat is live in our workspace and I put it through a real end-to-end test today — not a toy prompt, an actual project build. The short version: it’s promising, but it can’t yet do real work, and the most useful evidence for you came from the assistant itself. I’ve attached its own self-assessment. Here’s the full report.
QUICK CONTEXT ON HOW WE WORK
Rentman is our source of truth for projects, planning, crew and equipment. We drive it from Claude as our daily work hub: through your official MCP we query project status, planning and finance data all day, and on top of the public API we’ve built our own automations — a Rentman-to-Pipedrive mirror that keeps our CRM aligned with Rentman in near real time, warehouse prep tickets printed automatically two days before each confirmed shoot, and monthly P&L reporting. “AI + Rentman” is already how we operate, every day.
THE TEST
I asked AI Chat to build one of our real projects the way we actually would: add equipment, create a shoot period per day, add the Movement Designer + Robot Assistant crew functions on USA pricing, group them per day, and produce a quote. A normal project setup for us.
It got a long way — created the day-by-day periods, created the functions, grouped them — but hit a wall on nearly every step that wasn’t a plain create, and produced a lot of collateral damage on the way (20 duplicated/unnamed functions it then couldn’t clean up). What went wrong, concretely:
- No delete. It repeatedly asked ME to delete its own duplicates by hand. An assistant that can’t undo its own mistakes can’t be trusted to act unsupervised.
- searchItems can’t see project functions (Functie). After creating functions it couldn’t re-find them, so it worked blind with placeholder IDs and produced duplicates. (It also initially forgot that Standaardfunctie IS searchable, and only realized after I pushed.)
- It kept asking me for internal database IDs — equipment IDs, rate IDs — which users simply don’t have. It later admitted those IDs are visible only to it, not to us. Asking a human for a value only the database knows is a dead end; with searchItems it should resolve items by name/number itself.
- It couldn’t change an equipment’s pricing folder (FRA/ESP → USA), so a mispriced item stayed mispriced.
- It couldn’t create a quote or invoice at all — the workflow dead-ended at the final step.
- The session expired mid-operation more than once (“Your session has expired, please log in again”), and a refresh throws away the staged unsaved changes — so work is silently lost and has to be redone.
TWO THINGS WORTH SEPARATING
-
The execution/safety model is genuinely good. The agent stages changes in the form, an “Unsaved AI changes” banner appears, and nothing persists until a human clicks Save.
-
The model executes but doesn’t reason, and it has no way to learn how we work. It forgets its own tools, re-asks the same question, and can’t hold our conventions across a conversation. Combined with the lack of standing context/instructions per workspace, that makes it unusable for real work beyond single commands. This isn’t a criticism you need to fix in the chat — see the request.
THE ATTACHMENT
The most compelling part isn’t my write-up — it’s your own assistant’s. At the end of the session I asked it what it would need to complete a project A to Z. It produced a clean, structured gap analysis: its full tool list (23 tools, incl. two Zendesk ones) and the exact 8 tools it says it’s missing (deleteItem, createQuote, createInvoice, updateEquipmentFolder, searchItems for Functie and Leverancier, createSerialNumber, planEquipmentByPeriod).
THE REQUEST
Expose these write tools through the MCP. That single move solves both problems above at the root: you don’t have to build context management or a reasoning model into AI Chat — our agent already reasons and already carries our conventions, documentation and history. Give it the same tools your chat has, and the intelligence problem becomes ours, not yours. Your chat and our agent would share one write layer.
And one point I want to be very explicit about, because it follows directly from your own design: if the MCP write tools keep the same staging model as AI Chat (and they should), then the API also needs a SAVE / COMMIT command. An external agent can stage a change but cannot click your Save button. The full loop we need is: agent stages → human reviews → save is triggered, either by an explicit MCP-level save call after the user confirms in their own client (MCP clients already have per-tool-call human approval built in, so your confirm-before-apply guarantee is preserved), or by a review link a human opens in Rentman. A durable, programmatic save would also fix the lost-work problem above: staged changes wouldn’t depend on one browser tab staying alive. Without it, staged write tools leave every workflow stranded one click short of done.
Happy to be a test partner on this, as we are for AI Chat, and to walk you through the session or any of our workflows if that helps prioritize. This is the single highest-impact thing you could ship for us.
