You used to sign up to SaaS products to do a defined task, and that defined task used to stay defined. You learned the tool, the tool did the job, and the outcome was either done or it failed in a place you could point at.
A lot of the SaaS that owned those jobs have since grown an AI layer. Make, Monday, and plenty of others now sit a model on top of canvases that were already dense: modules, boards, branches, permissions. The pitch is convenience in managing the defined tools, but we find for a lot of people the result is more time, and more thinking, on work that used to be smaller.
What you should expect from AI
An AI layer on a product you already pay for should still behave like a tool.
- The job you signed up for stays defined
- The tool finishes that job, or it fails in a place you can point at
- Any path it suggests is one the product can actually run
- When it fails, the error names what failed
- You spend less time on the task, not more
- You are not left supervising another chatbot
Another decision layer
The reach for automation was meant to remove steps, but what many of these products shipped is a decision layer. Instead of learning the tools, you now describe the goal to their AI assistant and the model proposes a path. You are then left to judge whether that path is real, appropriate or good enough, just wothout ever learing the tool itself, so now you spend the time to make sure the path is good enough, and you are left to wonder if you have spent more time than you would have if you had just done the job yourself.
The old job was to configure a tool you could see undetsatnd the tool and terminoligy enough to get by, or employ someone who would already knows the tools. The new job is to understand why the assistant believes this will work, then discover whether it works out, with all the responsibility place on you, all the time tinking you are cutting out the expensive of delegating responsibity or learing the tool yourself.
Tools that were excellent at one specific task now ask you to supervise a generalist. When the one-shot misses, as it often does rght now, the miss is yours to debug and pay for. You spend the attention the product claimed it would save working out why the suggestion did not finish the job.
Unexpected behaviour
We see this when connecting Microsoft SharePoint to a Google Sheet in Make. The assistant recommended a route that dependeds on APIs the product does not support, and it failed. The errors that came back are generic, and so the remaining work is left to you to establish why a suggested path could not succeed.
A missing integration in a defined process is something you can plan around. A confident recommendation of an unsupported path, followed by an error that names neither, costs time on the already paid for extra AI steps. You are debugging the assistant as well as the workflow, and this is where a SaaS will wither rise or fall in the AI Agent market.
The AI wrapper
Putting AI in front of existing steps is a legitimate way to automate: Fill this field, pick this module, draft this status. It works when the step is real and grounded and not a on the fly recomendation from an LLM, the tool can perform it, and someone has tested the path the model will offer.
Great UX cannot skip that thinking and testing, or you inherit the Make mess. You added AI and it just became another surface to manage (a paid for service at that!).
The shortest path first
At Pendulum the AI we ship has to answer the problem in front of it. We start from the simplest path that still does the job and an agent earns its place when it removes a decision. More time spent choosing means the layer failed and this is fundametal to UX in a post AI world.
Thats how we build Arc. Agents work on the landscape you already have: measurements, repetitive passes, data driven workflows that solve probems we know you have already. The simplest route is the one taken first and the enhancement should leave you with less decision time and ultimalty a saving. Keep it simple applies to the user experience, not only the technical architecture. You should not simply get another chatbot you need to supervise.
A shorter path, results you can trust, and your attention on the outcome: that is what an AI-enhanced web promissed you, thats is waht we are building in Arc.