Two questions come up in nearly every customer conversation I have right now: “Which AI tool should we buy?” and “Where does Rovo fit into our stack?” Both are fair questions. Both are also, I think, the wrong ones, and I would rather say that plainly than sell you a comfortable answer.
The Tool Was Never the Point
Think about what it takes to answer one ordinary question at work. A customer asks why a release slipped. You open Jira to find the tickets. You open Confluence to read the decision log. You scroll back through Slack for the thread where someone flagged the dependency. Then you check the spreadsheet somebody maintains on the side, because there is always a spreadsheet.
Nobody designed that. It accumulated. The average company now runs 118 SaaS applications, up from 106 a year earlier, according to BetterCloud. Count licenses under management rather than apps in daily use and Zylo’s 2026 index puts the figure at 305.
Here is the part we rarely say out loud: the user interface was always a workaround. Menus, tabs, filters and buttons exist because software could not understand what you actually wanted. You had to translate your intent into a sequence of clicks the machine could follow. We got so good at that translation that we stopped noticing we were doing it, started calling the translation layer “the tool,” and then went and bought a hundred of them.
Agentic AI removes the need for the translation.
Agentic AI Turns the Interface Into a Colleague
Look at how Atlassian describes the anatomy of a Rovo agent: identity, behavior, knowledge and skills, in their own documentation. Read that list again without the product framing. A name and a description. Goals, guardrails and tone. The data it is allowed to see. The actions it is allowed to take.
That is not a feature list. That is a job description. You are not configuring software, you are writing an onboarding document for a colleague who happens to work very fast and never sleeps.
With a tool, you supply the process and the tool supplies the capability. With an agent, you supply the outcome and the agent works out the process, including which of your hundred-odd applications it needs to touch to get there.
I Have Mostly Stopped Opening Jira
I did not arrive at this from theory. It happened to me in practice.
I now do a large share of my Jira and Confluence work through Claude rather than through Jira and Confluence. I ask for what I want the way I would ask a colleague: pull the open items for this initiative, tell me where the migration project actually stands, find the page where we documented that decision in spring and tell me what has changed since. It reads both systems, reasons across them, and answers.
What surprised me was not the time saved, though there is plenty of that. It was noticing how much of my previous effort had nothing to do with the work itself. Choosing the right board. Remembering the JQL. Knowing that the thing I needed lived in Confluence and not in Jira, which is a piece of trivia about our tooling, not a fact about our business.
I have not stopped using Jira and Confluence. They hold the data, the permissions and the history, and they are still the system of record. I stopped operating them by hand.
Our IT Team Gets Asked in Plain English Now
The better example is not mine. It comes from our own IT team, and it runs every day.
We have an internal bot our people call @helpme. It lives in a Google Chat room. You do not file a ticket, you do not pick a request type from a dropdown, and you do not go looking for which form covers your situation. You describe what you need the way you would describe it to a person: I need access to this system, my license has vanished, can someone reset this for me.
Behind it sits Taskmate, our platform for building agents that operate SaaS tools the way a human would. The agent reads the request, works out what is actually being asked, carries out the steps across whichever systems are involved, and escalates to a human when it hits a decision it should not make alone. Every run is logged, with screenshots, so IT can see exactly what happened.
The interesting part is what disappeared. Not the IT team, who now spend more of their time on the work that needs judgment. What disappeared is the service catalog as an interface: the taxonomy our colleagues had to learn in order to ask for help correctly. The requests were always in human language. We used to make people translate them into our forms first.
Agentic AI Only Works if Everything’s Connected
None of this runs on enthusiasm. An agent that cannot reach your systems is a chatbot with opinions.
The reason agentic AI turned practical over the last eighteen months is deeply unglamorous: a standard connector. The Model Context Protocol, released by Anthropic in late 2024 and now governed under the Linux Foundation with OpenAI, Google, Microsoft and AWS behind it, gives agents one consistent, permissioned, auditable way to reach enterprise systems, and roughly 28% of Fortune 500 companies have already implemented it. Taskmate takes the other road where it has to: its agents drive the browser interface directly, which is how you reach the systems in every company that never got a usable API.
For anyone buying software, this changes the evaluation question. It used to be “does this tool have AI features.” It is closer now to:
- Can an agent reach this system through a standard, permissioned interface, or only a human through a browser?
- Does that access respect the permissions we have already defined, or does it require a second set nobody maintains?
- When the agent acts, do we get a log we would be comfortable handing to an auditor?
A system that fails those questions is not neutral. It is a system your people will go on operating by hand while the rest of the stack gets delegated around it. That is a real and growing cost, and it belongs on the evaluation sheet.
I Am Not Telling You to Hand Over the Keys
Here is the part where I argue against my own enthusiasm, because you should be suspicious of a CEO who only brings good news.
The gap between interest and working deployments is wide. Only 23% of organizations have actually scaled an agentic system into production, and Gartner expects more than 40% of agentic AI projects to be scrapped by 2027, largely over unclear business value and inadequate risk controls. Those will not be model failures. They will be organizations that bought agentic AI as a product instead of deciding which jobs they wanted done.
The risk profile genuinely changes, too. I wrote recently that AI safety should be architecture, not a prompt, and agentic AI is where that bill comes due. A tool that reads your data is one thing. An agent with permission to act inside your systems is another. Our @helpme bot has real access to real systems, which is exactly why it has scoped permissions, human escalation, and a screenshot log of every action it takes. Those were designed in at the start, not bolted on after a bad afternoon.
So start narrow. Pick one job that is high volume, low ambiguity and easy to check. Give the agent the least access that job requires. Read the logs, and measure the output rather than the enthusiasm: my colleague Martin ran two agents against a production codebase for a full workday and wrote up what he actually measured, including the run where the test was green and the page was unusable. Expand when you have earned the confidence, not before.
But start from the job.
Stop shopping for tools. Start defining the jobs.
Our Advisors Start With the Work, Not the Stack
So here is what we actually do. Before anyone talks about licenses, we map the work: which jobs happen often enough to be worth delegating, which ones are safe to hand over first, and which ones genuinely need a person in the loop. Then we build for the job, with the access and the guardrails designed in from the beginning, whether that means a Rovo agent inside your Atlassian stack or something running on Taskmate across everything else.
We run this on ourselves before we recommend it to anyone. The @helpme bot is not a demo, it is how our own IT team works. And I still keep Jira open in a tab. I just do not click around in it much anymore.