What I Learned From Six Months Of Using Agentic Assistants For Work

What Are Agentic Assistants, And Why Do They Matter?

Agentic assistants trace their origins directly to developments over the past couple of years in AI-assisted coding tools. As AI models improved, software development agents took on more complex tasks with less supervision, and some developers began using tools such as Cursor and Claude Code for work that had nothing to do with software. Until early 2026, one piece was missing: an agent that could bridge the everyday applications that knowledge workers live in — namely e-mail, productivity suites (Microsoft 365, Google Workspace) and the local file system. AI features existed inside those apps, but most of them were simple chat boxes that lacked the richness of developer-oriented tools.

For many non-coders, the first real example of an agentic assistant was Claude Cowork, which Anthropic launched as a research preview in January 2026. It embodied the agentic architecture behind Claude Code in a desktop app and connected to local files and applications to drive better context and results. Claude’s reputation for strong privacy and guardrails also made users more comfortable sharing sensitive data.

Cowork went viral and unlocked a new level of capability for power users who were not developers. The core benefit matched what developers had been enjoying: a real jump in productivity. It also drove a sharp increase in AI consumption and a clean upgrade path from free plans to $200-a-month Max-style tiers. Before long, similar tools appeared from other AI companies, including Perplexity Computer.

There is a catch. Claude Cowork, like an IDE or a command-line interface, is still a power-user tool. It rewards configuration and maintenance that most mainstream users are neither equipped nor inclined to take on. That gap is what pulled the enterprise software vendors into the market this year, each launching its own variant in a bid to become the primary AI interface for work. The prize is large. If every employee engaged an agentic assistant the way developers engage an agentic IDE, the economic benefits to the vendor (and, at least in theory, to customer enterprises) would be substantial. The result is a glut of products that, in my view, are still not well understood.

Are Enterprises Ready For Agentic Assistants?

I have been using Claude Cowork since it launched, and it has changed the way I work for the better. I have also been evaluating and testing some of the new enterprise-targeted products, which I cover in my next article. But the question for most organizations is not whether to deploy agentic assistants — sooner or later, they should. It is whether they have the human and technical foundation in place today to make them succeed. With that in mind, here are my five biggest lessons from six months of daily use.

1. It takes time to get dialed in. The most important thing I have learned is that my relationship to my own work has changed. Instead of always doing, I now more often act as an editor or a judge. The “assistant” analogy is apt: You provide goals and parameters, let the agent work, then collaborate and iterate with it toward the result you want. Even after half a year, sometimes it is still faster to do a task myself. The setup is also real work. I maintain an Obsidian vault as a personal context store (also known as a “second brain”), plus I have authored multiple Claude “skills” and I regularly connect to several custom MCP servers. If any of these terms or steps seem obscure to you, that just reinforces my point. My hope is that enterprise tools dramatically lower the bar for setup and operation.
2. Decide what you actually need the assistant to do. My biggest frustration is how vendors and influencers present these tools. Most demos revolve around small chores like organizing files or generating a daily dashboard — but those use cases are not worth the money. In my own work, the value is not that I save time; it is that the quality and depth of my output has increased. That leads me to what may be the single most important management insight for this technology. These tools have potentially enormous utility, so pick a starting point and define the change you want to see in the business up front, or else you can expect a degree of chaos.
3. Context is king. Appropriately applying context is the primary determinant of success with agentic assistants. Using memory features is a useful first step, but a well-organized personal context store and connections to the right applications and datasets are what separate a professional result from an embarrassing one. Hand someone an agent that is not wired to the right sources and the output will be poor, which sours both users and the team leads who support them. The stakes rise when these tools are used across a team or a company, where consistency matters. My colleague Melody Brue has written about the need to govern these agentic workspaces under a single operating model; at the platform level, this is where agent control planes such as Google’s Gemini Enterprise Agent Platform and Microsoft’s Work IQ come in.
4. Start thinking in pipelines. Because these tools descend from developer agents, another concept should carry over: the pipeline. In software, teams build automated processes so that improvement is continuous rather than a one-time event. Agentic assistants need the same. Consider a process for updating product collateral at launch, a process in which your product, marketing and legal teams all participate. An assistant can pick up the latest legal and marketing standards as content is produced. But what if legal needs to update a compliance statement or a trademark — what becomes of the older content? An agentic assistant can manage exactly that, provided that the pipeline is built with a bias toward continuous change rather than anchored to a point in time. My recent work on agents and ERP convinced me this is where many deployments will live or die.
5. One superstar will not carry a team. A pattern I’ve seen among early adopters: A single power user does not lift the whole group’s productivity, and can even create disruption. The developer who now produces ten times the code needs ten times the test coverage; if the testing team has not also automated its work, throughput barely budges. So far, we have mostly seen power users doing remarkable things on their own, but the bigger prize is making the entire team more productive.

So, What Should A Company Do With Agentic Assistants?

Most enterprises are not ready for full deployment of these assistants. This is not something simple like a new browser or even another SaaS tool. But do take action to deploy methodically — and start soon, because these agents take time to get used to. Three recommendations:

1. Start with a process, not a team. Choose a business process that crosses several teams. This will force you to understand cross-functional requirements and build the right pipelines and context sources. Not least, that will include getting leaders to agree on what those authoritative sources are.
2. Tie an agent platform into that first project. Agent platforms provide access controls, registries for agents and MCP servers, and observability and analytics functionality so you can monitor progress. My research on agentic control planes suggests these will become mandatory, so standing one up early keeps the right guardrails in place from the start.
3. Create a place for honest conversation. If you just hand people a pile of building blocks with a few skills and connectors, you will not get a good result. Power users love raw materials and will get rolling on their own, but most people need guidance on what to build and how. There is also genuine anxiety about AI in the workplace. Leaders do not need to have every answer, but in general people will get on board only if they feel supported.

I remain a strong proponent of this technology, and as an analyst I have done genuinely useful work using agentic assistants. But after countless demos and six months of hands-on use, my conclusion is simple: An agentic assistant is only as good as the human and technical foundation it is built on.

My next article digs into two enterprise agentic assistants that were launched this spring, how they compare and what I would still like to see incorporated into them and other competitors in this emerging space.