Most people picture AI at work as a chatbot. Something you open, ask a question, and close again.
Julia Nash, AI product manager at Impact, has spent the past year helping teams move past that picture. What she's found is that the more interesting shift isn't a new task Claude can do.
It's where Claude shows up while you're doing the work you were already doing.
"It used to be, okay, I'm doing all my work, now let me take it into this AI space," Nash said. "Whereas now, AI can touch almost every space that you're working in."
The distinction between Claude as a destination and Claude as a layer across existing tools shapes almost everything else Nash had to say about bringing AI into a company at scale.
Claude Working Where You Already Work
Nash described Claude's reach in terms of two tiers. There's the wider ecosystem an entire department builds and shares, and there's the personal workspace each person configures around their own role.
Both depend on the same idea. Claude connects directly into the platforms people already use, rather than requiring them to manually feed it information.
Nash pointed to Databox, the platform her former marketing team used to track performance across every client account, as one example. Instead of exporting data and uploading it, Claude pulls directly from the source.
"Across all of our platforms, if we have live data or live information of any kind coming into a space, Claude can go in and pull that information at any point," Nash said. It can push updates back into those same spaces too, depending on the task.
Projects Give Claude Long-Term Memory
The first lever Nash pointed to is projects, which she described as a contained space for ongoing work on a specific topic. A project can hold multiple related chats, uploaded documents, and, notably, a live connection to files on your own computer.
The file connection changes how document work actually happens. Rather than accumulating a folder of version after version, Claude can update a file directly.
It will also flag when a change affects other related documents. That alone, Nash said, has cut down significantly on the version sprawl that used to pile up around long-running client accounts.
Projects are also shareable across a team. Colleagues working the same account from different angles can contribute their own chats and deliverables into the shared project.
Their work becomes context for everyone else, without a status meeting required to pass it along. Worth noting is that nothing is shared automatically beyond what someone chooses to add.
Skills Turn One-Off Prompts Into Repeatable Processes
The second lever is skills, which Nash described as detailed, reusable instructions for a multi-step task. A skill can run inside a project, which gives it access to all the client-specific context already built up there.
Her example was quarterly client reporting. A skill can pull performance data from the relevant platforms, run the specific analysis her team has standardized on, and apply the narrative approach her team prefers, all tailored to the client sitting inside that project's context.
Skills can also chain together. Nash described a multi-stage website build process at Impact, covering everything from audience research through wireframing, copy, and development, where each stage has its own skill.
Each one understands what came before it. It can flag a missing step, such as an SEO audit that hasn't been run yet but is needed before wireframing begins.
"It's really easy to move somebody who's new to the process into long-form workflows that have all this expertise and knowledge baked in," Nash said. What's left for the person running it is judgment.
They decide what to do with what Claude surfaces, rather than chasing down every step by hand.
Building a skill, Nash added, doesn't require a technical background. She recommends starting with Anthropic's own skill-creator skill, which asks a series of questions about the task and assembles the instructions from there.
Automation Without the Follow-Up Reminder
Beyond running a skill on demand, Nash described scheduling one to run on its own. Her example was a weekly reporting task that used to mean logging into every platform and assembling numbers by hand.
Instead, a scheduled task can run automatically at a set time each week. It pulls from every relevant platform, runs the analysis, and packages the result in whatever format is useful, whether that's a polished slide deck or a live, continuously updated dashboard someone can scan at a glance.
"That entire process, from scheduling it to pulling the data to the final presentation, would all live within just a single skill," Nash said.
The Security Question Isn't Really About AI
Nash was upfront that this isn't the conversation she has most often. Her work is mostly with employees integrating Claude into their day-to-day tasks, not with the IT and security leaders deciding whether to roll it out.
Even so, one question comes up more than any other. It's whether the vendor will use what's put into the tool to train its models.
The answer varies by provider; according to Nash, on enterprise plans especially, it's generally not happening. Her advice is to confirm that in the contract rather than assume it.
The bigger issue, in her view, has less to do with AI itself and more to do with the access already sitting inside a company's own systems. These tools work within the permissions people already have.
They don't grant anything new; they just make what's already accessible easier to find.
Anything overshared inside a company over the years can surface once an AI tool is layered on top. Nash sees that as a data hygiene problem that predates AI, not one AI created.
Her other piece of advice is to pay attention to what people are already using. If the approved path into an AI tool is slow to set up, employees will route around it with personal accounts, and free tiers typically come with weaker data terms than an enterprise plan.
The Real Barrier Isn't a Mistake, It's a Missing Habit
Asked about the most common misstep she sees, Nash pushed back on the framing. She doesn't see it as a mistake so much as a natural learning curve for a genuinely new kind of tool.
It's one capable enough, she said, that most people haven't yet built the habit of exploring what it can do for their specific role.
Setting up projects, connecting applications, and building skills takes upfront time that can feel hard to justify against a deadline. Nash argues the payoff is worth it.
A system built around how someone actually works keeps paying off long after the setup is done.
"I would just say, be unafraid," Nash said. "You're not going to mess anything up. Just try, no matter what."
What This Means for the People Doing the Work
Nash doesn't think the fear that AI will take someone's job is unreasonable. She also doesn't think it helps anyone to pretend work isn't going to change.
What she's noticed is that the people who feel least afraid tend to be the ones already doing something new with the tool. They're actively pointing it toward what excites them, whether that's work they couldn't take on before or time reinvested back into the parts of the job that are inherently human.
That's part of why Nash's approach centers on people building their own skills rather than being handed a fixed set of instructions. It puts the decision of what's worth handing to Claude in the hands of the person doing the work, not a policy written for everyone at once.
Nash is clear that a lot of work is changing and will keep changing. She's equally clear that she sees the shift as opening up new kinds of work within people's roles, not closing them off.
At a company like Impact specifically, she doesn't see the goal as reducing headcount. She sees it as scaling what the business can take on without losing the people doing the work.
Key Takeaways
- Claude's biggest shift isn't a new task it performs. It's showing up inside the tools people already use, rather than requiring work to be exported into a separate AI space.
- Projects give Claude a persistent, shareable memory for a specific topic or account, including a live connection to files on your own computer.
- Skills turn a multi-step task into a repeatable, reusable process, and they can be chained so one skill knows what came before it.
- Scheduled tasks let a recurring process, like weekly reporting, run and package itself automatically, without anyone needing to remember to start it.
- The real security question is less about AI and more about internal permissions. These tools surface whatever's already accessible; they don't grant new access.
- Most of the value gets left on the table not from a wrong move, but from never carving out the time upfront to set the system up around how you actually work.
- The people who feel least threatened by AI tend to be the ones actively shaping how they use it, not waiting to see what happens to their role.
Wrapping Up on Building an AI Ecosystem That Works
The picture Nash describes isn't a chatbot bolted onto the side of a workday. It's Claude built into the applications, documents, and processes already in motion, configured around the specific way a team or a person works.
By taking the time a person or a team needs to invest upfront is what elevates this into a system that elevates the work, the individual, and the team at large.
Hear more on the role of AI, specifically in the marketing and creative space, in Impact's webinar, Is AI Amplifying or Erasing Creative Exploration in Marketing?
