One person was downloading multiple sales pipelines from Salesforce every week, combining them by hand, calculating what changed, then highlighting cells so the rest of us could see the movement. That is a real job in a lot of companies. It is also a poor use of a capable person’s week.
We connected Claude to Salesforce and taught a reusable skill from the report we already wanted. It pulled the data, assembled the weekly update, showed what moved, and reproduced the highlighting previously done by hand. It took roughly one or two rounds of human correction to get the output into shape. The manager still owned the pipeline review. The staffer no longer had to be a human integration between exports and a spreadsheet.
That is what an enterprise AI rollout should look like: the work changes, not just the software budget.
Start With the Operating Model
Taylor McKinley and Chris Cassidy of Mainsail Partners described an approach on the PE Value Creation podcast that deserves attention beyond private equity. Mainsail embeds operating specialists alongside its investment team and portfolio companies because smaller software businesses need the capacity to execute, not another recommendation in a slide deck. AI capability belongs close to the people and processes that create value.
Mainsail’s 2026 Talent Summit offered a different window into that model. People leaders built AI agents around candidate assignments, weekly KPI and deck preparation, and customer-call insights. These examples matter because none begins with a mandate that everyone become a software developer. They begin with ordinary work that the team already understands.
I would take that one step further. The operating team can supply the technical rails, but the person closest to the task has to help define the output. The expert in a weekly pipeline meeting knows which changes matter and what a useful report looks like. A programmer starting from a blank page does not.
Two Workflows, Two Different Stakes
The sales-pipeline example was about time and consistency. We had an existing output, an existing owner, and data in a system Claude could access once IT enabled the connection. We showed the model the format, corrected it, and made the result repeatable. There was no need to commission a custom application before removing the manual work.
We used a similar pattern in a very different workflow. Bank-account-change documentation sometimes entered a fraud-control queue that required human review. We created a Claude skill that let a team member upload a document and receive an assessment: what looked authentic or questionable, why, and which risk indicators deserved attention. The reviewer got a reasoned first pass instead of a blank screen.
We did not hand fraud decisions to Claude. Document review was one element of a broader control process, and the team retained human judgment and other checks. Automating an element of review is not surrendering the decision.
The Rollout Is a Management Job
I would organize adoption around five moves.
First, find a task that repeats and has an owner. Look for the person downloading files, reconciling fields, rewriting the same summary, or inspecting the same class of documents. Start with the actual inputs and finished output, including the exceptions. A generic request to use AI more does not tell anyone what to change on Monday morning.
Second, give the manager ownership of the workflow. In both examples, managers knew what a good result looked like and remained responsible for it. IT’s contribution was sometimes essential but narrow: provide access so the tool could reach the data. If the system is connected and the work is already being done manually, staff should be able to prototype a better way without waiting for a full software project. Work that actually requires engineering belongs with engineers.
Third, teach from the output. We did not begin the pipeline skill with a technical specification. We showed it the reports we wanted and reinforced the behavior when the first pass missed something. The fraud-review skill likewise had to give a rationale and flag concerns a reviewer could act on. People who know the work can teach a tool in plain language, provided they can inspect and correct its result.
Fourth, train for actual use. A demonstration at an all-hands meeting is not adoption. We used videos, examples, webinars, written how-tos, and direct training so regular staff could run the workflow without coding skills. Different people need to see the task performed, try it themselves, and know what to do when the answer is wrong.
Fifth, measure the work, not the licenses. How long did the weekly update take before and after? How much correction does the report still require? Are staff using the skill next month? In a higher-stakes process, does the AI assessment help reviewers surface the right questions while the control owner retains authority? My earlier piece on AI budgets explains why the baseline belongs at the task level, while the case for an embedded AI operating team shows how a central capability can support multiple operating teams.
Give Staff Room to Build
Some executives draw the line too early. They let employees use a chatbot to summarize a document, then require a technology project the moment someone wants it to touch a real workflow. That leaves the most valuable work on the other side of a ticket queue.
My line is different. With good people, good processes, and appropriate access, the staff already doing a manual task should be allowed to improve it with AI. Keep the accountable manager in charge, retain the human decision where the stakes require it, and involve engineers when the work actually needs engineering. Give the team the tools and the training, then let them work.
That is how AI becomes an operating capability instead of a collection of pilots.
If you run a portfolio company or an enterprise team and your AI program is still measured by seats purchased, start by showing me the weekly report somebody assembles by hand. The next step is probably sitting in that spreadsheet.
Sources: PE Value Creation podcast conversation with Taylor McKinley and Chris Cassidy; Mainsail Partners 2026 Talent Summit; Mainsail AI Labs.