What changes when people start building and using AI at work? These accounts follow the work, the people doing it and the obstacles they encounter.
Operator experiments, interview retellings and published interview summaries offer different kinds of evidence. Each entry names its source and separates reported experience from suggestions and questions still to resolve.
Sources reviewed 31 October 2026. These are external accounts, not Nuvepro customer stories.
Changed work, people and adoption friction
Interview retelling
A finance leader builds tools across four parts of the business
Unnamed Series B technology startup
Alex Lieberman recounts a finance leader's show-and-tell. The tools are reported by the interviewee; the company, artifacts and results were not independently inspected.
The work described
The account starts with an investor operating model and extends into customer success and compensation. The finance leader supplies business context and iterates the tools rather than handing the whole job to a single prompt.
Previous fundraising material becomes context for clarifying questions, an ARR waterfall and an expense budget.
Customer-reported revenue and contract cost feed customer dashboards; an internal version alerts customer success when a threshold is crossed.
Dashboard figures and charts become editable QBR slides in the company's brand.
Claude Code-generated Apps Script macros connect compensation scenarios to Google Sheets, Rippling data, manager-specific folders, slides and out-of-band flags.
People and responsibilities
The finance head supplies context and revises the model. Customer success receives alerts, and managers receive team-specific compensation documents. The account does not describe who approves final compensation decisions or maintains each tool after the first build.
The adoption work behind the tools
Investor-model preparation and recurring customer questions create the demand. Building requires clarifying questions and iteration. The account leaves testing, change ownership and the quality of the connected records open for further examination.
What the source establishes
Lieberman reports that the finance head completed the investor model in two hours, compared with one to two weeks without AI. That is an interviewee's comparison, with no inspected artifact or comparable baseline supplied.
The suggested ten-day to one-day month-end close is the interviewee's belief, not an achieved result. The story supports examining the four described tools; it does not establish their accuracy, access controls or independently measured business impact.
A question for your own work
Who owns each tool after the first build, and which checks protect financial accuracy, customer identity and manager access when the context changes?
Operator-reported experiment
Box starts with document work and engineering experiments
Box
Aaron Levie's January 2025 account explicitly describes early experiments inside his own company. It is not a report of Box's present-day deployment coverage.
The work described
Box teams begin with work that already has documents, requests and reviewable outputs. Internal agents work on RFP responses and contract data; engineering teams test external agents on code changes.
Internal agents on Box are being built to prepare RFP responses and extract important contract data for downstream work.
External engineering agents are being tested on bug fixes and SDK updates driven by feature requests.
People and responsibilities
Internal teams build and test the agents. Levie anticipates people breaking work into tasks, checking quality and applying domain judgment. Those expectations are not evidence that a specific review control was implemented in these experiments.
What still needs to improve
The author identifies limits in tool use, reasoning, cost and speed. Browser interaction and agents working with other agents are described as early capabilities. Adoption is presented as experimentation with these constraints, rather than a completed switch to autonomous work.
What the source establishes
The post establishes a named operator's report of concrete experiments. It supplies no measured productivity result, evaluation suite, merged fix or production-acceptance evidence.
Future examples involving insurance approvals or IPO material are scenarios, not demonstrated Box deployments. The January 2025 publication date matters when interpreting what the account establishes.
A question for your own work
What evidence and human sign-off would be needed before a useful experiment gains authority over a consequential action?
Adoption-barrier interview
A retailer encounters a central approval bottleneck
Unnamed large retailer
Alex Lieberman reports an AI enablement conversation about adoption friction. The post does not establish a successful repair or a measured operating outcome.
The work described
The reported arrangement puts AI initiatives through a central Center of Excellence owned by IT. The issue is how employees can develop and release useful workflows when IT must own and approve each initiative.
Employees outside IT reportedly lack the tooling and infrastructure to build beyond simple workflows.
The account describes an intake and approval bottleneck. No successful change to that arrangement is established.
People and responsibilities
IT owns approval in the reported model, while business employees seek more capacity to build. Lieberman recognizes IT's risk incentives while questioning the pace of centralized delivery; that interpretation remains the author's analysis.
The organizational constraint
The reported volume of work that can move forward is constrained by IT's capacity to own and approve it. The post raises ownership and tooling as adoption problems without demonstrating an alternative release process.
What the source establishes
This is an attributed account of a retail adoption barrier, with a defined ownership problem. Neither the retailer nor the enablement leader is named, and no initiative counts or approval times are supplied.
The story does not show that decentralization solved the problem or justify bypassing IT review. The possible division of responsibilities still needs to be designed and tested in the enterprise concerned.
A question for your own work
Which decisions can business owners make, and which shared tools, review evidence and release controls would let IT support more work safely?
Published interview summary
An interview summary connects adoption to verification
Aaron Levie's enterprise AI perspective, summarized by Mike Fishbein
A published summary of an interview or talk, not evidence of a new Box implementation. The full interview was not reviewed for this entry.
The work described
Fishbein's summary asks why adoption progresses unevenly across functions. It contrasts engineering outputs that can be tested with subjective or interpersonal work where acceptance is harder to reduce to a simple check.
The summary proposes verification capability as one factor in where AI work advances. It does not document a particular operating change.
Synthetic users and better initial drafts appear as possible directions, not validated solutions or deployed verification systems.
People and responsibilities
The summary retains more human involvement for subjective work while methods and output quality improve. A qualified reviewer still has to decide what an acceptable result means for the actual business decision.
The verification bottleneck
Some work has a testable digital output; other work includes interpretation or interpersonal negotiation. The summary's contrast is useful as a question about acceptance criteria, not a universal rule that non-engineering work cannot be checked.
What the source establishes
The X post and its thread establish an attributed interview summary and identify a full interview or talk. They provide no adoption rate, measured business outcome or working synthetic-user evaluator.
Only the published X summary was reviewed. The linked full interview may provide additional context.
A question for your own work
Which decisions have explicit evidence and acceptance checks, and where does a qualified human still need to judge the result?
