Replaced his team with AI — can't share it across the org

Interview metadata

Block 1 — Context

Consequential systems they named

• ADP Lyric (core human capital management system)
• ADP Health and Welfare Service Engine
• ADP Recruiting product
• Kronos UKG time system
• ADP Reporting
• DMS (dealer management system)
• Accounting software / CRM data
• Azure environment
• SQL database

Their language for “agent”

Will consistently avoided the term “agent” throughout most of the conversation, instead referring to “AI processes,” “skills,” “workflows,” and “LLM chat type stuff.” When directly asked about agents, he clarified: “agent to me is like what Connor was working on Andrew, which is like I’m trying to make an agent with copilot plus that like the accounting people can ask questions to and this agent is like customized with custom instructions.” He distinguished this from his current approach, saying he doesn’t use agents for “anything that’s important.”

Block 2 — The trigger event

What happened

Will has successfully deployed multiple AI workflows processing consequential business data, including weekly time audits for 1,800 hourly employees and automated incentive calculations for sales staff. However, he’s hit a scaling wall: “I’ve basically replaced my team with this AI process that I’m still the one managing and I’m still the one that’s having to basically get this information out to them… it almost makes it more complicated to try and share it with others because of how the ecosystem is kind of built within your own system.” The trigger was realizing that while he achieved 5–6x personal productivity, he created a single point of failure and couldn’t democratize the capabilities across his organization.

Named owner(s) of “what an agent is allowed to do”

Will Dellwo himself is the de facto owner, though he noted broader organizational concerns about AI licensing proliferation. He mentioned Andrew Walzer (likely a senior executive) and the IT team’s concerns about AI tools “like malware to a degree… duplicating across our organization.”

Security / compliance / risk conversation

Limited discussion at organizational level. Will mentioned IT team concerns about AI tools auto-inviting themselves to meetings and creating security risks, but no formal policy framework. His approach is self-imposed constraints: “I generally don’t use it for anything that’s important” and building extensive audit trails for everything he does deploy.

Block 3 — What they tried

Systems the agent touches (or would need to touch)

• ADP ecosystem (19 different connected products via APIs)
• UKG time system for punch data
• CSV exports from multiple systems
• Azure environment for data storage
• Shared folders and Git repositories for skill sharing

Vendors / tools evaluated and rejected

• Claude / Claude Code (currently using but frustrated with sharing limitations)
• Circle (mentioned as problematic, “inviting itself to like all sorts of things”)
• Various dashboarding companies using Snowflake
• ChatGPT (uses for non-critical tasks only)

Customer-operated vs SaaS preference

Strong preference for customer-operated: “I actually want to export, you know, 185,000 records, not 45,000 records because I don’t want to worry about the system trying to filter what the most recent record is and then screwing something up.” Wants full control over data processing and doesn’t trust external filtering.

Credential flow / approval surface

Currently manages all credentials personally. Expressed need for “gated approvals” and “action level audit” — wants system to show exactly what will be done before execution, especially for customer-facing or irreversible actions.

Their own connector / integration layer

Has built custom processes using Claude to ingest CSV files and connect data across ADP’s 19 sub-domains using employee ID as the key. Frustrated with the technical complexity of sharing these integrations with non-technical team members.

Block 4 — The pitch test

Reaction shape

Strong positive reaction with immediate recognition: “I agree” when presented with the concept of moving away from agent-centric thinking. Asked clarifying questions about how flight plans differ from markdown files and showed particular interest in the atomic, containerized approach. Engaged deeply with technical details about tool definitions and portability.

Of the four dimensions, ranked

1. Shared skills across teams — primary pain point: “how can we make sure that it’s being used the same way by everybody and like not drifting over time”
2. Action-level audit — critical requirement: “every time that we initiate a change to something I want you to generate a secondary markup file”
3. Gated approvals — important for irreversible actions: “do you want the gated approval before it pings the people about that?”
4. Deterministic execution — baseline expectation: already building this into his current workflows

Their one-sentence pitch to their VP

Could not compose one — framing not yet sticky. However, he articulated the value as: “We need like a tool that is deterministic, repeatable, auditable with approvals that like other people can engage with.”

Block 5 — Budget & authority

Named buyer

Andrew Walzer (name given) — appears to be senior executive reviewing AI licensing across organization. Will noted “Andrew Walzer or Pete Swenson” as decision makers for broader rollouts.

Proposed next step

Did not propose a specific next step, but expressed strong interest in continued conversation: “I’m very interested because… what you’re describing is what I was trying to like build out a proof of concept for.”

Budget signal

Indirect signals around current AI licensing costs being scrutinized. Mentioned concern about Claude licenses potentially costing “385 dollars a month instead of what it currently costs” and organization tracking “285 people” with various AI licenses. Suggested current spend is significant enough to warrant executive review.

Verbatim quotes

“I’ve basically replaced my team with this AI process that I’m still the one managing and I’m still the one that’s having to basically get this information out to them.” — Block 2

“It feels like there’s something missing from the way that it’s getting approached that doesn’t feel like it’s reaching down to the level of what would actually make a lot of companies more effective.” — Block 2

“I think we need a better tool that’s not claude code to like try and build certain things… it’s a technical mess for people who aren’t like they’re not developers.” — Block 3

“I think the value of AI right now is that people don’t need to be repeatable and that they’re totally fine to like spin something up, use it one or two times and then kick it down the road.” — Block 5

“The answer is if those 80 people actually use it to 30% efficiency, the answer is it’s probably still worth it. But the hard part is like you’re creating more effectiveness within the company, but not in a scalable way.” — Block 5

Action items

• Follow-up conversation warranted — Will is a strong design partner candidate with sophisticated understanding of the problem space
• Research Walser Automotive’s broader AI initiative and decision-making structure
• Explore specific use cases around payroll / incentive calculations as potential pilot scenarios
• Connect with Andrew Gordon to understand the broader organizational context and other potential stakeholders