We are completely missing the point of the AI revolution.
HR leaders give me an anxious face after telling me which AI model their team is using. As if to ask, "am I using the right one?" Wrong question. MIT research found that 95% of corporate AI pilots produce no measurable return. The 5% that do share one trait: they're built around workflow integration, not around which model got picked. We spent this year proving that inside our own walls at Lever Talent, Inc., asking a more beneficial question: "When does a workflow actually deserve an agent, and when are we just building an expensive shortcut around a job a human should keep?"
AI has introduced what I call the AI Wedge. This is the shift where Agentic AI stops assisting and starts executing entire workflows on its own. That shift eliminates what we call Tactical Middleware, or the manual routing and admin work that used to eat a huge chunk of a skilled person's week. The Wedge is real. And if appropriately addressed, it can create massive strategic leverage. But dropping a model into a workflow without governing it doesn't create leverage. It creates risk with a chat interface.
So, earlier this year, we built something for ourselves before we built it for a single client. It's called the Agentic Workflow Vetting Worksheet. When a Lever Talent team wants to hand a workflow off to an agent, it first has to clear four checkpoints. We call them the 4Cs: Context, Constraints, Connections, Control.
Before getting into the 4Cs below, I wanted to share what our AI tech stack looks like at Lever Talent. It's always subject to change, but the latest lineup is two platforms. First, we use Anthropic Claude Cowork as the default application for individual team members to use. Second, we use Google's Gemini Enterprise Agent Platform to create and host our autonomous agent fleet.
All team members are empowered to use Claude to optimize their work. Most workflows are prototyped in Cowork, and then we use the 4Cs to vet their company-wide application. We arrived at this AI tech stack for many reasons, including optionality (we use ChatGPT, Claude, Gemini, and Grok depending on the use case), which I can share in a future post. Your AI tech stack probably looks different, and that's totally expected since every business situation requires something different. Now, onto our Agentic Workflow Vetting process using the 4Cs below.
Context is what internal strategy, culture, and standards the agent needs to do the work the way we would. That means the playbooks it has to read, the brand and formatting bar its output has to clear, strategic plans, and any hard decisions it can't contradict. I'll use our real example here of applying the 4Cs against our Client Success team's account management work. Before building a Client Success AI agent, client research always happened before every client conversation.
A Client Success Manager (CSM) pulled together account history, engagement, usage, and prior conversations. Naturally, this important work competed with a full book of accounts and everything else on a CSM's calendar, so the depth of research varied CSM to CSM and rarely went as deep as the account deserved. We developed proprietary skill files to give the agent the standard operating procedures, protocols, and job knowledge it needs to produce the most valuable output.
Constraints address the process the agent operates inside. It governs what part of workflows it touches, what parts it stays out of, and the hard rules it can never break. It's also an honest read on the data it's acting on, whether that data is clean, structured, and ready to use, or a mess that needs conditioning first. Mitchell Hashimoto, the founder of HashiCorp, calls the rule-building part of this area "engineering the harness." What this means is every time an agent makes a mistake or does something unexpected, you build a fix or implement a hard rule so it doesn't happen again, instead of just hoping for a better result next time. The more the harness materializes, the better the output.
For our Client Success AI agent, the harness is built around a proprietary client retention model we built in-house to score a client's level of investment and engagement. The agent's research gets filtered through that model before it ever reaches a CSM. The workflow stops short of assigning tasks and action items. Sending a message or triggering another workflow isn't just discouraged, it's blocked in the code itself. The only way that changes is if someone deliberately rewrites it to allow it.
Connections are every technical system, or application, the agent has to plug into, what it pulls data from, what it pushes output to, and how clean each of those connections actually is. To conduct in-depth client research, our Client Success agent pulls account history from our HubSpot CRM, reviews the latest recorded meeting transcripts in Fathom, and researches the company through news outlets and its corporate website. It also goes deep on key contacts via Tavily search (a search engine for AI agents) to surface the most recent, relevant news. It then compiles that research and uses the context to generate a specific client playbook. This content lands in a Google Document that is linked to a task in HubSpot.
With the agentic workflow activated, playbooks are automatically created when a client score hits a certain alarming or celebratory threshold. A CSM now starts each week with a queue of the most critical client interactions to focus on and a playbook for each, which surfaces where the relationship actually stands and the best path forward to deliver value to the client.
Control is who's accountable for what the agent produces, who's actually in the loop reviewing it, and what that review has to carry for the people it touches. That includes empathy for those whose jobs may change because the agent is now doing part of it, and a clear answer on what gets disclosed and to whom. Researchers at Hugging Face published findings this year warning that human oversight of AI agents quietly breaks down over time: the more people lean on an agent, the more their own judgment atrophies, until the human review step turns into a rubber stamp.
We built Control to fight that. A named CSM reviews everything the agent surfaces. We call this the human-in-the-loop. The manager of our client success team is accountable for the team's playbook usage. CSMs deliver regular feedback to increase the efficacy and impact of the agentic workflow. I am personally accountable for the agentic workflow running as planned. The agent never has a single conversation with a client. The agent does the research. A human still owns the relationship.
Every vetting worksheet filled out creates a forcing function to double- and triple-check that the decision to build and assign a workflow to an agent makes the best sense. An agentic workflow project cannot be considered unless a worksheet is 100% complete. Even then, the worksheet is submitted to our operations team for review and assessment against our existing fleet of agents, skills, and workflows, to determine the best path for development or whether it should be discussed further before adding it to our roadmap. And in cases where the business may not be ready to adopt it company-wide, individual team members are empowered to prototype the process using their own Claude Cowork instance.
Before implementing the Client Success agent project, we set an objective to "predict the churn risk of accounts and empower the Client Success Team to proactively address issues, deliver on client needs, and yield the highest client engagement possible." Conducting a thorough round of client research took up to thirty minutes per account, more once you factor in the context switching to get into it.
The implementation would be successful if it freed up the CSM to have more time with clients and increased the quality of client interactions. Across our roughly 400 client accounts, each needing that treatment at least once a year, that added up to about five weeks of someone's time annually, time a CSM was spending on research instead of clients.
We're already seeing results in our internal tracking. CSMs aren't spending their week rebuilding account history or considering the best course of action from scratch anymore. They're spending it in the conversations that actually decide whether a client renews or expands. The agent didn't replace the CSM's job. It took the Tactical Middleware off their plate so they could do the part of the job only a human can do.
This is the same discipline behind everything we're building at Lever Talent right now. We're becoming the Orchestrator of Record for the human and AI Workforce, the layer that decides the right mix of people, programs, and technology for a business and then deploys it. The 4Cs are how we make that call on ourselves before we make it for a client.
This is not about whether you can build an agent. With the latest development tools, anyone can build an agent. This is about whether the workflow will benefit from one and what you owe your team and your clients once it's running. The leaders who win over the next few years won't be the ones who bought the most AI. They'll be the Human Orchestrators who built the tightest 4Cs around it, starting with themselves.
Want to run this on one of your own workflows? Download the 4Cs Worksheet and try it yourself.
Want to run your whole team through it together? Register your team for our AI Fluency for HR masterclass.
Want help to build a strategic harness and vetting process for your company? Speak to a Workforce Orchestration Expert today!