The Executive Agent Lab · On-site, Metro Detroit · Q4 dates

Most corporate AI training is a slideshow about prompts. In this one, every department builds something that runs.

One afternoon on-site. Four departments, each sending a leader and the person who does the work. Every pair leaves with a working agent on a process they own, tested in front of you on work it has never seen, and a 30-day pilot with a named owner. Paid from training budget, delivered before the books close.

$24,500 for the leadership team Runs in your environment, not mine Training budget, not consulting budget

You cannot hire for this. So grow it.

Nearly half of North American manufacturers are still at the awareness stage on AI. Not because the technology is hard. Because the person who would own it does not exist on the org chart, and the req to create them has been open all year. The fix is not one hire. It is four department heads who have each watched an agent do their own department's work, correctly, on a Tuesday afternoon.

This is for you if

  • You lead operations, finance, IT, continuous improvement, or sales support at a $50M to $500M company, and "figure out AI" has been on your list all year
  • You can name four departments where somebody keys, matches, drafts, or routes the same kind of thing every day: invoices, quotes, supplier email, orders, scheduling requests, claims
  • Four department leaders will sit next to their best operator for one afternoon and build, not watch
  • You have training or L&D budget that expires December 31

This is not for you if

  • You want a keynote for two hundred people. This is eight seats and everyone works
  • Your team cannot get access to a business-tier AI account (Claude, ChatGPT, or Copilot) before the Lab. That is a prerequisite, and I will help you sort it
  • You want one big cross-department agent first. Those die in the room because nobody owns them. Start with four that somebody does
  • You want someone else to build it. That is a different engagement, and I offer that one too

What you get, exactly

Two weeks of sourcing, four hours in the room, thirty days of pilot support, and nobody leaves with a certificate instead of a result.

The engagement
The Executive Agent Lab
$24,500 up to 8 people, four pairs · Net 30
FormatTwo weeks of sourcing and prep, one four-hour lab on-site, 30 days of weekly office hours and a pilot review
CohortFour leader-and-operator pairs from four departments
EnvironmentYour AI accounts, your data, your machines. Nothing leaves the building
ApprovalTraining budget, one signature
What the leadership team walks out with
  • 1Four working agents, one per department, each tested on work it had never seenInvoice coding, quote assembly, supplier email triage, order entry, scheduling requests, report prep. Each pair builds one on three real examples, then runs it on two more it was never shown and scores the result against what the operator did by hand. That comparison is the whole Lab. Everything before it is setup.
  • 2A delegation brief and a governance card for each oneThe brief is what the leader wrote: what it does, when it runs, what it never touches. The card is the decision: runs on its own, runs and waits for approval, or stays on a human's desk. Plus a named owner, a start date, a review date, and a stop condition. It is how a workflow moves up the trust ladder one step at a time, and it is what makes your CFO comfortable saying yes.
  • 3The two layers everyone skips, startedA context layer (what your business knows about itself: pricing logic, service rules, brand voice, past decisions) and the first procedures written the way an agent can execute them. I draft the context file from discovery so the room does not spend hour one on setup. The model is a commodity you rent. These layers are what make it your AI, and they compound with every agent built after I leave.
  • 4An opportunity inventory across four departmentsBefore the Lab, each pair runs a structured sourcing interview on their own department and brings back a ranked shortlist: process, trigger, volume, minutes per instance, what goes wrong. Stacked, the four shortlists are the first honest map of where the manual hours go. It is the raw material for next year's roadmap, and most companies your size have never had it.
  • 5People who can build the next one without meThe point of the thirty days. Every week the pairs bring what they built and what broke. At day 30 we review the four pilots against their own stop conditions. By then they are not asking how, they are asking which.

Two weeks before, four hours in the room, thirty days after

The afternoon only works because of what happens on either side of it. That is what the fee covers.

Weeks 1 to 2 · Source

Find the four processes

A 45-minute sponsor call to pick the departments. Each pair then runs a sourcing prompt I give them inside your own AI account. It interviews the operator, applies the screening rules (document in, draft out, no logins, repeats weekly, five real examples exist), and returns a ranked shortlist. Discovery confirms the pick, redacts the examples, and gets accounts sorted. I write the context file.

Lab day · Build

Four hours, four agents

Every pair runs the same seven blocks on the clock, each on their own department's process. The operator types. I never touch their keyboard. By the demos, four agents exist, each has been tested on work it never saw, and each has an approval step written in front of it by the leader who will own it.

Days 1 to 30 · Pilot

Run it for real, with a stop condition

One hour a week, remote, whole cohort. Bring what ran and what broke. I review, unblock, and push. Day 30 is the pilot review: each pair reports against the number they said it would move. Keep, fix, or stop. All three are fine answers.

The four hours, block by block

Same clock for every pair. Different process for each.
  1. 0:00FrameWhat an agent is and is not. The trust ladder: drafts, drafts and waits, acts. The one rule: nothing leaves this room without an approval step.
  2. 0:20BriefThe leader writes the delegation brief on a one-page template. The operator edits it until it matches how the work actually happens.
  3. 0:45Build v1The operator builds it in your AI tool (a Claude project, a custom GPT, or a Copilot agent) on examples 1 to 3 and the context file. No code. One approved source: a folder of documents, not an API.
  4. 1:45Test on held-out workRun it on examples 4 and 5, which it has never seen. Compare to what the operator did by hand. Score it. This is the moment it stops being a demo.
  5. 2:15Break it on purposeFeed it the ugly example, the missing field, the angry customer. Fix the brief, not the model. Put the approval gate where it failed.
  6. 2:45Governance cardRuns alone, runs and waits, or stays human. Named owner, start date, review date, stop condition. Written by the leader, for their workflow.
  7. 3:10DemosEach pair shows the other three. Eight minutes each. The leader presents, the operator drives. Finance watching sales support's agent assemble a quote is what turns four tricks into a company capability.
  8. 3:50CloseFour pilot commitments on the wall. Office-hour dates set. Nobody leaves with notes.

Why this books this year, when most AI spend does not

Year-end AI projects usually fail for accounting reasons, not technology reasons. This one is shaped to get past them.

L&D

It comes out of the training budget. Which is the most classically use-it-or-lose-it line on a corporate P&L, and it is not the external consulting line that 40% of mid-market companies say they are cutting to fund AI. Different pocket, different conversation, usually a different approver.

1 afternoon

The expense books when the work is delivered, not when it is signed or paid. A Lab in November is delivered in November. Four department heads can clear one afternoon; nobody can clear two days. Net 30 means you pay it in December or January while it lands in this year's budget.

Zero

Logins for me. Everything runs in your environment on your team's accounts. I never touch your systems, which means there is no security review of a vendor, and that is the step that quietly kills most year-end deals.

✓

If a pair does not leave the Lab with an agent that handled work it had never seen, the fee is refunded.

Not a demo agent. Not a sample. One of your workflows, with your data, doing the work on examples it was not built on. I have never had to honor this, because the sourcing weeks are where the bad picks get killed. The thirty days after are the part that takes discipline, and that is why they are included.

Twenty years inside Fortune 500 technology programs. I have sat through the training that produces nothing. This is the other kind.

I'm Matt Sitek. Before Rivet I spent two decades in enterprise technology consulting, and a good share of it in rooms where people were being taught about a tool instead of building with one. Rivet is where I build automations for companies in Metro Detroit, and the Lab is how I teach leadership teams to do it themselves.

The approach is simple and it is the same one I use on every engagement: rent the commodity layers, build the proprietary ones. The model and the tools are the same for everyone. Your business's context and your way of working are not, and once your team learns to encode those, every agent they build after is faster than the last.

I'm local on purpose. On-site means I am in your building, with your people, and you can call me the week after.

Six questions your CIO will ask

Who should be in the room?
Four pairs. Each pair is a department leader and the operator who actually does the work: finance with the AP clerk, sales support with the person who assembles quotes, customer service with whoever answers the inbox, operations with the scheduler. The leader writes the brief and owns the pilot. The operator builds and knows what "done right" looks like. IT attends as an enabler, not a builder: accounts, data policy, the "does this train on our data" answer. Eight is the maximum because past that, nobody gets unstuck fast enough.
What do we need before the Lab?
Business-tier AI accounts for each participant (Claude Team or Enterprise, ChatGPT Team or Enterprise, or Microsoft Copilot all work). A conference room with a screen. Each pair runs the sourcing prompt I send them, which takes twenty to thirty minutes, and sends back the shortlist plus five real examples of their top pick with anything sensitive redacted. That is the whole list, and I will walk your IT lead through the account setup on a 20-minute call.
What does "tested on work it has never seen" mean?
Each pair brings five real examples. The agent is built using only the first three. Examples four and five are held back. Then the pair runs the finished agent on those two and compares its output to what the operator actually did with them by hand. An agent that only looks good on the examples it was tuned on falls over Monday morning. Holding two back is how you find that out in the room instead of in production.
Is our data safe?
Everything runs on your accounts, on your machines, inside your existing data agreements with your AI provider. I never take a copy of anything and I never log into your systems. Business-tier AI accounts do not train on your data by default; I will show your IT lead where that setting lives so they can confirm it themselves.
What are the payment terms and what paperwork is involved?
Fixed fee for the cohort, invoiced after the Lab, Net 30. W-9 and certificate of insurance ($2M general liability, $2M professional liability) come attached to the proposal. Two-page agreement, or your paper if that is faster. Most companies process this as a training vendor, which is usually the lightest onboarding path you have.
What happens after the thirty days?
Most teams keep going on their own, which is the goal. Some keep a monthly office hour running. Some take the four shortlists and the four pilot results and ask me to turn them into a sequenced roadmap for next year, with a payback number on each item. Some hit a workflow that is bigger than the team can build in-house and bring me in to build that one. All of those are fine, and none of them are required. You will have people who know how to do this, and that does not expire.

Q4 dates are open. Thirty minutes to see if the room fits.

Tell me the four departments and what each one is stuck doing by hand. I will tell you whether an afternoon is enough, and I will tell you honestly if it is not.