AI Training for Employees and Commercial Teams: Delegate, Judge the Output, Build Your First Workflow

Corporate AI training in our practice teaches four competencies in a fixed order: delegating a defined task to AI, describing the output you want, judging the result against the deciding test in your market, and staying diligent once the first workflow runs. Each is practiced on the team's own documents, never on sample material.

Delegating comes first because most teams have never written down what a task is. An analyst who watches a public filings feed knows the job by feel: pull the notices, skip the ones that never apply, read the rest, flag the ones that touch something on our watch list. The first exercise is to write that down as a task with an input, an output, and a rule for what gets flagged. Describing follows: the output format, the fields, the tone, the length, and what to do when the source is ambiguous. Teams are surprised how much of the work was never specified, and how quickly the model performs once it is.

Judging the output is the competency that separates a team that can use AI from a team that can be blamed for it. A model produces fluent, confident text whether the answer is right or wrong, and the only defense is a person who knows what a wrong answer looks like in this market. That knowledge is domain-specific: a coordinator knows which names on the watch list matter, a proposal manager knows which clause a reviewer will strike. Knowing what a wrong answer looks like in your market is the scarce skill, and it cannot be bought; it is learned on the desk. Our exercises put a wrong answer in front of the team and ask them to find it.

Staying diligent is the competency most programs never teach, because it only matters after the first win. Once a workflow runs, the temptation is to stop reading its output. We teach the team to write the check down, name the owner, log every run, and review the incomplete executions on a fixed cadence. Every agent stops for a person, and that person has to keep showing up. The competency checklist that closes the course makes this concrete: a competent user on this desk can delegate a task, describe the format, find the wrong answer, and switch the workflow off. Managers score the team against it before and after.

The three prompts, three answers lesson shows a team that the same facts and the same model produce three different answers under three prompts, and that the triage stays wrong until the deciding test comes first as a literal checklist. It is the first lesson in every format because it teaches judging and describing at once.

The lesson came out of our own build. Our regulatory watch agent watches a public filings API and ranks each notice HIGH or ROUTINE against the desk's own watch list. During the build, three versions of the ranking prompt were run against the same notices with the same model, and each version returned a different triage. The facts never changed. The model never changed. Only the prompt did, and every version read like a reasonable instruction to a colleague. The triage was wrong until the deciding test, does this notice touch something on our watch list, came first in the prompt as a literal checklist.

In the room, the team runs the same sequence on its own documents. An analyst brings a notice, a proposal manager brings an RFP clause, a marketing manager brings a review to answer. Each writes the prompt they would have written before the lesson, runs it, and reads the output aloud. Then the group names the deciding test for that document, moves it to the top of the prompt as a numbered checklist, and runs it again. The second output is closer, and the team can say why. That is the moment describing and judging become one skill instead of two.

The lesson goes first in every format for a practical reason. Executives who see it stop asking whether the model is smart enough and start asking whether the desk has written its deciding tests down. Analysts who see it stop collecting prompt patterns and start collecting checklists. And the sponsor gets a clear picture of the work ahead: every workflow the team will build needs its deciding test written before the build starts, which is exactly what the first-workflow lab asks for. The prompt is not the product. The written check is.

We deliver AI training for business teams in six formats: a 90-minute executive briefing, a half-day workshop, a full-day workshop, a week-long workshop, a six-week first-workflow cohort, and a train-the-owner session after any agent or workflow build. The format follows who is in the room and whether the mandate is awareness, judgment, or a running workflow.

The 90-minute executive briefing is for the CEO, the division head, the CFO, and the leaders who sponsor the work. No laptops required. It covers what an agent is, an automated workflow with one thinking step that stops for a person, the three prompts lesson, the three rules, and the five moves to the first workflow. Leaders leave able to ask the right questions of any build proposal: which platform, which data, who owns it, where does it stop. The half-day workshop is for a working team of six to twenty. It covers all four competencies with exercises on the team's own documents and ends with the competency checklist and three candidate workflows named. The full-day workshop is for the same working team when the mandate includes a build. The morning covers the four competencies and the three prompts lesson; the afternoon sketches each candidate workflow on the trigger, gather, think, act, stop template, names an owner for each, and writes the deciding test before anyone builds.

The week-long workshop is for a team that wants the whole sequence on site in one block: the four competencies, the three prompts lesson, the first-workflow lab on the sanctioned platform, and the owner handoff. Each small group leaves with a workflow that has run end to end, a named owner, and a written check. The six-week first-workflow cohort covers the same ground with working time between sessions, for the team that will own workflows. The opening sessions cover delegating and describing, the next runs the three prompts lesson on the team's documents, the middle sessions are the first-workflow lab with a build check at each one, and the closing session is the owner handoff, the written check, and the competency assessment. Each participant, not each group, leaves with a workflow that runs, their name on it, and the next one sketched. The cohort is the format we recommend when the mandate is a running workflow per person rather than awareness.

Train-the-owner is the shortest format and the one we never skip. It follows any agent or workflow our practice builds. The named owner sits with the builder for a working session and learns to read the scenario, edit the prompt, read the run log, review incomplete executions, and switch the workflow off. Nothing is handed off until the owner has done each of those once without assistance. It is the difference between a client-owned build and a build the client is afraid to touch, and it is included in every scope because a workflow without a trained owner is a workflow that stops being reviewed.

The first-workflow lab takes a team from a sanctioned platform to a running agent. Each participant builds a trigger, one thinking step, a defined action, and a stop point where a person reviews, on the low-code or no-code platform IT sanctions: Make.com, n8n, Zapier, or Microsoft Copilot Studio. If it needs a developer to start, it is the wrong first workflow.

The lab starts with platform sanction, because the first rule is sanction the platform. Before anyone builds, IT has named which of Make.com, n8n, Zapier, or Microsoft Copilot Studio the organization will run, on which account, with which connections approved. If your team already pays for Microsoft 365, Copilot Studio is often the shortest path. If the desk lives in Google Workspace and HubSpot, Make.com or Zapier connect directly. n8n suits teams that want the workflow hosted inside their own environment. We do not pick for you; we run the lab on what you sanctioned, and every workflow built in the room lives on the company account, never a personal one.

The second rule is keep confidential data inside the firewall. The lab uses public and approved sources only: a public filings API, an agency or government docket, a procurement or RFP portal, a shared inbox IT has cleared. Anything the data policy has not cleared stays out of the exercise, and the workflow logs what it read. The third rule is every agent stops for a person. Each workflow built in the lab ends with a review step: an email that asks for RELEASE or HOLD, a task in the CRM, a row in a sheet with an owner column. Nothing the workflow drafts moves downstream until a named person approves it.

The build itself follows the same shape every time. Trigger: a schedule, a webhook, a new email, a new row. Gather: the source call, the dedupe check against a data store. Think: one model step with the deciding test at the top of the prompt as a checklist. Act: draft, flag, route, log. Stop: the review request to the owner. Participants build it on their own source, run it against real data, and read the run log together. The lab closes when the workflow has run at least once end to end, the owner has switched it off and on, and the next workflow is sketched on the same template. That is what client-owned means.

AI literacy training in our practice builds on free courses rather than repeating them: Anthropic Academy AI Fluency, Google Prompting Essentials, Oracle Agentic AI Foundations 2026, Harvard CS50 AI, and MIT 6.S191, plus Make Academy and the n8n, Zapier, and Microsoft Copilot Studio learning libraries. Our training adds the team's own workflows on top.

The five courses are free and they cover the foundations well enough that no organization needs to pay to have them taught. Anthropic Academy AI Fluency teaches the delegation, description, discernment, and diligence frame that our four competencies are built on. Google Prompting Essentials is the shortest path to writing a usable prompt. Oracle Agentic AI Foundations 2026 explains how an agent is structured, which is the trigger, gather, think, act, stop shape from the lab. Harvard CS50 AI and MIT 6.S191 are for the analyst or engineer who wants the mechanics underneath, and are optional for a commercial desk. We sequence them by role in the free-course pathway deliverable.

The platform academies cover the build side. Make Academy walks through scenarios, modules, data stores, and error handling. The n8n, Zapier, and Microsoft Copilot Studio learning libraries do the same for their platforms. All of them are free, and all of them teach the tool on the platform's own sample data. That is the gap our training fills. A team can finish every course on the list and still not know which filings matter to its own watch list, which RFQ fields its estimators need, or which exception should page a person at 2 AM. The courses teach the tool. We teach the tool on your work, with your deciding tests, and a named owner.

We assign the pathway before the first session, not after. Cohort participants complete Anthropic Academy AI Fluency and Google Prompting Essentials before the first session, so the room starts at the exercises rather than the vocabulary. Owners in the first-workflow lab complete the relevant platform academy modules before the lab sessions. Leaders in the executive briefing get the pathway as a leave-behind with a note on which courses their teams should take first. The education is free and the first workflow needs no developer, so the remaining work is the team's own sources and judgment.