AI for Managers

A woman executive in a charcoal suit with a gold brooch stands at a large office printer and presses a button on its panel. A sheet with coloured bars and lines is coming out of the printer.

AI training for executives produces results when you work through a specific task step by step. In this example, a manager decides whether to change how a weekly consolidated report from three teams gets prepared. The deliverable is a comparative memo - five options compared against the same criteria. The data is composed for the exercise; no real tables appear.

What is the weekly report task behind this exercise?

Every Friday a coordinator collects status updates from three teams and prepares a consolidated report. The manager reviews the document before an internal meeting. On the surface it looks routine, but the coordinator names a specific problem: a lot of time goes into clarifying date periods and decoding statuses.

There is no precise time measurement. But the discrepancies are concrete:

  • One team's table sends statuses with no date attached.
  • In another, "done" means the internal work is finished, not that the client received the handoff.
  • In the third, the period is recorded as "week 41" - without a clear definition of what that covers.

Constraints for the task: the coordinator may prepare a draft from anonymized copies of the tables. Sending the report and changing systems are not allowed. There is no license pricing and no savings estimate.

The manager will evaluate results against four criteria:

  • accuracy of statuses and periods in the finished report;
  • preparation time including review - unknown for now;
  • additional ongoing costs;
  • ability to revert to the current process.

What four steps does a manager take - and how is the prompt structured?

The training program describes twelve sequential actions. For a working task, they collapse into four blocks - things the manager does, not the tool.

Step 1. Lock in the decision and the known facts.

Write down exactly what is being decided: change the process or leave it as is. Set a deadline and name who is responsible. Separate what you know for certain from what is still unknown.

In this task, what is known: three tables with discrepancies, the current report format, and the coordinator's role. What is unknown: exact preparation time, the cost of any change.

Step 2. Set the alternatives and criteria before opening the AI tool.

Write down all the reasonable options before you see any results. In this task there are five: keep as-is, introduce a unified template without AI, learn to use an approved tool, buy an off-the-shelf solution, and build a custom integration.

Comparison criteria are fixed before you look at the output. If you choose them after, they start matching the option you already liked.

Step 3. Test AI on one specific, narrow operation.

In the teaching example there are no full tables - only a task description, three discrepancies, and the constraints. The manager gives the model a task description and asks it to build a comparative memo across five options. Then the manager checks every conclusion against the original conditions.

The goal of this step is to test one narrow action: can a task description be turned into a structured memo without losing any meaning?

Here is the prompt you can copy into any approved tool:

"Here is a task description. A coordinator collects status updates from three teams every Friday and prepares a weekly consolidated report for an internal meeting. The manager reviews the document.

Three discrepancies in the source data: in Table A, statuses arrive without dates; in Table B, 'done' means internal work is finished, not client handoff; in Table C, the period is labeled 'week 41.'

No measurement of current preparation time exists. No license prices or savings estimates exist. Sending the report and changing systems are not allowed.

Prepare a comparative memo on five ways to change the process: keep as-is, introduce a single unified template without AI, learn to use an approved tool, buy an off-the-shelf solution, build a custom integration.

For each option describe: what changes, what the costs are (initial and recurring), and what is needed to revert to the old way. Mark what is unknown and cannot be assessed from this data. Do not invent prices, timelines, or volumes."

The prompt structure: role, three discrepancies, absence of measurement and pricing, a restriction on external actions, and a specific output format. Vague commands like "analyze," "give recommendations," or "formulate a strategy" are left out.

Step 4. Decide with recorded grounds.

Write down which facts were weighed, which assumptions were accepted, and what remained unknown. The next step should be small and reversible.

It matters to separate two levels here. Preparing the report each week is operational work. Deciding to change the process is a management action - one that is taken on the basis of several observations, not a single episode.

A middle-aged man with grey temples in a white shirt stands at an easel holding two office floor plans framed in cyan and magenta. He holds a requirements sheet in one hand and points a pencil at a room on the magenta plan.

What does a comparative memo look like across five options?

Below is a sample of the full memo. All facts come from the teaching task description. Estimates and unknowns are marked explicitly.

Task: determine whether to change the weekly preparation of a consolidated report from three teams.

Current state. The coordinator collects statuses; the manager reviews. The bottleneck is clarifying periods and decoding statuses. No time measurement exists. One table arrives without dates; in another, "done" carries a different meaning than expected. How much this hurts has not been measured.

Option 1. Keep as-is. The report gets prepared; the meeting happens. But exact preparation time is unknown, so it is unclear how serious the problem is. The discrepancies are there; only their size is unknown. Risk: if a wrong status makes it into the report, the decision made at the meeting may be wrong.

Option 2. Single unified template, no AI. The coordinator agrees with all three teams on one format: date is mandatory; "done" must specify whether it means client handoff or only internal completion; period is stated in days. This is an organizational decision, with no technology involved.

Initial costs: the coordinator's time negotiating with teams, and the manager's time reviewing the first reports in the new format. Volume unknown. Ongoing costs: no license needed, but the coordinator's checking work remains. Reverting: another round of agreement with the teams.

Option 3. Learn to use an approved tool. The coordinator or manager learns to use an available AI tool to draft the report from anonymized copies. Costs: time for training and adaptation. Ongoing: a license if the tool is paid (price unknown in this exercise). Revert: stop using it.

Option 4. Buy or configure an off-the-shelf solution. A specialized tool for consolidating statuses. Initial costs: implementation, configuration, training. Recurring: license, support. Consider vendor dependency: what remains if terms change. In this exercise, price and terms are unknown.

Option 5. Build a custom integration. An in-house system for collecting statuses. Initial costs are typically higher than for training. This creates an owned asset if the architecture supports it. Volume and cost are undefined in this exercise.

Comparison against the criteria (fixed before viewing options):

Criterion Keep Template Training Purchase Integration
Accuracy of statuses Depends on coordinator Unified format lowers mismatch risk (estimate) Depends on human check Depends on configuration Depends on development
Preparation time Unknown Unknown Unknown Unknown Unknown
Ongoing costs Coordinator labor Coordinator labor License (price unknown) License, support Operations
Revert to old way Not required Re-negotiate with teams Stop using Depends on contract Depends on architecture

Most cells say "unknown." That is a direct result of no measurement having been done.

Preference in this exercise: Option 2 first, then a trial of Option 3. The logic: discrepancies in terms and dates can be fixed organizationally, with no tools at all. AI makes sense once the format is already unified and the task shifts from decoding to processing a large volume of data. This is not a universal choice - in a different situation with a hundred statuses in ten formats, the logic might differ.

How do you decide when there is no time measurement?

In this teaching example there is no measurement of how long preparation takes. There is no guarantee that checking an AI draft against the sources will not take just as long as manual collection.

The claim "automation will save half the time" is an example of an unverified estimate. Nobody has measured it. In the memo it can appear only as a hypothesis, marked "no measurement performed." If the manager decides to buy a license on this basis, the decision rests on a number nobody checked.

The concrete action within the task constraints:

  1. On the next two Fridays, the coordinator and manager separately record: time spent on preparation (collecting statuses, decoding, checking), time spent on the manager's review, and the number of discrepancies between the source tables and the finished report.
  2. In parallel, over those same two weeks, the manager asks AI to draft the report from anonymized copies and checks the result against the sources.
  3. Sending the report is not allowed. Integration is not allowed. If AI produces a wrong status, the draft is corrected by hand or the trial stops and the previous method is used.

Two reports is the number proposed for this exercise. It is not a scientific norm and not a minimum sufficient sample. It is simply enough to see whether the same discrepancies repeat and how long checking takes.

For example, if after two weeks the difference turns out to be three minutes - that is a basis for comparing the result against the criterion the manager set in advance. The number alone says nothing about payback. For that you need license price, cost of working time, and volume of use. The decision to continue or stop belongs to the manager - they compare what they observed against a threshold they set themselves.

An older man with glasses, a white beard and a navy waistcoat stands at a wooden lectern by a window. On a sheet with five boxed steps, a red magnet sits on the fourth step, which he corrects with a pen. Another sheet with lines is clipped beside it. A minaret and a dome are visible through the window.

When does a task passport help?

For the weekly report, a task passport is overkill. It is useful where a decision touches several processes or requires a budget.

Decision and deadline. What is being decided, by what date, and who is responsible.

Baseline process. How it is done now, which resources, which metrics are measurable.

Approved sources. Which data may be used, its version and origin. What is explicitly excluded.

Known facts and unknown parameters. For each unknown, assign a type:

  • Dispute resolved by source. Two teams gave contradictory statuses - request the primary record from each, compare, and record the result.
  • Parameter is measurable. Preparation time, license cost, data volume - measure it or request it from the vendor.
  • Priority decided by the manager. If two options give comparable results but require different levels of attention, this is a resource question, not a data question.
  • Forecast checked by sensitivity. "We will save an hour a week" is a forecast. What happens if the saving turns out to be twenty minutes? Does the decision change? If yes, measurement is mandatory. If no, the decision is robust.

Five options and criteria. Fixed before viewing results. Each criterion names its data source.

Allowed model actions. What the model does independently, what requires human confirmation, what is forbidden.

Option comparison. Table by criteria: initial and recurring costs, dependencies, what remains an owned asset.

Vulnerable assumptions and counterarguments. What changes if a key assumption proves wrong. For example: the assumption "the coordinator will manage the new template" may not hold. Review condition: if the teams have not adopted the format within two weeks, revert to Option 1 and look at Option 3.

Boundary of hours and money. How much time the manager and coordinator are willing to spend on negotiation and checking before the decision becomes uneconomical. In this exercise, that boundary is not set - no measurement was performed.

Vendor dependency. What happens if the license price rises or support ends. This matters most for Options 4 and 5.

Professional expertise. Where the decision touches labor relations, regulatory requirements, or significant financial commitments, the model's draft is a basis for working with a specialist, not a finished conclusion.

Next testable step. A small trial. Criteria for whether to proceed. Who decides to continue, adjust, or stop.

Log. Date and substance of every change to the framing or sources. Who made it and why.

This template is needed where the decision is expensive, repeatable, or touches multiple teams.

What does research say about AI training for executives?

Harvard Business School AI Institute research (September 2023, 758 consultants) found that AI's usefulness depends heavily on task type. For some tasks, working jointly with AI improves quality. For others, it does not. The boundary is not obvious in advance. This covers a specific tool and specific tasks from 2023 - it is not a generalization about all tools in 2026.

A survey study by Lee et al. (CHI 2025, Microsoft Research) covered 319 knowledge workers and 936 work-situation examples. The authors found a correlation: the higher a person's confidence in the model's answer, the less critical thinking effort they reported exerting. This is a correlation in self-reported data - a co-occurrence of signals, not proof that AI directly reduces critical thinking.

The practical takeaway from both sources: when criteria are set before viewing results, the risk of accepting a polished answer without checking goes down. This is a proposed measure, not a proven outcome. Two studies do not guarantee the conclusion will hold in your situation - they give early data on which to build a decision.

Which training format gives which result?

A program where the manager watches a demonstration and maps priorities across their own tasks differs in outcome from a program with hands-on construction on a real task. The management format is the first kind: the manager understands what is possible and where the limits are. Hands-on construction is the second: the manager leaves with a concrete draft and a recorded list of discrepancies.

The management format excludes independent technical setup by participants. The manager watches real demonstrations, builds a map of what can be checked now, what requires development, and where AI is not a good fit. Hands-on construction is a different format with a different result: instead of a decision map, the manager leaves with a specific draft and a list of discrepancies.

Hours saved on report preparation do not automatically become money. Freed time may go to other tasks or stay unfilled. Converting it to a monetary value makes sense only with explicit justification of where the resource actually goes.

A prototype that works on two anonymized reports is not a system ready for production. Moving from "works in test" to "works consistently" requires a separate decision with explicit criteria. Where the decision touches labor relations or significant financial commitments, the model's draft is a basis for working with a specialist, not a finished conclusion.

Questions usually asked

Do you need real tables to start with this structure?

No. The teaching example uses only composed conditions: roles, three discrepancies, and constraints. The prompt is passed to the model by task description, with no actual data. Real anonymized copies of your own tables are needed at the trial-assembly stage - which is not claimed here. The right first step is to test the structure on composed conditions and understand how the prompt works and how the result is checked.

Is this a client case or a real model run?

This is a teaching example composed specifically to work through the structure of a management decision. No model was run on real data; no report was sent. The task data - roles, tables, discrepancies - are composed for the exercise. The goal is to show what each step looks like and what the manager records at each one, not to demonstrate a specific client result.

Does this structure work for other management tasks?

Yes. Four steps, five options, criteria before viewing results, types of unknowns, next small trial - this is a general frame. Only the subject changes: instead of a weekly report it could be another report, a contract, hiring, or a vendor assessment. Criteria and constraints are formulated for the specific task. The task passport is used where the decision is expensive or touches multiple teams.

Why should a manager set criteria before using AI?

If you choose criteria after seeing the table of options, they start matching whichever option you already liked. This happens quietly: the manager thinks they are deciding by criteria, but they are actually fitting the criteria to the option they already prefer. Research by Lee et al. (CHI 2025) confirms this pattern for knowledge workers. Criteria fixed before viewing allow you to compare options objectively.

What if the teams do not agree on a unified template?

Record this as a review condition in the task passport before starting. For Option 2, write: if three teams have not adopted the unified format within two weeks, Option 2 has failed. Then look at Option 3 - learning to use an approved tool - or return to Option 1 and wait for new data. The disagreement itself is information: the organizational task is harder than it appeared, and needs either a dedicated resource or a different approach.

How do you check that AI did not add data that was not there?

Check every conclusion in the memo against the original conditions. If the model wrote specific cost figures or timelines that were not in the conditions, that is an addition without basis - it needs to be removed or separately verified. The prompt explicitly contains the phrase "do not invent prices, timelines, or volumes" - this reduces the risk but does not eliminate it. If there is no basis for a figure, "unknown" is more accurate than any number.

How does the management training format differ from the technical one?

The management training format is a program where the manager watches real AI demonstrations and maps their tasks across three categories: what can be tested right now, what requires technical development, and where AI is not a good fit. The outcome is a decision map and an understanding of boundaries. The manager does not install tools or configure integrations. The technical format is different: its output is a ready draft or a configured automation, not a priority map.

Four steps for any management task that involves AI

AI training for executives is built on a specific task, not on a general demonstration of capabilities. The four steps from this example apply to any management decision: lock in the task and known facts, set alternatives and criteria before using AI, test the tool on one narrow operation, and decide with recorded grounds. The comparative memo for a weekly report shows what this looks like from start to finish.

If you want to work through your own task with this structure or discuss how to embed such a process in team operations, write to hello@majento.ai or Telegram.

On systematic AI work at the function level, see the executive training page.

Sources

  1. aiinstitute.hbs.edu publication
  2. microsoft.com publication

View as Markdown