AI Training Programs

A young woman with bright pink hair and a red-orange jacket stands by wooden mail cubbies, holding an envelope and a sheet of paper. In front of her are two separate pockets holding papers.

Choose corporate AI training by the specific learning outcome you need, because a course title says little about its content. A "practical AI course" from one provider is ninety minutes of demonstration, from another a three-day hackathon that ends in a prototype, and from a third a video course with a quiz. All three make sense for different roles, but trouble starts when the buyer and the provider do not agree on expectations before the contract is signed.

Corporate AI training formats

Before you look for a provider, decide what type of product you need. Below are four types of program. This is a practical grouping for comparing programs, put together for this article.

Literacy and orientation. A short course for the whole company. It explains what modern language models are (neural networks that answer questions and work with text), where they give reliable results, where they can be wrong, and what risks arise when they are used at work. It suits a first step before deeper training, or employees who need AI only as background for work conversations. After such a course a participant understands what is being discussed in meetings and can ask a sensible question about an implementation proposal. They will not yet configure tools themselves - that is the next format.

Management briefing. A format for managers and decision-makers. The goal is to teach a manager to set the limits of delegation, assign owners, set priorities and define acceptance criteria. The program usually includes a review of the company's processes, a map of possible changes and a structured decision on next steps. After the course, the manager sets tasks and assesses the output - that is, the result the tool produces. Judge such a program by how well it fits the manager's role; hands-on work with a tool belongs to a different format.

Role-based practice. Training for a specific role on its real tasks. A marketer learns on their own texts, an analyst on their own data, a product manager on their own briefs. Such a course needs preparation: real working files and a description of the current process. The verifiable result is that the participant performs a specific operation with the tool independently and explains where the tool helped and where its output needed checking.

Prototype and hackathon. A hackathon is an intensive workshop of one or several days in which a team builds a working prototype of a solution and defends it. The result is visible at once: there is a prototype, the task is described and the limitations are recorded. A prototype shows that the task can in principle be solved with the tool. Production use needs reliability, integration with existing processes, access management and technical support. That is separate development work, which training cannot replace, even if the prototype looks finished. Defending a prototype shows that the team took the task through to the end; whether a participant can repeat it alone a month later has to be checked by a separate exercise.

A Harvard Business School study with BCG (758 consultants, GPT-4, 2023; described on the HBS AI Institute site) found that on tasks inside the AI frontier, speed rose by more than 25% and quality, as rated by experts, by more than 40%. On tasks outside that frontier there was no such gain (Navigating the Jagged Technological Frontier). These data come from specific tasks and the tool that was studied; the study did not measure any training course. The takeaway for choosing a format: define the task first, then pick the format.

How to write a request to a provider

Negotiations often open with "what do you have?", and then the provider decides what to offer you. It is better to write the request yourself and send the same package to several providers. This takes time, but it lets you compare their real answers on equal terms instead of comparing marketing pages.

Include in the request:

  1. Participant roles. Individual contributors, middle managers, senior management or a mixed group. For a mixed group, decide between one program with separate modules and separate programs for each role.
  2. Current process. How the task you plan to change is done today. Without this the provider cannot judge whether their program fits.
  3. Desired result. A specific wording. "A content manager independently prepares the first draft of an article with the tool" or "a department head can formulate a task for an AI contractor and assess the result" are verifiable results. "Employees will know AI" is not.
  4. Source materials and access. Which files and tools participants have. If data is closed for security reasons, say so at once.
  5. Prohibited data. What must not be used in training: customers' personal data, trade secrets, financial documents.
  6. Format and time. In person or remote, how many hours the team can set aside, and whether there is a deadline.
  7. Criterion for independent work. What the participant does without a trainer after the course. The criterion differs by role. For role-based practice, the participant performs the operation with the tool. For a management briefing, the manager formulates the task and assesses the proposed solution, and running the tool is not part of their criterion. Difficulty answering this question shows how clearly the expected result has been defined.
  8. Documents expected after the course. What will remain: a working guide, a prototype, a set of templates, a process map. These are sometimes called artifacts - any working material that records the outcome of the training. If the provider plans nothing of the kind, ask what exactly carries over into the work.
  9. Need for follow-up support. Whether support is needed after the training, who provides it and in what format.
  10. Budget and timeline. Without these the provider cannot propose a proportionate solution.

Worked example: a manager's request to compare two versions of an internal guideline

Below is a teaching example, made up for this article, on a neutral process. It shows what a completed request looks like and what follows from it for choosing a program. The conclusions apply to this modeled situation only.

The manager's task: employees prepare, for their colleagues, a summary of the differences between the old and new versions of an internal guideline.

What gets lost today. The employee reads both versions one after the other and writes out the differences in free form. Changes to deadlines and conditions that are stated in a single phrase inside a paragraph drop out of the summary. A section that was moved looks as if it was deleted and added at the same time. The summary has no references to sections, so the manager can check completeness only by rereading both versions.

What the employee learns to do. The employee works with AI in three steps. First they set the comparison rules: match the versions section by section and record each difference as a row in a standard format - section, was, now, whether it affects colleagues' actions. Then they get a draft summary from the tool and mark doubtful places separately. Before handing it in, they check every row against the texts of both versions themselves: the AI draft does not count as a source.

What the manager accepts. The manager does not run the tool. They check the summary against four criteria: every row names its section, "was" and "now" are quoted verbatim, the effect on actions is marked, and doubts are set apart. Then they decide whether to send the summary out or return it for revision.

Who should take part. The employees who write the summaries and the manager who accepts them. The guideline owner takes part in preparation: they approve the row format and the list of permitted materials. Format: role-based practice for the employees on anonymized pairs of versions, plus a short separate acceptance module for the manager.

Completed request

Field Answer
Participant roles Employees who write summaries; the manager who accepts them; the guideline owner
Current process Manual reading of two versions, differences in free form, no references to sections
Desired result The employee independently writes a summary in the standard row format; the manager accepts it or returns it for revision against four criteria
Source materials and access Pairs of versions prepared for training; workstations with access to the tool
Prohibited data Personal data, trade secrets, real customer documents
Format and time Practice for the employees, a short module for the manager; hours to be agreed
Criterion for independent work A new pair of versions that the group did not work through in class; no help from the trainer
Documents after the course The standard row format for summaries, an acceptance checklist
Follow-up support Individual review of participants' work difficulties - as a separate item
Budget and timeline A separate estimate is needed: training and individual review as separate lines; start after materials are prepared and the schedule is agreed

Result and check by role

Role Result Check
Employee A summary of a new pair of versions The reviewer checks every row against both versions and compares the summary with a reference list of differences, prepared in advance by the exercise author
Manager A decision to "accept" or "return", with a reason Receives two summaries, one of which misses a deadline change; finds the gap and writes a note stating why that summary should be returned for revision
Guideline owner An approved row format The guideline owner approves the format before the sessions start

Sample row: "Section 4.2 | was: reply within five working days | now: reply within three working days | affects colleagues' actions: yes, the reply deadline is shortened".

What the shared program includes. The standard row format, matching versions by section, checking rows against the text, and the manager's acceptance of the summary.

What belongs to separate mentoring. Individual review of a particular participant's work difficulty, for example a pair of versions they could not handle after the course. A review has its own goal and limits: what exactly the participant should be able to do alone as a result. Helping colleagues and keeping the shared row format in use can be an internal responsibility of the trained employees; the buyer decides what that role involves. Who runs the mentoring, how much time is allotted and in what format are agreed separately.

What belongs to separate integration. Automatic preparation of the summary from the system where guidelines are stored, and publication of the result. This is a technical task. Any integration should be scoped and quoted separately: training helps define its requirements, while development is separate work.

Open questions for the provider and how the answers change the choice

  1. Which pairs of versions does the provider need from the buyer before the start? If the provider asks the buyer to prepare pairs with a known list of differences in advance, the guideline owner's time goes into the budget. If the provider prepares anonymized pairs itself, preparation is simpler, but the practice moves further from real documents.
  2. How is independent performance checked? For introductory orientation, a terminology quiz is enough. A manager's ability to accept a summary or return it for revision is checked by an acceptance task: two summaries, one of which misses a deadline change. For the employees, you need a new pair of versions and a check of their rows.
  3. Is the manager's acceptance module part of the scope of work or a separate program? A separate module can be delivered to the same group and included in the same quote. Ask the provider to state whether the module is included in the scope and how it appears in the full estimate.
  4. Is individual review of participants' work difficulties needed after the course? If it is, describe it separately, with its goal and limits.

Three cases where a course will not help, or will help only partly

Some requests look like "we need training" but point to a different or additional problem. Check these before choosing a format.

The tool is available, but the team does not use it. The cause may vary: there is no convenient place for it in the process, the interface is awkward, people distrust the results, or the benefit for a particular task is unclear. It is worth talking to the team first. You may find that a process audit is needed in addition to training, or that a course built around the real work process removes the barrier by itself.

The tool is used, but the results are not accepted. If managers or customers do not accept what employees produce with the tool, this is a question of quality and checking. Work out where quality is lost: in how the task is set, in editing the output, or in a mismatch with what the recipient expects. Clear acceptance criteria and a check of the result at each step help here.

Integration with company data is needed. Connecting tools to internal databases, a CRM (customer relationship management system) or document workflow is a technical task. A course explains the principles and helps formulate requirements for the integration; a separate team does the development.

Comparing AI training offers: a matrix for the buyer

When you have the providers' answers, compare them in a matrix - a table with the same criteria for each provider. For every criterion, record three things: what exactly the provider answered, what document or example supports it, and what remains unclear.

An empty cell means the answer has not been confirmed. The provider may have left the information out of the first answer and may give it on request, so turn every gap into a follow-up question.

Record the full cost separately: the price of the program, group time (how many working hours it will take), preparation of materials, possible licenses, and the coordinator's time for organizing. Comparing price lists alone gives a distorted picture.

One more row covers the boundaries of the service: what the provider does and what stays with the buyer. Some programs include a preliminary audit of tasks, others start straight with training. Some check participants' results, others stop at a quiz. It is important to understand what you are taking on yourself.

Vague answers to specific questions lower a provider's score, regardless of how good its website is or what the program is called.

A woman with pink hair and a red-orange jacket stands next to a board with two photos of an office with equipment. The right photo has its right edge torn off, and she is holding the torn piece.

Checking a program before you buy: how to verify quality

Before signing, ask the provider to run one exercise on specific material from your process. It lets you judge how well the trainer understands the context of the task, whether the approach carries over to real work, and how the provider reacts to non-standard situations.

If the provider is not ready for this format, ask why: sometimes it is a matter of resources, sometimes a feature of the method. Refusing a trial is a reason to ask a question. If two direct follow-up questions get no answer, that is grounds for ranking the provider lower in the comparison. This is a recommendation to the buyer, not a scientifically measured norm.

Check separately how the program measures results. A multiple-choice test shows knowledge of terminology and rules, but does not confirm that the participant performed a working operation with the tool themselves. The most informative task is the one closest to a real working operation, with a clear acceptance criterion agreed in advance.

Request template and an example of comparing two offers

Below is a table to fill in and send to providers, then a comparison matrix and a walkthrough of two offers.

Request to the provider

Field Your answer
Participant roles
Current process (briefly)
Desired result (specific)
Source materials and access
Prohibited data
Format and time limits
Criterion for independent work
Documents expected after the course
Need for follow-up support
Budget and timeline

Provider comparison matrix

In each cell write three items: the provider's answer; a supporting document or example with a concrete pointer (a link, a document title, or a review with its source); and an open question or unclear point. In the supporting item, write either real material (a program sample, an exercise description) or the note "answer needed".

Criterion Provider A Provider B
Type of learning product
Hands-on practice or demonstration
Result for a specific role
How the result is checked
Carry-over to a working task
Artifact or prototype
What the buyer prepares
Full cost (including group time)
Boundaries of the service

Make the final decision on the whole picture: how well the product type fits the needed role, how complete the answers are, whether the method for checking results is clear, and whether the full cost is realistic.

Example of using the matrix

The following example was made up for this article: it contains no offers from real providers and no customer data. It shows how the matrix makes gaps visible and how a precise follow-up question replaces a repeated general request.

The situation: the buyer needs participants to prepare an answer on their own from a document they are permitted to work with in the exercise. Offer A: a demonstration of the approach and a terminology quiz. Offer B: individual completion of a task with a check of the result. However, B does not describe what exactly the buyer must prepare before the course.

Prompt to the AI for filling in the matrix:

The buyer needs participants to prepare an answer independently from a permitted document. Offer A: demonstration of the approach, terminology quiz. Offer B: individual completion of a task, output check. B does not say what the buyer prepares.

Compare on three rows: participant role, independent result, buyer preparation. Where a provider's answer lacks information, mark it "question needed". Do not name a winner automatically.

Result:

Row A B
Participant role Gains an understanding of the approach; independent practice is not part of the format Performs the operation independently; the trainer checks the result
Independent result Knowledge of concepts is checked by a quiz; performing a working operation is not part of the format The check of the output shows whether the participant completed the task
Buyer preparation Not described; clarification needed Not described; clarification needed

The precise follow-up question for B: "Please state which materials, access and preliminary actions you need from the buyer before the exercise begins."

A explains how the tool works, and independent practice is not part of its format, so it fits the stated goal less well. B is closer to the needed result. The purchase is postponed until B answers about buyer preparation and until the exercise has been tried on a real task of the team. Until B describes that preparation, it stays an open question to be closed before purchase.

Now consider a different goal: introductory awareness for managers, meaning they learn what AI is and where it helps or fails. In that case A would be a suitable choice, because a demonstration and a terminology quiz check that limited goal. They would not show that a manager can set tasks and accept results; for the management briefing described earlier, keep the practical criterion and use an acceptance task like the one in question 2 above. The matrix records fit to the stated goal; it does not assess the absolute quality of a product. AI structures the answers and fills in the cells; the choice and the checking of the result stay with the buyer.

In a warehouse area, a man in a blue work jacket stands at a printer and compares a sheet with a printed label; boxes and an open wheeled equipment case are nearby. A woman with pink hair and a red-orange jacket watches from the side.

Frequently asked questions

Do managers and employees need different training

It depends on what should change in each person's work. A manager needs to understand the limits of applicability, set tasks and assess the output of the tool. An employee needs to perform a specific operation themselves. Combining these goals in one program is possible if the modules are clearly separated: content, examples and acceptance criteria must differ for each role. Without that separation, neither goal is reached in enough depth.

How do you check that the training worked

Define the criterion before training starts. The most informative way is a specific task, close to a real working operation, which the participant performs without the trainer's help. Agree the acceptance criterion with the provider in advance, while discussing the program.

What if the provider does not answer some of the questions in the request

Ask a follow-up question directly: the detail may simply have been left out of the first description. If two direct follow-up questions get no answer, that is grounds for ranking the provider lower in the comparison. This is a recommendation to the buyer, and a refusal to run a trial exercise remains a reason to ask why.

Can open courses replace a corporate AI training program

Yes, for a number of tasks this is a reasonable choice. An open course fits when you need general literacy, when the employee is motivated to learn on their own and when the task does not require work with closed company data. Add role-based practice on real tasks inside the company and an open course comes closer to the format you need. Without such adaptation, the transfer of knowledge to a specific role is limited: the course is built on general examples, without your processes and data.

Why a prototype is not the same as implementation

A prototype confirms that the task can be solved; its limits are described above, in the section on training formats. Production use will need separate development, which training cannot replace.

How to choose corporate AI training: summary

Choosing corporate AI training comes down to matching the format to the role and to a concretely stated result before you buy. Write the request, compare the offers in the matrix and test the program on a real task of your team.

What we publish about our program

Three facts are published about our AI training program: groups of up to 15 people; sessions are held online or in person by agreement; the program is built on the team's working tasks and permitted data, which are agreed in advance. After the sessions, projects and internal helpers are identified. The role of internal specialists after the course is discussed separately. It is convenient to check these facts against the "Boundaries of the service" row in your matrix.

Contact us: Telegram, hello@majento.ai

Sources

  1. aiinstitute.hbs.edu publication

View as Markdown