Three slides. One decision.

Why executives do not need another attractive report

The team brings forty slides to the meeting: quarterly sales, color charts, departmental metrics, list of initiatives. Twenty minutes later the most important question arises: “What do you propose to do?” If you have to go back to the first table to answer, the presentation has conveyed information but not prepared a solution.

A short presentation solves a different problem. It focuses on one management decision, shows the essential basis, honestly identifies the uncertainty and ends with a specific fork in the road. Three slides are a convenient work discipline, not a universal law. If there is a complex transaction, legal risk or investment committee, more pages and a separate evidence appendix may be required.

AI's job here is not to "come up with a strategy." Its reasonable role is to help structure the validated material, make the language clear, and show the cause-and-effect line without losing the constraints.

Define the decision in one sentence

Before the first prompt, write down the question so that it can be answered with an action. "How's the support doing?" - a topic to watch. “Is it worth changing the procedure for transferring requests between shifts for two weeks?” - a decision that has a deadline, a person responsible and a verification criterion.

A good brief contains six fields: addressee, decision, data period, verified facts, options for action, choice period. If there are few facts, this also needs to be written down, and not mask the missing lines with general words like “the dynamics are positive.”

Example: “For the owner of a service company. You need to decide whether to run a two-week experiment on transferring requests between shifts. The basis is the unloading of requests for September. We show the share of SLA violations and do not draw a conclusion about the reason without checking. The decision will be made by Friday." Such a request is much more accurate than “make a premium presentation on effectiveness.”

Separate facts, calculations, hypotheses and missing data

The most dangerous mistake is not a typo, but turning an assumption into a confident statement. A strong headline, “Shift Handover Causes 60% of Delays,” may appear from data that only notes that a portion of late claims were passed through the shift. This is a coincidence of signs, not a proven cause.

It is useful to do this classification before generating the slides. For each number, record the source name, period boundaries, upload time, unit and owner of the data. If you are using tables from different systems, first agree on the definition of “request” and “overdue.”

Worked example: the fictional service company Mayak

Let's imagine a fictional company called Mayak with two support shifts. All metrics in the example are created to explain the method and are not taken from AGENTUM, the client's CRM, or actual research. According to the training upload for September, there are 240 requests, of which 60 were closed in violation of the internal target deadline of 24 hours. Among these 60 cases, 36 are labeled as intershift transmission.

The share of requests with violation of the deadline: 60 ÷ 240 × 100 = 25%. The share of requests with a transfer mark among overdue ones: 36 ÷ 60 × 100 = 60%. The last number does not mean that shift handover caused 60% of delays: to verify causality, you need to study the routing order, load, complexity of requests and the time the mark appeared.

The owner does not want to immediately buy a new system. His question is more specific: “Does it make sense to try a simple procedure for transferring unfinished requests for two weeks and see the result?” It is for this choice that we will build three slides.

Slide 1 — What is happening?

Headline with conclusion: “Every fourth request in the illustrative sample exceeded the target response time.” Instead of a set of abstract KPIs, we show the number 60 out of 240, the formula 25% and two brief captions: “September” and “internal 24-hour SLA”. Next to it is a neutral explanation: “36 overdue requests had a shift change mark.”

For visualization, a simple scale of 240 units or one column with the share of overdue payments is suitable. There is no need to depict a complex analytics dashboard that the company does not have. Below is briefly: “Source: educational register of appeals; illustrative indicators; reconciliation of primary records is required.”

What should not be: a fictitious 37% growth, a missing sales plan, a beautiful but not relevant chart or a caption “the reason is poor shift handover.” The first slide is responsible only for the observation and quality of the fact.

Slide 2 — What are the options?

The next slide compares three actions and their cost in an organizational sense. There is no need for fictitious financial forecasts here. The goal is to show where the uncertainty actually lies.

For option B, propose a specific process: at the end of each shift, the employee notes the unfinished case, its owner and the next step; The manager checks the list twice a working day for ten minutes. Over ten working days this is 200 minutes, or 3 hours 20 minutes of supervisory review. This is a calculation of the time spent on viewing, and not the full cost of the experiment: staff time for maintaining the register must be measured separately.

Mandatory phrase on the second slide: “An experiment tests a hypothesis; improvement has not yet been proven.” It reduces the risk of mistaking a beautiful plan for the result already achieved.

Slide 3 — What decision should we take?

The final page should not end with the generic “increase efficiency.” Instead, you can make a draft decision: “Run option B for ten working days after approval by the owner.” Below are four fields: the person responsible for the registry, the launch date, the control indicator and the date of analysis of the results.

The training objective is to see if the delinquency rate can be reduced from 25% to 12.5% or lower in a comparable period. This means no more than 30 late hits only if the total volume is again 240. If the volume is different, compare the percentages and difficulty distribution rather than the absolute number of 30.

Stopping conditions are also important: if the register creates a leak of personal data, employees cannot reliably fill it out, or the selection of requests changes so that the periods are not comparable, the experiment needs to be reconsidered. The decision to change the process is made by the manager; the presence of text on a slide does not mean that the task has already been assigned in the system.

One reproducible briefing prompt

Below is a working template. It does not require AI to have access to CRM: it is permissible to use a pre-prepared and verified set of aggregated information.

“Prepare an executive presentation of 3 slides in English, 16:9 format, for the owner of the fictional service company Mayak.” The goal is to discuss a decision on a two-week experiment in transferring requests between shifts. Use only these training facts: September, 240 support requests, 60 violations of the internal 24-hour SLA, 36 of the overdue ones have a shift flag. Calculate the share of delays of 25%; don't conclude that shift handover is the cause. Slide 1: observation and source. Slide 2: A/B/C options, pros and limitations. Slide 3: draft decision, responsible person - “to be specified”, goal - to check the share of delays no higher than 12.5%, revision date - “agreed”. Don’t invent sales, savings, ROI, names of employees, clients, agreements and successfully completed actions. If there is a lack of facts, directly indicate what needs to be checked. Use premium, but meaningful graphics and readable headings; Don’t turn your slides into advertising posters.”

If the problem concerns a real business, replace training data with only permitted facts and submit it to a tool consistent with your organization's access policy. Confidential documents and client identifiers should not be included in the “just in case” prompt.

Visual hierarchy: one clear point per slide

Visual quality is not equal to quantity of detail. On the first slide, the main indicator can be one indicator with an honest denominator. The second contains three comparable alternatives without an artificially selected “winner”. On the third - the proposed action, the person responsible and the moment of re-checking. This plot can be understood even when viewing a PDF on a phone in Telegram.

Check readability on a real scale: are the headings cut off, are there noticeable disclaimers, is the brand spelled correctly, is there a random number in the illustration. Be especially wary of text drawn inside an image: a visually compelling AI slide may contain an extra word, an erroneous unit, or an incorrect caption.

Do not use “realistic” CRM screenshots unless they are genuine, authorized images. For a management presentation, correct infographics, an actual product and a well-staged scene that do not imitate a non-existent interface are enough.

What AI can and cannot verify independently

AI is able to propose a structure, check an arithmetic expression, reformulate the conclusion, and compare the draft with a given list of constraints. But he does not know whether the upload is true, whether the specified contract exists, or whether the manager actually completed the assignment - unless it is confirmed by an authorized source.

Separately check four boundaries: origin of numbers, access to information, conventionality of conclusions and fact of execution. If, after a presentation, someone wants to create an order, send a document to a client, or change an entry in CRM, this is an independent operation with verification of authority, and not a consequence of the fact that the AI ​​wrote “recommended.”

NIST's guidance on the risks of generative AI specifically addresses information reliability, source verification, and the need for human oversight. These guidelines are useful for editorial review of AI presentations, but are not a certification of a specific product or a guarantee of error-free generation.

Why a short PDF is not evidence of an outcome

AGENTUM provides a scenario for preparing a visual AI presentation based on a user request: plan → generation of individual slide images → PDF assembly → delivery to Telegram. As documented as of October 19, 2026, three slides are offered by default, valid request is one to eight, RU, UK and EN are supported. This PDF consists of images: text within slides should not be considered automatically editable or reliably searchable.

Important Availability Boundary: In production this scenario is limited to the Owner Circuit, mass enablement for all users is not confirmed. Processing is based on material from the user's request; Automatic generation of a presentation from connected company reports cannot yet be promised as a published feature. Detailed live quality testing of the new pipeline and end-to-end Telegram → PDF request has not yet been confirmed in the canonical technical documentation as of the date of this article.

Therefore, the guide above is useful regardless of the chosen tool, and the example prompt is a sample problem statement, not a demonstration of the guaranteed result of AGENTUM.

Choose the number of slides to match the decision

Three slides are suitable when talking about one solution, a compact evidence base and clear alternatives. If budget protection, market comparisons, legal disclaimers, or multiple related projects are required, additional pages are appropriate. However, increasing to eight slides does not in itself solve the problem of missing data.

A useful criterion: for each additional page there should be a separate question, without an answer to which the decision will be premature. For example, after “fact - choice - decision”, you can add a separate application with a calculation methodology, a range of uncertainty and an implementation plan. If a new page only decorates the previous one, it is better to remove it.

Before starting, agree which slides are for discussion and which will be kept as supporting materials. Don't try to put a primary table with hundreds of rows into a PDF: attach a valid source file for the internal process, without distorting it in the picture.

Twelve checks before sending it to the executive

1. Decision. Is it possible to formulate the question of the meeting in one sentence?
2. Addressee. Is it clear who makes the decision and who only recommends?
3. Source. Is an authorized primary source indicated for each fact?
4. Period. Do the dates, sampling boundaries and the moment of unloading coincide?
5. Formula. Are the numerator, denominator, units, and rounding correct?
6. Hypotheses Has a hypothesis become a “proven reason”?
7. Predictions. Is the goal of the experiment presented as a guaranteed effect?
8. Alternatives Are the real benefits and limitations of each choice shown?
9. Rights. Do the slides reveal unnecessary commercial or personal information?
10. Visual. Has AI changed the packaging, numbers, inscriptions and meaning of the diagram?
11. Action. Has the person responsible, the deadline, and the metric for the next review been specified?
12. Fact of execution. Is the proposal called an order that has already been executed?

Best Practice: Assign each item a status of “checked,” “needs edit,” or “not applicable,” along with a reason. A visually attractive file with incorrect numbers does not pass this acceptance, even if it looks like the work of an expensive agency.

Frequently asked questions

Is it possible to make a good presentation without AI access to CRM? Yes. A vetted, human-prepared summary is often sufficient to discuss a particular decision. The absence of a direct connection does not give the right to invent missing information.

Does a manager always need exactly three slides? No. This is a working composition for one solution. Complex topics require independent evidence; extra pages are not needed just for the sake of space.

Can AI guarantee the accuracy of graphs and text on images? No. Formulas, signatures, numerical facts and the resulting PDF require human verification. Look especially carefully at the numbers drawn directly in the picture.

Can the final PDF be considered a verified corporate report from corporate data? Only if its numbers are actually verified with authorized sources. The mere appearance of a PDF in Telegram does not confirm the relevance of the CRM, access authorization or completion of any recommended action.

Does the mention of the AGENTUM feature mean that it is already available to every client? No. As of the date of preparation of the material, the new AI presentation pipeline was deployed for the owner's restricted access only, without a confirmed mass launch. The availability of a specific function must be checked separately.

Conclusion: one presentation, one verifiable next step

A good management presentation is not a substitute for a leader. It makes the choice clearer: it shows the fact being verified, the limits of confidence, several options and a specific action for approval. If after the third slide you know what to decide, who to assign, when to double-check, and where the data came from, the presentation has done its job.

The approach can be used when preparing queries to AGENTUM and other tools. Start with one real question, a minimum set of data allowed, and a simple fact → choice → action scenario. Agree on the meaning first, then ask the AI ​​to shape it.

Sources and context of verification (October 20, 2026): AGENTUM - description of the current AI presentation pipeline; AGENTUM - current product status; NIST - Generative AI Profile; NIST AI Risk Management Framework. The Lighthouse case study is completely fictional; calculations are given to explain the method and do not constitute AGENTUM data or a promise of results.