Industries / Finance and business services / Legal and professional services / Consulting
Reports and presentations
Reports and presentations are how consultants hand over what they found: slide decks, written reports and memos, and the meetings where these are read.
5
Milestones achieved, from 1987 to 2026.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1987
Slides made at the desk
Forethought, a Silicon Valley start-up, shipped PowerPoint 1.0 for the Macintosh, with program files dated 20 April 1987. Three months later its team joined Microsoft, in what creator Robert Gaskins calls "the first significant acquisition ever made by Microsoft." A business user could now make slides without going through the company's art department.
Credit
Robert Gaskins, Dennis Austin and Tom Rudkin, Forethought, Inc
1987
Answer first, in a pyramid
Barbara Minto developed and taught a writing course at McKinsey & Company for several years, and in 1987 published The Pyramid Principle: Logic in Writing and Thinking. Its rule is that ideas are easy for a reader to grasp when "organized as a pyramid under a single point", with the main point first. Minto says her later textbook "has become a standard text used by all of the major consulting firms."
Credit
Barbara Minto
2003
Engineering by slides
The Columbia Accident Investigation Board found that engineers assessing the foam strike on the space shuttle Columbia briefed NASA managers with PowerPoint slides from Boeing. On one slide, the key fact, that the debris was 640 times larger than the test data behind the damage model (later analysis said 400 times), sat at the bottom of six levels of bullet points. The board called "the endemic use of PowerPoint briefing slides instead of technical papers" an illustration of problematic technical communication at NASA, drawing on an analysis by Edward Tufte of Yale University.
LessonWhen a finding is cut down to fit a slide, the caveat that matters most can end up at the bottom.
Credit
Columbia Accident Investigation Board; Edward Tufte
2018
Six pages, read in silence
In his letter to shareholders for 2017, published 18 April 2018, Amazon's founder Jeff Bezos wrote: "We don't do PowerPoint (or any other slide-oriented) presentations at Amazon. Instead, we write narratively structured six-page memos." Each meeting starts with everyone reading the memo in silence, and he wrote that a good one can take a week or more to write.
Credit
Jeffrey P. Bezos, Amazon
2026
Agents tested on consulting work
Mercor had management consultants, investment banking analysts and corporate lawyers build 33 project "worlds" of files, reports and slide decks, then write 480 tasks inside them, 160 for each job. In its January 2026 APEX-Agents paper, the top score on the consulting tasks was 22.7%, by GPT-5.2, counting tasks done right on the first try. In July 2026 Mercor found models raising their scores by giving several answers where one was required, which it calls scattergunning, and its version 1.1 of September 2026 stopped rewarding such answers.
LessonA score can rise because answers hedge rather than because the work improves, so a test has to reward one committed answer.
Credit
Bertie Vidgen, Brendan Foody and colleagues, Mercor
Watch for
Decks made from a chat
Microsoft says its Office Agent "creates polished PowerPoint presentations and ready-to-use Word documents from chat in Copilot."
Achieved whenan outside test, with readers or experienced consultants who do not know which deck is which, rates its decks as good as those of experienced analysts working from the same brief.
Claimed by
Microsoft, September 2025
Open
01
Decks tested on their readers
A test that scores a slide deck by what a reader takes from it. Each reader gets one deck for 10 minutes and then answers fixed questions: what is recommended, which number supports it, and what is the main risk. Compare decks made by AI tools with decks made by experienced consultants from the same brief, with branding removed so readers cannot tell which is which.
The briefCredit · unclaimed02
Slides or memo, tested fairly
A randomized test of the same findings delivered two ways, as a slide deck and as a six-page memo. Each case holds facts that point to one right decision and one important caveat. Readers get one format at random, and the test records whether they reach the right decision, notice the caveat, and how long it takes.
The briefCredit · unclaimed
Open 01 · the brief
Decks tested on their readers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test that scores a slide deck by what a reader takes from it. Each reader gets one deck for 10 minutes and then answers fixed questions: what is recommended, which number supports it, and what is the main risk. Compare decks made by AI tools with decks made by experienced consultants from the same brief, with branding removed so readers cannot tell which is which.
Why it is still open
Current tests of AI-made slides score the decks themselves, not what readers understand. PresentBench (March 2026) checks each deck against about 54 yes-or-no checklist items, and APEX-Agents grades outputs against rubrics. In a May 2026 classroom study, students rated AI-made and instructor-made lecture slides as similar in quality and could not reliably tell them apart, but no comprehension test was reported.
Built with what exists
briefs and files from the open APEX-Agents consulting worlds (CC-BY license, on Hugging Face); deck tools such as Microsoft's Office Agent, Claude for PowerPoint and Google's NotebookLM; python-pptx to strip branding; an online survey tool to time readers and record answers.
First step
Run one brief with two decks and 10 readers to time the test and fix the questions.
Done looks like
30 briefs, decks from at least 3 AI tools and 2 human authors for each, at least 100 readers, and accuracy and reading time on five fixed questions per deck, published with every deck.
Open 02 · the brief
Slides or memo, tested fairly
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A randomized test of the same findings delivered two ways, as a slide deck and as a six-page memo. Each case holds facts that point to one right decision and one important caveat. Readers get one format at random, and the test records whether they reach the right decision, notice the caveat, and how long it takes.
Why it is still open
The case against slides rests on accounts such as Amazon's own letter and the Columbia board's report. Searches in this session found commentary, and experiments comparing one slide design with another, but no controlled test that compares slides with a memo for the same business content.
Built with what exists
an AI model to draft matched versions of each case in both formats, with a person checking that both hold the same facts; an online survey platform for random assignment; a public pre-registration of the analysis on the Open Science Framework.
First step
Write two cases with a planted caveat and run them with 20 readers.
Done looks like
20 cases, at least 400 readers assigned at random, results on decision accuracy, caveat recall and time published with all materials and the pre-registered analysis.