How to Build a McKinsey-Style Presentation with AI: 10 Habits from 45 Real Consulting Decks

A McKinsey-style presentation is a deck you can understand from its slide titles alone. The name is shorthand for the style the top strategy and consulting firms share: the answer comes first, every title states a conclusion, every chart proves one point, and every number shows where it came from.

Three titles from BCG's deck for Dallas City Council show the idea: "Dallas dogs can be conceptualized as buckets and flows", then "Issue is difficult to fix because it requires coordinating efforts", and finally "Dallas' issues are solvable, but require community action". You know the argument before you look at a single chart.

Can AI build a deck like this? Yes, but not from "make me a McKinsey-style deck". A deck that holds up under a partner's or a client's scrutiny comes from three things working together. The standard: knowing what a great deck looks like, which is what the habits below set out. The context: the research, evidence and decisions your team builds up over the engagement, which are the raw material every slide is made from. And the discipline: how you brief the AI, check its understanding and correct it before it drafts. This article covers all three.

We read 45 decks that McKinsey, BCG, Bain, Deloitte, PwC, EY, KPMG and Accenture produced for real clients, and that clients, regulators, courts and legislatures later made public. Below are the 10 habits the strongest of them share. Each comes with what it gets you, a real slide you can open, and how to get AI to produce slides that follow it. You can also browse all 45 decks.

Try the quickest test now: copy your deck's slide titles onto one page and read them in order. If they don't tell the story, start with habit 2.

The 10 habits follow the order you use them across an engagement: plan the storyline from the first week, build slides as the analysis lands, make the deck usable before the decision meeting, then check it.

Plan the storyline from week one

1. State the question, and what the deck will and won't do
2. Write the storyline first, as a page of leads
3. Hang the deck on one framework and number everything
4. Put the answer on page one

Build slides as the analysis lands

1. Make each chart prove one point
2. Keep facts and interpretation apart
3. Show your working
4. Label how firm each number is

Make it usable before the decision meeting

1. Compare options on named criteria, and show what you rejected
2. End with owners, timing, size and the decisions needed

Then check the deck before it goes to the partner, and use the whole process on one page.

Plan the storyline from week one

A final consulting deck is the last version of a document the team has been writing since the first week. In our consulting work, the storyline started as a page of hypotheses, the answer we expected, before most of the analysis was done. Each round of interviews, data work and client conversation then confirmed some leads, changed others and added new ones. By the steering committee, the deck reflected weeks of analysis, synthesis and judgement; it wasn't written in one night like a university assignment. That's why the strongest decks feel inevitable: the argument was settled, and tested, before anyone built a slide. Changing a page of titles is quick; rebuilding a finished deck the night before a steering committee isn't.

All of that work leaves a trail of evidence, and that trail is what makes the final deck defensible. It's also what AI needs to produce one. Keep it in one place as the work progresses: the brief and workplan, a log of sources, interview notes and syntheses, analysis outputs, the decisions the client has made, and each version of the storyline with the feedback on it. Every claim on a slide should trace back to something in that trail. If it can't, neither you nor the AI should put it in the deck.

The habits below are your standard: they tell you, as the person responsible for the deck, what every slide must meet, whoever or whatever drafts it. Each ends with how AI can help, and every one of those steps works from the material above.

1. State the question, and what the deck will and won't do

What it gets you: every slide can be tested against the question, and arguments about what the work covered are settled before the numbers appear. It prevents the deck that answers a slightly different question from the one the client asked.

How: name the question early, in the client's terms. McKinsey's study for US Ski and Snowboard puts it in a side box near the front, headed "Core question that McKinsey was engaged to address". Then say what's out of scope. Bain's diagnostic for Syracuse University has a column headed "THIS REPORT DOES NOT...", which includes making recommendations or decisions. That one line can reassure a stakeholder who worries the consultant is deciding their future.

Why it works: a stated question gives every slide a test (does this help answer it?) and gives the client an early chance to correct it. A stated scope makes disputes about coverage surface before the numbers are presented, not in the middle of them.

Common slip: the question drifts as the work goes on, and the final deck answers what the team found interesting rather than what the client asked.

How AI can help: give it the proposal or statement of work, the kickoff notes and the stakeholder interview syntheses. Ask it to restate the client's question in their own words, list what's in and out of scope, and flag where stakeholders describe the question differently, which are the points to settle with the sponsor.

2. Write the storyline first, as a page of leads

What it gets you: a partner can check the whole argument on one page, in minutes, before the team spends days on slides. It catches a wrong answer, a missing step or a weak link while they're still cheap to fix, and it keeps the analysis pointed at what the deck needs to prove.

How: write only the leads, one full-sentence title for each slide you plan, often called an action title. Then put every lead on a single page, in order. In our consulting work we called this the storyline slide; other teams call the leads-only draft a ghost deck or a skeleton deck. Start it in the first week from your hypotheses, review it with the engagement lead, and update it each time new analysis lands. The result reads like Deloitte's service efficiency study for Toronto's 311 division, where the titles alone tell the story: "Operations appears to account for most of 311's FTE costs".

Why it works: the storyline is the argument with everything else stripped away, so its flaws are easy to see and cheap to fix: a missing step, a lead the evidence doesn't support, a conclusion that doesn't follow from the points above it. Cole Nussbaumer Knaflic calls reading only the titles in order "horizontal logic": if the titles alone tell the story, the deck will too (Knaflic). Writing each lead as a full sentence matters as well, because a sentence tells the reader what to take from the slide. In a classroom study, students taught from slides with sentence headlines averaged 79% on recall questions, against 69% for phrase headlines (Alley et al., 2006). The gain is modest, and the groups were taught in different semesters, but a later randomised study of assertion-style slides points the same way (Garner and Alley, 2013).

Common slip: jumping straight into building slides, or writing the storyline once and never updating it as the evidence changes. And leads that name a topic ("Governance findings", "Market overview") instead of stating what you found.

How AI can help: keep the storyline current as the analysis lands

Give it the client's question and the decision needed, your current storyline page, this round's analysis outputs and interview syntheses, and the notes from the last storyline review. Then ask it to update the storyline, not to write it:

Update our storyline using the attached material.
1. For each lead, mark it [SUPPORTED], [CHANGED] or [EVIDENCE NEEDED], citing the analysis or interview note behind it.
2. Rewrite only the leads the new evidence changes. Keep each a full sentence of 15 words or fewer.
3. Suggest leads the new evidence adds, and leads it no longer supports.
4. List the analyses still needed to support the storyline.
Don't strengthen any claim beyond the evidence, and don't invent numbers.

Illustrative output (fictional engagement: week three of a margin review for a building-materials distributor):

Answer (unchanged): Margins are falling because small trade orders now cost more to serve than they earn.

• "Orders under $500 now make up 46% of deliveries but 11% of revenue" [SUPPORTED]: order data extract, week 2.
• "Delivery cost per small order has risen faster than order value" [CHANGED]: the cost model shows the rise is in picking labour, not delivery. Proposed lead: "Picking small orders now costs more than the margin they earn."
• "Branch managers price small orders below cost to win larger ones" [EVIDENCE NEEDED]: interviews raise it; no pricing data yet.

Expect the AI to round off caveats and make claims sound firmer than the evidence allows. The engagement lead still decides what the answer is.

3. Hang the deck on one framework and number everything

What it gets you: the reader learns the structure once and always knows where they are, even in a 100-page deck opened a month later. Numbered parts let everyone, including the client, refer to "opportunity 9" instead of "that slide about managers".

How: pick one structure, such as an issue map (see our guide to hypothesis trees), a set of barriers or a lifecycle, and reuse it from the overview to the deep dives to the close. BCG's cross-technology summary introduces six barriers in one grid, gives each a deep dive, and closes with a grid of actions using the same icons in the same order.

Make the parts distinct and, together, complete: consultants call this MECE (mutually exclusive, collectively exhaustive). Then number the parts and carry the numbers everywhere, as Bain's Syracuse diagnostic does with its 16 numbered opportunity areas: every finding page that follows carries the matching number badge.

Why it works: every slide has a visible place in the argument, so readers spend their attention on the content rather than on finding their way.

Common slip: a framework shown in the overview and never seen again, or one whose parts overlap so that findings could sit in two places. And the opposite problem: a full-page divider or agenda before every short section. Keep dividers for major sections and use a slim tracker inside them.

How AI can help: give it your hypothesis tree or issue map and the current storyline. Ask it to check that the framework's parts don't overlap and together cover the client's question, to show which leads sit under each part, and to flag leads that fit nowhere, which usually means either a gap in the framework or a slide that doesn't belong.

4. Put the answer on page one

What it gets you: busy readers leave with the conclusion even if they stop after one page, and the meeting starts on what to do rather than on what you found. It also prevents a common reason decks get sent back: the point is hard to find.

How: open with the answer and the points that support it, before background or method. PwC Strategy&'s demand assessment for Ireland's international connectivity project opens with nine numbered key messages, each a bold claim with two or three supporting lines.

Why it works: this is the core of Barbara Minto's Pyramid Principle, in which the ideas form a pyramid "held together by a single thought" (Minto). The evidence that follows confirms the answer rather than building suspense.

Common slip: an executive summary that's a wall of bullets, with the answer somewhere in the middle. Our guide to the consulting executive summary slide covers this page in depth.

How AI can help: once the storyline has been reviewed, give it the storyline page, the key exhibits and the client's own wording from steering committee notes. Ask it to draft the summary page from them and to flag any summary point that doesn't match a section title, in the same order.

Pick your deck structure

What it gets you: the deck makes the case in the order your audience needs to hear it, so they reach the decision you want without getting lost or pushing back early.

Lead with the answer, unless the audience isn't ready for it. Busy readers want the conclusion first. But if your audience is sceptical, or doesn't yet understand the problem, build the case before you reveal the answer. McKinsey's deck for the US Postal Service does this deliberately: it shows that the Postal Service's own actions can't close its financial gap before it argues that legislation is needed. If you build up to the answer, still give readers a one-page summary they can read first.

Decide whether it will be presented or read. The same McKinsey engagement produced a detailed deck that covers the postal network's fixed costs on one dense page, and a version for a wider audience, dated the same day, that gives the same point a divider and three single-message slides.

Then order the sections by what you need the client to do:

The orders follow guidance on cabinet papers, business cases, project reporting and evaluation reports (listed in the sources at the end), adapted for consulting decks.

Build slides as the analysis lands

Once a lead is supported by the analysis, it can become a slide. This is where AI saves the most time, and where a short prompt does the most damage. The quality of a consulting slide comes from context: the client's question in their words, what the interviews said, what the numbers showed, what the partner pushed back on and what the client has already agreed. An AI given five lines returns a plausible, generic slide. Given the engagement's material, and worked with the way you'd brief a capable new team member, it produces strong drafts quickly:

1. Brief it with the context: the slide's lead, where it sits in the storyline, the audience, the analysis and sources behind it, and your template's conventions.
2. Ask it to play back its understanding and plan before it drafts: what the slide must prove, which evidence it will use, and what slide type it will choose.
3. Correct course with questions: "Which analysis shows that?", "Will the partner believe that holds for every region?", "Is that a finding or our view?"
4. Let it draft, then review the draft against habits 5 to 8. They double as your review criteria. Revise until the slide passes them.

Habit 5 shows this loop end to end. Check your firm's rules before you start: many consultants can't put client data into public AI tools, so use approved tools or anonymise the inputs.

5. Make each chart prove one point

What it gets you: the reader sees the finding in seconds instead of decoding the chart, and is far less likely to take away a different point from the one you meant. It prevents the meeting stalling while people work out what a chart is saying.

How: give every chart a title that states what it shows, a callout on the number that matters, and one highlight colour for the client. BCG's economic impact study for Ford titles a chart "Ford assembled twice as many full-size pickups in the US as any competitor in 2019" and colours Ford alone. In Bain's Syracuse diagnostic, a red callout on a chart does the work: "211 managers (30%) have only 1 direct report".

Why it works: a chart can say many things, and the reader shouldn't have to guess which one you mean. Knaflic's companion idea, vertical logic, is that everything on a slide should support its title (Knaflic).

Common slip: a dense table or three charts on one slide in tiny type. Show the rows that matter and put the full table in the appendix.

How AI can help: build the slide with it, not from one request

This is the loop from the start of this section, applied to one slide. Illustrative, using the fictional distributor from habit 2.

Brief: the lead "Picking small orders now costs more than the margin they earn", its place in the storyline (the second supporting point), the audience (the client's executive team), the week-three cost model output by order size and branch, the source details, and the deck's chart conventions.

Ask for a playback first:

Before drafting, play back: what this slide must prove, the evidence you'll use from the attached model output, the slide type and chart you'd choose, what you'd highlight, and the callout and source line. Don't draft the slide yet.

Playback: The slide must show that for orders under $500, picking cost per order exceeds gross margin per order. I'll use picking cost and gross margin by order-size band from the cost model. One bar chart by band, margin against picking cost, with the bands where cost exceeds margin highlighted. Callout: the order size where cost overtakes margin. Source: cost model, week 3.

Correct course with questions: "The CFO will ask whether this holds in every branch. What does the branch data show?"

Answer: 9 of 12 branches show the same crossover; the other 3 have higher order values. I'd note that in a footnote rather than a second chart, so the slide keeps one message.

Then let it draft, and review the draft against this habit and habits 6 to 8: one point, a callout on the number that matters, facts separate from your view, a source line, and labels on anything that's an estimate. The AI makes the drafting fast; the playback and questions make the slide right.

6. Keep facts and interpretation apart

What it gets you: the client can accept your facts while debating your view, so disagreement stays on the judgement calls instead of spilling onto the evidence. It prevents a single contested opinion from discrediting a whole slide of solid data.

How: show the reader which parts are evidence and which parts are your view. Bain's Syracuse diagnostic splits its finding pages into "Key facts and data" on the left and "Perspectives" on the right.

Why it works: readers weigh evidence and opinion differently. Separating them lets each be judged on its own terms.

Common slip: report prose pasted onto slides, so that findings, opinions and recommendations run together in one paragraph.

How AI can help: give it the draft slide and the analysis behind it. Ask it to mark each statement as evidence (with its source) or interpretation, and to flag any interpretation presented as if it were a fact.

7. Show your working

What it gets you: readers trust numbers they can trace, and the meeting discusses your recommendation instead of interrogating your spreadsheet. It prevents the moment when one unexplained figure undermines everything else in the deck.

How: put a source line under every chart, show how big numbers were built, and give ranges where the inputs are uncertain. McKinsey's future business model for the US Postal Service rebuilds one chart as the argument moves: the 2020 loss falls from $33 billion in the base case to $15 billion after the Postal Service's own actions, and the next slides show what fundamental change must close.

Why it works: a reader who can follow how a number was built can challenge one assumption without rejecting the result.

Common slip: a headline figure that doesn't match the number in the body. Reconcile every figure before the deck goes out (see the checks below).

How AI can help: give it the slide, the model or analysis export it came from, and your source list. Ask it to trace each number to its source, flag any it can't trace, and flag estimates that aren't labelled as estimates. Ask it to find the numbers that need a person to check, not to fix them.

8. Label how firm each number is

What it gets you: you can share early thinking without it being quoted later as a final finding, and readers know which numbers they can rely on. It prevents a preliminary estimate from turning into a commitment by the time it reaches the board.

How: say when a slide is preliminary, illustrative or an estimate, and put limits next to the conclusions they affect. McKinsey's evaluation of the Global TB Drug Facility labels a waterfall chart "TOP-DOWN ESTIMATE", so nobody mistakes it for a bottom-up count.

Why it works: limits stated beside a conclusion get read; limits in an appendix are easy to miss.

Common slip: the same caveat box repeated on every slide until nobody reads it. State it once, then use a short label.

How AI can help: give it the draft deck and the analysis notes. Ask it to list every estimate, projection and preliminary figure, and to flag any that isn't labelled as such or whose limits aren't stated next to the conclusion they affect.

Make it usable before the decision meeting

By now the slides prove the argument. These last two habits make the deck work for the people who must act on it, including the many who will read it after the meeting rather than in it.

9. Compare options on named criteria, and show what you rejected

What it gets you: the recommendation reads as the result of a fair comparison, not your preference, and you answer "Did you consider…?" before anyone asks. It prevents the meeting reopening options you've already ruled out.

How: set the criteria before showing the options, then judge every option against them. Deloitte's policing review for Nova Scotia sets out six design principles before its options, presents every option on the same template, and marks the ones it doesn't recommend in their titles.

Why it works: criteria agreed up front turn a debate about opinions into a check against standards.

Common slip: options described in prose with no common criteria, so the reader can't compare them.

How AI can help: give it the long list of options, the client's objectives from the brief, the criteria agreed at the last steering committee and your evidence on each option. Ask it to build the comparison table, rating each cell only from that evidence and marking gaps, and to draft one line on why each rejected option lost. Choosing the recommendation stays with the team.

10. End with owners, timing, size and the decisions needed

What it gets you: the meeting ends with commitments, and the deck becomes the client's tracker after you leave. It prevents recommendations that everyone agrees with and nobody owns.

How: close on what happens next: each recommendation with an owner, a timing and an expected size, and the decisions you need from the audience. BCG's deck for Dallas City Council closes with "Dallas' issues are solvable, but require community action" and sets out what each group needs to do.

Why it works: a decision that someone is asked to make in the room is far more likely to get made than a recommendation with no owner.

Common slip: recommendations with no owner, timing or size. A range is fine; nothing is not.

How AI can help: give it the recommendations, the analysis behind each, and the owners and timings discussed in workshops. Ask it to turn each into an action card and leave owner, timing and size marked "to confirm" wherever your notes don't settle them, then to list the decisions the meeting must make. The blanks are the point: they show the team what still needs agreeing with the client.

Check the deck before it goes to the partner

The last stage before a deck leaves the team.

What it gets you: the partner reviews the argument, not your typos, and the client never sees a mismatched number. These checks are mechanical and slow for people, which is where AI is useful and easy for you to confirm.

1. Title-only read. Extract every title in order and read them as a page. Does the story still match the storyline the partner approved?
2. Numbers and IDs. Every figure, unit and recommendation number should match wherever it appears, and match the model.
3. Release sweep. Remove placeholders ("XX", "TBC", "insert text here"), comments, internal file names, blank pages and text too small to read.
4. Explain-back. Ask a colleague who hasn't seen the deck to explain the argument back to you. AI can act as a first cold reader, summarising each slide so you can compare its message with your title, but it doesn't replace a person.

How AI can help: give it the full deck, the approved storyline page and the model export. Ask it to run checks 1 to 3 and the cold-reader summary, and to report problems only, not rewrite the deck.

What makes the deck hold up

The final deck holds up because of three things: the evidence trail behind every title, built over weeks of research; the standard you held each slide to; and the discipline with which you briefed, questioned and checked the AI along the way. Get those right, and AI becomes a fast, capable partner in producing a McKinsey-grade deck. Skip them, and it will produce a confident, polished deck that falls apart at the first hard question.

The whole process on one page

Where AI helps, stage by stage. Each step works on the engagement's own material. The more of that material you give it, the more useful the output.

Checklist before the deck goes out

Plan

• The storyline was started early, reviewed by the engagement lead and updated as the analysis came in
• The client's question and the decision are stated on an early slide
• Scope says what the deck won't do
• One framework organises the deck, and its parts are numbered

Build

• The first content page gives the answer and its supporting points
• Every content slide title is an action title: it states a conclusion, not a topic
• Every chart has a takeaway, a callout where needed and a source line
• Facts and interpretation are visibly separate
• Estimates, ranges and preliminary slides are labelled
• Options are compared on criteria set out before them, and rejected options are shown

Close and check

• Each recommendation has an owner, a timing and a size (or a range)
• The last section lists the decisions needed
• The titles read as a story on one page, and still match the approved storyline
• Every repeated number matches the model
• No placeholders, comments or internal notes remain
• Someone who wasn't on the team has explained the argument back

More habits worth stealing

If you want to go further, these habits also appeared in the strongest decks.

Slide numbers match the page numbers shown in the PDF.

Doing this as a team in Collivo

In Collivo, the engagement's material, the storyline and the deck live in one project, so the team and its AI agents work from the same context as the analysis develops:

• Draft the storyline in a shared document in the engagement's project folder (how to set one up) and reference the source files with @, so readers and the agent can see what each lead rests on (documents).
• Build the deck from the approved storyline: once the storyline has been reviewed, ask an agent to create the PowerPoint file from it and the source files, and save it to your deliverables folder. Then review it slide by slide against the habits above, as you would any draft (generated files).
• Review on the slide itself: comments can anchor to a region of a slide, and an agent can review as a commenter without editing. Expert Panel mode gives a review from several perspectives (comments and AI-assisted review; agent modes).
• Edit the deck however suits you: ask an agent to update it, and Collivo saves each change as a new version you can compare or restore. Or open it in PowerPoint or Google Slides, edit it yourself and bring the updated file back into the project (generated files; view files).

How we did this research

We looked for slide decks that consulting firms delivered to a client, that are at least 10 slides long, and that were made public on an official website by the client, a regulator, a court or a legislature, with the firm's authorship shown in the document (in one case, only in the published file name). We excluded prose reports, conference talks, firm research and briefings, and decks not in English. That left 45 decks from eight firms: McKinsey 7, BCG 10, Bain 3, Deloitte 7, PwC 4, KPMG 7, EY 5 and Accenture 2. They date from 2000 to 2025, and most were prepared for governments and public bodies.

We read every deck in full, noted the practices each one uses well, and checked every slide cited in this article by viewing the page. We compared what we found with published research and expert guidance on slide design. Where a habit comes from the decks alone, we haven't attributed it to research.

Limits: these are decks that someone chose to make public, which may not represent client work in general. We picked examples by editorial judgement, not by scoring the decks, and this isn't a ranking of firms. Firm names identify the authors of each deck; none of the firms reviewed or endorsed this article.

Sources

• Alley, M., Schreiber, M., Ramsdell, K., Muffo, J. and Borrego, M. (2006). Testing the Effect of Sentence Headlines in Teaching Slides. ASEE Annual Conference.
• Garner, J. K. and Alley, M. P. (2013). How the Design of Presentation Slides Affects Audience Comprehension: A Case for the Assertion–Evidence Approach. International Journal of Engineering Education, 29(6).
• Knaflic, C. N. Horizontal logic and vertical logic. storytelling with data.
• Minto, B. The Minto Pyramid Principle: concept.
• Consensus Building Institute. Multi-stakeholder consensus building.
• Federal Chancellery of Switzerland. HERMES: project status report.
• GOV.UK. Open Policy Making toolkit: quick tools.
• Harvard Kennedy School Communications Program. How to write a policy memo.
• HM Treasury and Welsh Government. Guidance on developing business cases (Five Case Model).
• Queensland Government. Cabinet Handbook: information requirements.
• Treasury Board of Canada Secretariat. Guide to reviewing evaluation reports.
• UK Government. Functional Standard GovS 002: Project delivery.
• UNICEF. UNICEF-adapted UNEG evaluation report standards.
• University of Virginia Darden School. Strategic communication: to inform or persuade.

The 45 consulting decks we analysed

All decks were made public on an official website by the client, a regulator, a court or a legislature. We also link the shorter USPS deck we compared with the detailed one; it isn't counted. Firm names identify the authors; none of the firms reviewed or endorsed this article.