How to Build a Consulting Executive Summary Slide with AI That Survives Partner Review

What we learned from 20 published Big Four, McKinsey and BCG reports

Two of the 20 executive summaries we scored, from reports by the Big Four, McKinsey and BCG, put the answer first. BCG's review for the New York City Housing Authority opens with its diagnosis, states a claim in its section headings and backs its points with numbers. McKinsey's summary slide for New York's Metropolitan Transportation Authority leads with the result: a total 2020 gap of $7.0–8.5 billion, broken down by source. A reader who stops at the first line still knows what matters and why. That is exactly the habit AI drafts miss by default, and it can be learned.

This guide shows how to build a consulting executive summary slide with AI: what the finished slide needs, why AI drafts miss it by default, and a workflow that helps you close the gap reliably every time.

Quick answer: A consulting executive summary slide comes straight after the title slide and gives the whole argument on one page. It has an action title that states the answer, three to five supporting points backed by evidence in the deck, the decision or next step you need, and a source line. A reader should be able to act on it without reading further.

What a good consulting executive summary slide looks like

This fictional example shows a private hospital group deciding whether to outsource its pathology testing. The callouts identify the four parts a finished slide needs.

Put the answer before the reasoning, following Barbara Minto's The Pyramid Principle. Each supporting point should explain why the answer is right, with figures the reader can check. If the answer is still unsettled, start with a hypothesis tree that tests the working answer.

Unlike a closing key-takeaways slide, the executive summary opens the deck and argues for one answer. The detailed recommendation comes later.

Why AI drafts of this slide get sent back

The strongest published summaries show what to aim for. We scored the executive summaries of 20 consulting reports that governments commissioned and published, dated 2012 to 2024. (How we scored, and the full list of reports.) Two habits were common: specific numbers and a clear ask. Three were rare, and they are what set the best summaries apart: the answer first, headings that state claims, and a visible trail to the evidence.

Executive summaries of 20 government-published consulting reports dated 2012–2024, scored against a five-criterion rubric. Counts, not percentages.

AI drafts struggle with exactly those rare habits, for a predictable reason. A model drafts from whatever you give it and fills the gaps with plausible defaults. Each failure below has a cause you can address in the inputs or the review.

It recaps instead of answering. Given a deck or a set of findings, a model tends to summarise in the order it received them: background, analysis, findings. That produces a summary of the work, not the answer. The two published summaries that stood out show the alternative: one opens with its diagnosis, the other with its headline result. Ask for the answer first, and give the model the decision the slide must support.

It titles the slide with a topic. "Executive summary" is the conventional heading, so that is what a model writes. BCG's NYCHA review shows the alternative: each summary section opens with a claim, such as "NYCHA facing several significant challenges, many of which are mutually reinforcing". The MTA deck's analysis slides do the same for charts, with every title stating the finding. Tell the model the title must state the answer.

It states numbers the deck doesn't contain. Models round, combine and occasionally invent figures, especially when asked for a "headline number". A plausible "$5m" on the summary that appears nowhere in the deck is the fastest way to lose a partner's trust. The strongest published summaries make checking easy. PwC's post-incident review of the cyber attack on Ireland's Health Service Executive footnotes its sources throughout the summary and points readers to the sections behind its conclusions. Every figure on the slide needs a source slide.

Its "so what" is generic. Without the audience and the decision, a model falls back on safe phrases such as "significant opportunities" and "key considerations". Specific inputs produce specific implications.

It leaves out the ask. This is where most published reports shine: 16 of the 20 asked for a specific decision or action. A model only does that if your prompt says what the audience must decide. Otherwise it ends with "next steps: further analysis", which leaves the audience with nothing to decide.

It hedges. Models qualify claims when the evidence they were given is incomplete or ambiguous. Some qualification is right. But hedging every point hides which findings are firm.

It writes for the wrong reader. A draft built from analysis slides often keeps the analyst's vocabulary and detail. A steering committee needs the implication, not the method.

None of these failures means you shouldn't use AI. They mean the AI needs the right inputs, and its output needs a check designed to catch these particular problems.

The workflow: from findings to a slide that survives review

The steps below follow one running example. It's a fictional regional building-supplies distributor deciding whether to consolidate its three warehouses. The company, figures, draft responses and review comments are illustrations written for this guide, not a client engagement or recorded AI output.

The deck behind the summary contains these findings:

The board meets next month to decide whether to proceed.

Step 1: Give AI the inputs that decide quality

A better prompt matters less than better inputs. Before asking for a draft, give the model four things:

1. The decision: what the audience must decide, and when.
2. The audience: who reads the slide and what they care about.
3. The findings: the supporting slides or their content, with slide numbers.
4. The constraints: one slide, three to five points, no figures that aren't in the deck.

I need an executive summary slide for a board deck.

Decision: whether to approve consolidating three warehouses into one,
at the board meeting next month.
Audience: the board, including the CFO. They care about cost, payback,
customer service and risk.
Evidence: the attached slides 4–12. Each finding has a slide number.

Rules:
- Use only figures that appear in the attached slides.
- Cite the slide number for every figure.
- If the slides don't support a point, say so instead of filling the gap.

Step 2: Ask for the answer first

Draft the executive summary slide.

- Write the action title as one sentence that states the answer
  and its main benefit.
- Add three to five supporting points. Each must give one reason
  the answer is right, with the evidence and its slide number.
- End with the decision the board must make today.
- Do not describe the analysis we did. State what it found.

A bare "summarise this deck" prompt, without Steps 1 and 2, typically produces something like this.

Before (illustrative AI draft):

Executive summary

• We reviewed the current warehouse network across three sites
• Utilisation is low at around 60%
• Consolidation could save up to $5m per year
• Customer service impacts need further consideration
• Next steps: further analysis of options

It reads well and it would be sent back. The title is a topic label. The first point describes the work. The "$5m" appears nowhere in the deck. The service point ignores slide 9, which resolves it. And there is no decision.

Step 3: Tighten the action title and supporting points

Rewrite the action title so a board member who reads nothing else
knows the recommendation, the main benefit and the main condition.
Keep it under 25 words.

Then check each supporting point:
- Does it support the title, or is it background?
- Does it state an implication, not just a fact?
- Is it one idea?
Rewrite any point that fails, and tell me what you changed.

Step 4: Run the trace test

The trace test is the step most summaries skip. Every claim on the slide has to trace to a slide or source in the deck. Anything that doesn't is fixed or removed.

List every claim and figure on this executive summary slide.
For each one, give:
- The supporting slide number, quoting the exact figure or finding.
- A status: Supported, Imprecise, Unsupported, Contradicted,
  or Omitted (an important finding in the deck that the summary leaves out).
- The fix.
Do not assume a figure is supported because it looks plausible.
If you cannot find it in the slides, mark it Unsupported.

Run on the Before draft, the result looks like this:

Two lessons from this table apply to any deck. First, the most dangerous error, the unsupported $5m, looked the most authoritative. Second, the draft was missing the cost, the risk and the ask. A summary can be accurate and still fail because it leaves out what the decision depends on.

Check the AI's trace table yourself, too. It can misread a slide or miss a figure. The table makes that check fast, because each row points to one slide.

The trace test's limit: it checks the summary, not the deck

A summary can trace perfectly to a deck that is wrong. The Victorian Government's business case for the 2026 Commonwealth Games shows the gap. It was prepared by the then Department of Jobs, Precincts and Regions in coordination with EY and others, with EY carrying out the economic and financial analysis. Its opening box gets to the point quickly: after a one-sentence definition, it states the endorsement sought. It sets out the ask, a funding envelope of $2.7–3.2 billion, and quantifies the benefits. Even if every figure on a summary like that traces to the body, the summary is only as sound as the analysis behind it. Here, the Victorian Auditor-General later found that the department's business case "underestimated the costs and overstated the benefits". It also found the business case was inadequate to support an informed decision (VAGO, Withdrawal from the 2026 Commonwealth Games, March 2024). The summary's own benefit-cost ratio range was 0.7 to 1.6, and the audit notes that projects with a ratio below 1.0 do not usually proceed.

So after the trace test, ask one more question of the figures the decision rests on:

Which figures on this slide would change the decision if they were wrong?
For each one, summarise the evidence and assumptions behind it,
say how confident the deck is, and flag any range that crosses
a decision threshold (for example, a benefit-cost ratio below 1.0).

The trace test keeps the summary honest to the deck. This question tests whether the deck can carry the decision.

Step 5: Ask AI to review it as your audience would

Run this on the revised draft, after fixing the trace-test findings.

Review this executive summary slide three times, once as each reviewer:
1. The engagement partner, who will present it.
2. The client CFO, who will test every number.
3. A sceptical board member, who doubts the recommendation.

For each reviewer, list the two or three questions or objections
they would raise. Flag any question the deck itself cannot answer.
Do not rewrite the slide.

Illustrative output:

Partner: The title gives the saving but not the condition. Lead with the recommendation and say it keeps next-day delivery.

CFO: Is the $3.6m net of the cross-dock? What is the payback including the lease break?

Board member: Why expand Northgate rather than build a new site? The deck doesn't compare these options.

The last comment is the most useful. AI can't settle whether the board needs an option comparison, but it can show you a question the deck can't answer before the board asks it.

Step 6: Make the human pass

Some judgments belong to the person presenting the slide:

• What this audience will challenge. You know which board members care about service and which about cost.
• Which points to lead with. The order of the supporting points is an argument, not a list.
• The final numbers. Confirm figures against the latest model or finance sign-off, not just the deck.
• What to leave out. Some true findings don't belong on the summary.

In the example, the presenter decides to add a backup slide comparing Northgate with a new site, and to put the service condition in the title, because the board has lost customers before.

After:

Consolidating into an expanded Northgate warehouse saves $3.6m a year and keeps next-day delivery for every customer

1. The network is under-used. The three sites average 58% utilisation, and an expanded Northgate can absorb all current volume plus 15% growth (slides 4–5).
2. Net saving of $3.6m a year. Operating cost falls from $13.0m to $8.8m. A small Eastfield cross-dock costs $0.6m a year and keeps next-day delivery for the 8% of customers who would otherwise lose it (slides 6, 8–9, 12).
3. The $10.3m one-off cost pays back within three years of completion. It covers $7.1m of expansion, $2.4m for people and moving, and $0.8m to break the Riverside lease (slides 7, 10, 12).
4. The main risk is timing. The 14-month programme must avoid the October–December peak, so detailed design needs to start in the first quarter (slide 11).

Decision required today: approve $10.3m and start detailed design in the first quarter.

Source: network utilisation and cost model, slides 4–12.

Every figure traces to a slide. The title states the answer and its condition. The reader knows what to decide.

Adapting the workflow to your situation

The steps stay the same for every executive summary. What changes is what you give the AI, what the title has to say, and which extra check matters most.

Recommendation

You've assessed several options and need the audience to back one, such as a market to enter, a supplier to choose or an operating model to adopt.

• Give the AI: the options considered, the evaluation criteria and how each option scored.
• The title must say: which option you recommend and the main reason it beats the others.
• Extra check: ask the AI whether the scores really separate the preferred option from the next best one, or whether the recommendation rests on a judgment the slide should state openly.

EY's options assessment for Scotland's ferry network, prepared for Transport Scotland in 2022, has a feature worth borrowing: a colour-rated matrix and a page of key recommendations that make the options easy to compare at a glance. Put a one-line recommendation above a matrix like that, and the reader gets both the answer and the comparison.

Review finding

You've investigated something that went wrong or underperformed, such as a cost overrun, a failed project or a compliance breach, and need to explain why.

• Give the AI: the findings and the amounts or evidence attributed to each cause.
• The title must say: the verdict on what went wrong and why.
• Extra check: run the trace test on every attributed amount, because these are the figures a reader will quote.

PwC's National Children's Hospital review is a model for attributing cost. It assigns each part of a €450m increase to a cause and shows the build-up in a waterfall chart. Pair a chart like that with a title that states the verdict on cause, and the slide does both jobs. The same discipline of linking findings to source documents applies earlier in the work. See how to turn deal documents into findings your team can use.

Status update

You're reporting progress on work already approved, such as a transformation programme or an implementation plan, to a steering committee or sponsor.

• Give the AI: the original commitments and the evidence of progress against each.
• The title must say: an overall verdict, such as "on track except X".
• Extra check: ask the AI to flag any workstream reported as on track without evidence of progress.

EY's Muskrat Falls assessment for the Government of Newfoundland and Labrador (2017) shows how to report status cleanly: a progress verdict for each earlier recommendation, and a named owner for each new one. Add one overall verdict line above a list like that, and it becomes a status slide.

Decision paper

You need a board, cabinet or investment committee to approve something specific, usually money, at a set meeting.

• Give the AI: the decision requested, the amount, and the analysis behind the main estimates.
• The title must say: what is being asked for and the main reason to approve it.
• Extra check: run the decision-threshold question (under Step 4, "The trace test's limit") before anything else.

The Victorian Commonwealth Games business case has the opening to borrow: its first box states the endorsement sought, the funding envelope and the expected benefits. Step 4 shows the check to add, because its benefit-cost ratio range crossed 1.0.

Doing this as a team in Collivo

The workflow above works with any AI tool. In Collivo, the deck, the evidence and the review can sit in the same project, so the trace test and the review happen where the work is.

• Agents create and update PowerPoint files. An agent can draft the summary slide in the deck itself, within the edits supported for Office files. You can't edit Office files directly in Collivo. The agent updates the file for you, or you can download it, or use Open in to work on a copy in your connected account and then import it again (Generated files).
• Comments attach to a region of a slide. Reviewers can draw a rectangle around a chart or a bullet and comment on that part. Resolve with AI then asks an agent to work on the feedback (Comments and AI-assisted review).
• Agents can review without editing. Where supported, you can ask an agent to check the summary against the deck and comment on any claim the evidence doesn't support, without changing the file. That is the trace test as a review.
• Expert Panel compares perspectives. For the reviewer challenge in Step 5, Expert Panel is designed for comparing perspectives and structured reviews (Agent modes).

Checklist before the slide goes to the partner

Each item matches a step in the workflow.

Inputs (Step 1)

• The decision, the audience and the supporting slides were all part of the brief.

Structure (Steps 2–3)

• The title is a full sentence that states the answer and its main condition.
• There are three to five supporting points, and each gives a reason, not background.
• The decision or next step is explicit, with an owner and a date where relevant.
• The language suits the audience, not the analysts.
• Hedging is kept for findings that are genuinely uncertain.

Evidence (Step 4)

• Every figure traces to a slide or source in the deck.
• Nothing the decision depends on is missing: cost, risk and timing.
• The figures that would change the decision have been tested, including any range that crosses a threshold.
• A source line shows where the evidence sits.

Review (Steps 5–6)

• You can answer the questions the reviewer challenge raised, or you have added backup slides.

Run the checklist on the next summary you draft, AI-assisted or not.

About the scorecard

The scorecard covers 20 executive summaries from consulting reports that governments in the United States, Canada, Scotland, Ireland, New Zealand and Australia commissioned and published, dated 2012 to 2024. Nineteen were written by consulting firms: KPMG (8), EY (5), PwC (2), Deloitte (2), BCG (1) and McKinsey (1). The twentieth is a Victorian Government business case prepared in coordination with EY. Summary length ran from a single slide to a 29-page register of findings, with a median of five pages. Reports reviewed and links checked in October 2026.

How we scored. We scored each summary Yes, Partial or No on each criterion, working from the summary pages themselves. We recorded the supporting page and a short quote for key scores, and re-checked selected scores against the pages.

• "Answer first" means the main conclusion appears before background, scope or method on the first page.
• "Claim headings" means most summary headings state a takeaway.
• "Numbers" means key findings, not process counts such as the number of interviews.
• "Ask" means a specific decision or action, not a general direction.
• "Evidence trail" means sources or pointers to supporting sections.

Limits. This is a small convenience sample of reports written for public release, and only 8 of the 20 are landscape slide decks. It isn't representative of client decks. Several scores were borderline calls between Partial and No; a stricter reading would make the answer-first and heading results worse, not better.

All reports were published by the commissioning government or public body, and the links go to the official copy. Firm names identify the authors of each report; none of the firms reviewed or endorsed this article.