How a Chief of Staff researches a decision

It's 6 p.m. on a Friday and your principal sends one line: "Can you look into whether we should expand into Germany? Need something for Monday's leadership meeting." You could spend the weekend reading market reports and arrive on Monday with forty pages of notes and no answer, because nobody said what the leadership team was deciding or how sure they needed to be.

Most research that disappoints goes wrong in the first hour: a vague question, no agreed depth, and public sources first when the most useful facts sat in the company's own sales calls. Done well, a clear page on a hard question saves the leadership team a meeting, and it's often the work that turns a Chief of Staff into the person the CEO thinks with. This is part of our Chief of Staff toolkit.

What research does a Chief of Staff do?

A Chief of Staff researches whatever stands between the principal and a decision: quick factual answers, deeper decision research, competitor and market scans, vendor and build-or-buy choices, internal diagnostics on why something isn't working, and sometimes due diligence on an acquisition or investment. What ties them together is that each one ends in a decision someone has to make.

Principals think out loud, and a lot of what they think comes out as a question. Teddy Owen, who was Chief of Staff at Moonhub when we spoke, described on our podcast about building a team around yourself how many discrete research questions his principal sent him, and how he'd started taking each one straight to an AI tool first. The bigger jobs sit inside the most requested parts of the role: in the Chief of Staff Network's 2026 Pathways Report, 43% of more than 1,800 postings ask for strategy and annual planning, 37% for a thought partner on decisions, 27% for special projects and 12% for M&A or corporate development.

You don't need a consulting background for any of this. 27% of the postings in our Pathways data ask for management consulting experience, but only 7% of the Chiefs of Staff in our career histories came straight from consulting. Our Leveling Framework, built from a survey of 200+ Chiefs of Staff, describes the first three levels still doing tactical analysis themselves; from director level up, you frame the questions, set the standard and decide what the research means.

How do you scope a research request before you start?

Before you open a single source, agree five things with your principal: the decision the research serves, the question in one sentence, the deadline, how deep to go, and what's out of scope. Ten minutes of scoping is the highest-return part of any research job, because it decides what you'll read and what you'll ignore.

Principals ask in shorthand. "Look into Germany" might mean "should we hire a sales team there next year" or "the board asked about Europe and I need a view by Thursday", and each needs different sources and a different deliverable. Send this note back within minutes, in whatever channel they asked.

The most useful scoping question is "what would change your mind?" If your principal says "I'd go ahead if at least five of our top twenty customers have German subsidiaries," you can answer it from your CRM in an afternoon. Our post on building a compelling project makes the same point about any piece of Chief of Staff work.

Then agree depth out loud, using three levels: an hour (a quick answer with sources, good enough to talk about), a day (a scan with a view, good enough to discuss) and a week (a memo with evidence, good enough to decide on). Most of the time the answer is shorter than you'd assume. I'd always put "recommendation expected: yes" on the note, because it tells you to read with a view in mind, which makes the reading faster.

A worked example: scoping "look into Germany"

Back to the Friday message. This is a labeled illustration. You reply within ten minutes: "Happy to. Is this for a yes or no on hiring a sales lead there next year, or for the board's Europe question? And what would make it a yes?"

Your principal answers that the board asked, and that she'd back a German sales hire if enough current customers already operate there. You agree a number: at least five of the top twenty accounts with German entities. Depth is a day, due Monday at 9. Out of scope: other EU countries, local entity setup, tax. The note now asks something your CRM and two calls can answer by Saturday lunchtime.

Judgment calls when scoping

If your principal can't say what would change their mind, the decision isn't framed yet; offer two or three framings and let them pick. If the question arrives with its answer attached ("find data showing we should..."), do the work honestly and say early if the evidence points the other way. A principal is better served hearing it from you on Tuesday than from the board in a month. If the request comes from a board member through your principal, ask what the asker needs to see, since a director's question often wants two paragraphs and one chart.

The next time your principal asks you to "look into" something, reply with the scoping note before you start, and notice how often the question changes.

Where do the best answers come from?

Start inside the company: your own data, then the people closest to customers, then customers themselves, then people outside who've made the same decision, and only then public sources and AI summaries of the web. The inside sources are faster, more specific to your situation, and something no competitor can read.

Public research tells you about a market; your own data tells you about your position in it. For the Germany question, the useful facts are how many current customers already operate there, what your sales team hears from German prospects, and what share of support tickets come from European time zones. A fifteen-minute call with the sales lead who covers Europe often answers more than a day of reading. Our post on data-driven decision making suggests short recurring reviews with each department head, so you already know which numbers exist and who owns them; if they don't exist yet, start with our piece on becoming data-driven.

Then talk to people who've been there. Someone at another company who expanded into Germany last year will tell you in thirty minutes what it really cost and which assumption was wrong. Public sources and AI come last, and they're good for context and numbers you can cite. They can't tell you which of the many true things matter for your decision.

That choice is where you come in. Anna Aden, Chief of Staff to the Managing Director for Asia Pacific at the Tony Blair Institute for Global Change, said on our podcast about AI policy that as information gets easier to produce, the scarce skill is finding the kernels of truth that matter to one stakeholder versus another.

How do you use AI deep research without getting burned?

Use AI deep research for the first pass of reading: give it a scoped brief, let it gather and summarize sources, then open every source behind any claim or number you plan to keep. Treat the report as a well-organized pile of leads, check it like a junior analyst's first draft, and keep confidential company data out of any tool your company hasn't approved.

A question that used to take a day of reading now takes ten minutes of waiting and an hour of checking. Abhishek Vora, who was Chief of Staff to the CEO at LaunchDarkly when we spoke, described on our podcast about leading AI adoption using a deep research tool on competitors' public material after a run of acquisitions and funding rounds. A ten-minute run produced a seventeen-page report that helped them plan a competitive response. Then came the caveat:

Most of the quality comes from the brief. It should name the decision, ask for primary sources with dates, and ask the tool to label what's well supported, what rests on a single source and what's its own inference. That last instruction tells you where to start checking.

I'm a Chief of Staff researching a question for a decision my CEO will make. Run deep research and write a report I can verify.

Decision: [what will be decided, by whom, by when]
Question: [one sentence]
What would change the decision: [the evidence that matters most]
Scope: [markets, segments, time period, competitors]. Out of scope: [list]

Instructions:
- Prefer primary sources: company filings, official statistics, product documentation, pricing pages, regulators. Avoid content marketing, listicles and undated pages.
- For every number, give the source, the date of the figure and a direct link. If sources disagree, show both.
- Separate what is well supported from what is a single-source claim or your inference. Label each.
- End with: the 5 facts that matter most for the decision, the biggest open questions, and what you could not find.
- Under 2,000 words. Use tables where comparing more than two things.

The steps are similar in Claude and ChatGPT. I'd use Claude, but you can pick any tool you prefer or your company allows.

Why deep research reports go wrong

A deep research tool reads many pages of mixed quality and writes one confident summary, so a figure from an undated blog post gets the same tone as one from a company filing. Citations can point to a page that mentions the topic without supporting the claim, and figures go stale without a year attached. Andrew Rea, co-founder and CEO of Taxwire, said on our podcast that in precise work nobody wants a tool that's wrong a tenth of the time, and that AI works best turning unstructured information into structure. For research, let the tool gather and organize, and anchor every number you keep to a source you can open.

A worked example: checking a report in an hour

A labeled illustration. The deep research run on German expansion returns 14 pages. You ask for a table of every factual claim with its source, and get 41 rows. Most are background you won't use. You mark 11 that would appear in your memo or change the decision, and open each source. Seven check out. Two cite a page that doesn't contain the number. One market-size figure is from 2021, presented without a date. One competitor "office in Berlin" turns out to be a reseller. Your memo uses the seven, drops the three, and notes the reseller as a single-source claim. The run took 12 minutes; the checking took 55.

Judgment calls on AI research

Verify every number in the memo, every claim that would change the decision, and anything that surprised you; background you won't cite can stay unchecked. Keep three kinds of question away from deep research: anything answerable from your own data, anything confidential your company hasn't cleared for AI tools, and events from the last few days. Confidential material goes only into tools your company has approved, signed into the company account; our guide to setting up AI covers those approvals, and our prompt library has more research prompts.

Once your principal sees what these tools produce, they'll start forwarding their own reports with "thoughts?". Treat a forwarded report like any first draft: find the claims that matter, verify them, and send back what's true and what you'd do.

This week, take one question you'd normally spend a day on. Write the brief, run it, and time how long verification takes. That number is your real cost of AI research.

How do you turn findings into a recommendation your principal can use?

Write one page that leads with the answer and your recommendation, states how confident you are and why, separates what you know from what you're assuming, lays out the options, and puts the evidence in an appendix. Your principal asked a question to make a decision, so the first lines should help them make it.

The common failure is the report that walks through everything you found in the order you found it, leaving your principal to do the synthesis on their phone between meetings. Venture firms solve this with a fixed format. Alexa Gellman, who was Chief of Staff at Inspired Capital when we spoke, described on our podcast about the VC Chief of Staff how much of her role was deal work, including writing investment memos and presenting them to the investment team. An investment memo is a research memo with a decision attached, written to a standard everyone knows how to read.

Put your answer, your recommendation and your confidence level in the first three lines of every research memo.

Confidence is the line most memos leave out. Say it plainly, high, medium or low, with the reason: "Medium: three of five customer calls support this, but our usage data covers only six months." Your principal can then decide how much more evidence they need before acting.

[SCOTT: a research question the Spring CEO asked you, and what the answer changed?]

When the decision itself needs a formal write-up for a group, the decision memo in our guide to briefing your principal takes over, and this memo becomes its appendix.

# [Question, in one sentence]
To: [principal] · From: [Chief of Staff] · Date: [date] · Decision: [what, by when]

## Answer
[Two sentences: the answer to the question.]

## Recommendation
[What I recommend we do, and the first step.]

Confidence: [high / medium / low], because [one line].

## What we know
- [Fact that matters most for the decision] ([source])
- [Fact] ([source])

## What we're assuming
- [Assumption] (how we'd test it: [test])

## Options
| Option | For | Against | What it costs |
| --- | --- | --- | --- |
| [A] | | | |
| [B] | | | |
| Do nothing | | | |

## What would change my recommendation
- [Evidence that would flip it]

## Open questions
- [What we couldn't find, and whether it matters]

## Appendix
- Sources (with dates), data pulls, interview notes (facts only)

# Should we hire a sales lead in Germany next year?
To: CEO · From: Chief of Staff · Date: Monday, 9:00 · Decision: Germany hire in or out of the Q1 headcount plan, by Jan 15

## Answer
Yes, on the test you set. Seven of our top 20 accounts already run German entities, and your bar was five.

## Recommendation
Hire one senior account executive based in Germany in Q1, reporting to the VP Sales, and review after two quarters. First step: the VP Sales drafts the role this week.

Confidence: medium, because the account data is solid but only two of the seven have told us they'd buy more there.

## What we know
- 7 of the top 20 accounts have German entities (CRM and company filings, pulled Saturday)
- 3 of those 7 already pay for seats used by German staff (billing export, last 12 months)
- German prospects ask for local-language contracts and EU data hosting (VP Sales, two recent calls)

## What we're assuming
- These accounts buy faster from a local seller (test: ask all seven in this quarter's account reviews)
- We can hire a senior seller in 90 days (test: recruiter estimate this week)

## Options
| Option | For | Against | What it costs |
| --- | --- | --- | --- |
| One senior AE in Germany | Covers the seven accounts in their time zone; small bet | One person carries the whole test | One salary, travel, recruiter fee |
| Cover Germany from the US team | No hire; can start now | Six-hour time gap; no local contracts | Sales time taken from US pipeline |
| Do nothing | Keeps focus on the US plan | The seven keep growing in Germany without us | Expansion revenue we never see |

## What would change my recommendation
- Fewer than three of the seven say they'd buy more with local support
- EU data hosting turns out to be a hard requirement we can't meet next year

## Open questions
- Entity and tax setup: out of scope here; counsel scopes it if we go ahead

## Appendix
- CRM pull, billing export, call notes (VP Sales, two prospects), deep research run (7 checked claims kept)

Keep it to one page above the appendix. If you can't, the question was probably two questions, and your principal should see them separately.

Judgment calls on the recommendation

When the evidence is mixed, recommend anyway and say the confidence is low. "Low confidence: I'd run a three-month pilot before hiring" is something your principal can act on. When your answer cuts against what your principal expected, lead with it anyway, and put the strongest evidence for their view in the options table. When the memo will travel to the board or the leadership team, agree with your principal whether the recommendation goes with it or stays with them.

How do you research a competitor or a market in a day?

Work from a fixed source list: the competitor's own website, pricing and documentation, their release notes, their job postings, customer reviews, your own sales and win-loss notes, and their filings or funding news. Run AI deep research across the public sources, add what your sales team knows, and finish with a one-page scan that says what changed and what it means for your company.

A fixed list makes the scan repeatable: the first takes a day, the next quarter a few hours. Job postings are the source people forget. A competitor posting for five enterprise account executives in Germany has answered part of your expansion question for you, which is the same reason our 2026 Pathways Report reads job postings: they show what employers need before they say it anywhere else.

Run a competitor scan on [competitor] for [our company, one line on what we do]. Focus on changes in the last [90 days / 12 months].

Use only public sources: their website, pricing page, product documentation, release notes or changelog, job postings, press releases, funding or acquisition news, filings if public, and customer review sites. Give a link and date for every finding.

Report in this structure:
1. What changed: product, pricing, positioning, team (from job postings), funding or M&A
2. What their job postings suggest they're building or entering next
3. Where they now claim to beat us, in their own words
4. What customers praise and complain about most
5. Three implications for us, each labeled as fact-based or inference
6. What you couldn't find

Keep it under 1,000 words. Don't speculate beyond the sources; label anything that is inference.

Then add what only you have. Paste the scan next to your win-loss notes and sales call summaries, and ask what the public picture is missing; ten recent deal notes often say more about a competitor's real position than their whole website. Keep it honest: use public information and what customers tell you freely, never ask a new hire to share a former employer's confidential material, and never misrepresent who you are.

How do you keep research from eating your week?

Time-box every request to the depth you agreed, keep a research log so repeat questions take minutes, hand recurring research to the teams who own the data, and offer the one-hour version when a request doesn't justify more. Research expands to fill whatever time you give it.

Research is interesting, visible to your principal and has a clear end product, which is why it crowds out the operating rhythm and follow-ups. If research took more than a fifth of your week for a month, something is off. Treat the depth on the scoping note as a budget: when you hit it, send what you have with your confidence level and a line on what another day would add.

Keep a research log, one line per request: the question, the date, the answer, the confidence and a link to the memo. When the CEO asks about Germany again in nine months, you start from your own answer, and the log sits naturally next to the decision log in our guide to context management. If you pull the same churn analysis every month, it belongs in a dashboard the data or finance team owns; our post on vendor selection shows the same pattern of moving recurring work to an owner.

I'd offer the one-hour version first for most requests. "I can give you a quick read by end of day, or a proper answer by Thursday. Which do you need?" Most principals pick the quick read. This week, start the research log with the last five questions you answered.