The AI workspace for consultants: decks, frameworks, and research at partner speed

What AI actually replaces on an engagement

Consulting work is not one job. In a single week you structure a problem, interview stakeholders, clean a data extract, build a storyline, write the pages, and defend the recommendation. Those tasks reward completely different model behavior, which is why a single AI subscription always feels 70 percent useful and 30 percent frustrating.

The honest framing is this: AI does not produce the answer on an engagement. It removes the two hours you spend getting to a first draft of something you already know how to build, and it catches the gap you would have found on Thursday if you had more time. Used that way it is worth several hours a week. Used as an answer machine it produces exactly the generic output partners have learned to spot. The highest-leverage instance of that on an engagement is the call recap, which has its own workflow in how to summarize meetings with AI.

Here is the routing most consultants land on after a few weeks:

Task	Best model	Why
Problem structuring, issue trees, 2x2s	GPT	Follows explicit structural constraints, holds MECE logic, returns clean hierarchies
Slide headlines and storyline	Claude	Writes in a human register, resists filler adjectives, keeps a consistent argument thread
Executive summary and client email	Claude	Best at tone control and at saying difficult things diplomatically
Market and competitor scans	Gemini	Strongest at recent web material with sources you can check
Long document and transcript reading	Gemini	1M token context window, so a 120 page annual report fits in one pass
Data extraction into tables or JSON	GPT	Most reliable at strict output formats you can paste into Excel
Sanity checking a recommendation	Whatever you did not draft in	A second model is the cheapest red team you will ever run

You do not need to memorize that. The practical version is: draft structure in GPT, draft prose in Claude, research in Gemini, and check the final logic in a model that did not write it.

Workflow 1: framework generation that is not generic

Asking for "a framework for this client problem" returns a Porter's Five Forces with the client name pasted in. The fix is to give the model the constraint set a manager would give you: the decision the client is actually making, the axes that are allowed to matter, and what has already been ruled out.

Prompt: issue tree

You are structuring a problem for a consulting engagement. Client decision: [the specific decision, for example "whether to consolidate three regional distribution centres into one"]. Build a MECE issue tree, three levels deep. Level 1 must be the 3 to 5 questions that, if answered, resolve the decision. For each leaf, state the analysis required and the data source needed. Do not include branches we cannot test with available data. Flag any branch where the answer is likely already known.

Prompt: 2x2 that earns its place

Propose three candidate 2x2 matrices for this problem. For each, name both axes, explain what a position in each quadrant would imply for the client, and state the single decision the matrix drives. Reject any axis pair that is a restatement of the same variable. Then recommend one and explain why the other two are weaker here. Context: [paste your situation notes].

Prompt: pressure test a framework you already built

Here is a framework I built: [paste]. Act as a sceptical engagement manager. Identify overlapping branches, missing branches, branches that cannot be analysed in the time available, and any place where the structure presumes the answer. Do not rewrite it. List the problems only.

The last one is the highest value prompt on this page. Models are far better at critiquing structure than at inventing it, and a critique pass takes 30 seconds. Run your own framework through it before you take it into a manager review, not after.

Workflow 2: deck narrative that survives a partner review

The failure mode of AI written slides is not grammar. It is that every headline is a topic rather than an assertion. "Market overview" is a topic. "The category grew 11 percent while our client lost 3 points of share, driven entirely by the discount channel" is an assertion, and a deck of assertions reads as thinking rather than as reporting.

Prompt: storyline first, pages second

Build a horizontal storyline for a client deck. Situation: [2 sentences]. Complication: [2 sentences]. Question the client is asking: [1 sentence]. Evidence I have: [bullets]. Recommendation: [1 sentence]. Return 8 to 12 slide headlines, each a full assertion with a subject and a verb, that read in sequence as a single argument. No topic labels. After the headlines, list any assertion I do not yet have evidence for.

That final instruction is the one that saves you. It gives you an evidence gap list before you build pages, which is exactly the list a partner will generate in the review anyway.

Prompt: write the page

Write the body content for this slide. Headline: [paste headline]. Supporting data: [paste numbers or findings]. Audience: [for example "COO and two direct reports, operationally fluent, sceptical of consultants"]. Constraints: three supporting points maximum, each under 20 words, no adjectives that cannot be measured, no claim that goes beyond the data given. Add a one line "so what" for the client.

Prompt: the tone pass

Rewrite this to match the voice sample. Keep every number and claim identical. Remove any sentence that would appear unchanged in a deck for a different client. Voice sample: [paste two paragraphs from a deck your partner liked]. Draft: [paste].

The voice sample matters more than any instruction about tone. Two paragraphs of your firm's actual writing will move the output further than a paragraph of adjectives describing how you want it to sound.

Before you send anything, run the deck through the reverse test: paste your headlines back into a fresh chat and ask What is the argument being made here, and what is the weakest link in it? If the model cannot restate your argument, neither can the client.

Workflow 3: market scans you can actually cite

Research is where AI is most useful and most dangerous on an engagement. A confident, fabricated market size number that reaches a client deck is a career problem, not an inconvenience. The discipline is simple: use a web connected model, demand sources, and treat every number as unverified until you have opened the source yourself.

Prompt: category scan

Research the [category] market. Return: (1) market size estimates from the last 24 months, each with the publishing organisation, the year, and the URL; (2) the 5 to 8 largest players with any publicly reported revenue or share figures; (3) material changes in the last 12 months, meaning entries, exits, funding rounds, acquisitions, and regulatory changes; (4) the two or three structural forces most likely to change the competitive picture in the next 24 months. Where sources disagree, show both numbers rather than averaging them. Mark anything you cannot source as UNVERIFIED.

Prompt: competitor teardown

Build a positioning teardown of [competitor]. Cover: stated value proposition in their own words, target segments, pricing where public, distribution channels, recent product or messaging changes with dates, and the customer complaint themes visible in public reviews. Cite each claim. Separate what is stated by the company from what is observed by third parties.

Prompt: reading the long document

Read this document and extract only what matters for [the client decision]. Return a table with: claim, page or section reference, and whether it supports, contradicts, or is neutral to the hypothesis that [state the hypothesis]. Then list the three things in this document that most surprise you against industry norms.

The last prompt is why a large context window matters in this job. Loading an entire annual report, a full regulatory filing, or a set of 15 interview transcripts and asking one question against all of it is a genuinely different capability from summarizing a page at a time. Gemini 3.7 Flash reads a 1M token context on Whizi, roughly 2,000 pages of text, so it is the default for that pass.

For a deeper walkthrough of sourcing and verification, see the guide on AI for market research.

Client confidentiality: what to redact before you paste

Every consultant asks this before the second prompt, and the honest answer has two parts. Whizi does not train on your conversations, and each provider's data handling policy is available before you enable it. That covers the platform. It does not cover your client agreement, which is usually stricter than any vendor policy and is the document that actually governs you.

The workable habit is to strip identity without stripping structure. AI reasoning does not need the client name to be useful, and removing it costs you nothing:

• Replace the client and competitor names with tokens: Client A, Competitor B, Region 1.
• Round or index figures. "Revenue indexed to 100 in 2023, 94 in 2025" preserves every bit of the analytical shape.
• Remove names, emails, and titles from interview transcripts before uploading. Roles are enough, and roles are what the analysis uses.
• Keep anything covered by a specific NDA clause, personal data, or material non public information out of any AI tool entirely, including this one.
• If the engagement is under heightened restrictions, check with your firm's risk team before the engagement, not after.

A redacted prompt is usually a better prompt anyway. Forcing yourself to describe the situation in structural terms removes the incidental detail that was diluting the model's attention.

Why one workspace beats three subscriptions here

The specific advantage on an engagement is not price, although the arithmetic is real: ChatGPT Plus, Claude Pro, and Google AI Pro run about $20 each, roughly $60 a month, against $29.99 for Whizi Pro, and a Claude Sonnet 5 message spends 10 credits of the 2,000 Pro carries each month, so a deck's worth of headline passes barely dents the allowance. Independents feel a sharper version of that arithmetic, worked through in AI tools for freelancers. The real advantage is that context survives the model switch. You upload the client's data extract, structure the problem with GPT, draft the storyline with Claude, and pull the market context with Gemini, all in one thread that remembers what came before. Doing the same across three products means re-explaining the engagement three times and holding the version history in your head.

Side-by-side comparison is the second advantage, and it maps directly onto how consulting quality is judged. Running the same executive summary through two models and keeping the better paragraph is a two minute habit that measurably raises the floor of what you send. It also reduces the "this sounds like AI" reaction, because you are shipping the best output for each section rather than one model's house style stretched across an entire document.

The third is the red team pass. Ask a model that did not draft the recommendation to argue against it: Argue the strongest case that this recommendation is wrong. What would have to be true for the opposite conclusion to hold? Whatever it finds, the client's CFO would have found too.

What AI will not do for you

It will not tell you what the client actually wants, which is usually different from what the RFP says and is only learnable from the room. It will not know that the COO who has to implement your recommendation blocked the same idea two years ago. It will not carry accountability for a number, so every figure that reaches a page is yours to verify.

It is also weak at the specific thing consultants are paid for: judgment under incomplete information with reputational consequences. A model will happily give you a confident answer from thin evidence, because confidence is cheap for it and expensive for you. Treat every output as a well read analyst's first draft, produced by someone who has never met the client and cannot be blamed.