The short answer
Consultants get the most from AI on research, synthesis and first drafts: sourced desk research in Perplexity, interview synthesis and proposals in Claude, data analysis in ChatGPT or Copilot in Excel, and draft decks in Gamma or ChatGPT Slides. Keep the recommendation, every number and client relationships with people. Use business plans that do not train on client data, and open every source: in October 2025 Deloitte Australia partly refunded a government report containing fabricated references.
| Start with | Perplexity for research, Claude for synthesis |
|---|---|
| Best free option | Perplexity Free with AI data retention off, for public research |
| Keep human | Recommendations, numbers, client conversations |
| Biggest risk | Invented citations in a deliverable |
| Check first | Confidentiality clauses in your engagement letters |
How consultants are using AI in 2026
Consulting is research, synthesis and persuasion under time pressure, which is why consultants adopted AI early and why the mistakes are so public. The work AI speeds up is real: a first scan of a market in an hour instead of a day, twenty interview transcripts boiled down to the themes that matter, a storyline tested before anyone builds slides. The work it cannot do is the part clients pay for: judgement about their situation, and the responsibility for being right.
A small set of tools covers most of the job. Perplexity searches first and numbers every source. Claude is strong at structured thinking and long documents, with Claude Opus 5.5 on Pro and above. ChatGPT is the all-rounder, with data analysis and ChatGPT Slides, which is rolling out to eligible Pro, Business and Enterprise users and exports to PowerPoint. Gemini Notebook keeps answers tied to the documents a client gave you, and Gamma turns an outline into a draft deck quickly.
The 8 consulting tasks worth handing to AI
1. First-pass desk research with sources (Perplexity)
Ask for the market structure, recent moves by named competitors and regulation in play, with a source for every claim. Perplexity's cited answers make checking quick. Turn off AI data retention under Account settings, Preferences on Free, Pro and Max, because it is on by default, and phrase questions without naming your client.
2. Synthesise interview notes into themes (Claude)
Paste anonymised notes from stakeholder or customer interviews and ask for themes with the number of interviews supporting each and direct quotes. Ask separately for the points where interviewees disagreed. Claude handles long inputs carefully and is honest when the evidence is thin, if you ask it to be.
3. Build and stress-test an issue tree (Claude or ChatGPT)
Give the client question and ask for a MECE issue tree with hypotheses and the analysis that would prove or kill each one. Then ask the model to attack its own tree: which branches overlap, what is missing. It is a fast way to get the team arguing about the right things on day one.
4. Analyse a client dataset (ChatGPT or Copilot in Excel)
ChatGPT's data analysis reads a spreadsheet, runs the calculations and draws charts; Copilot in Excel works inside the client's workbook if they use Microsoft 365. Use anonymised data or a client-approved environment, and recompute any number that will appear in a deliverable.
5. Draft the storyline and slides (Gamma, Claude or ChatGPT)
Write the storyline yourself as action titles, then ask AI to build draft slides. Gamma is fast for a first deck; Claude Slides and ChatGPT Slides export to PowerPoint for your template. Gamma's individual plans can use your content for training unless you switch it off, while Business and Enterprise are excluded. Our best AI for presentations compares them.
6. Capture client meetings (Fireflies.ai, with consent)
Fireflies.ai says it never uses meeting content to train AI models. Ask permission before recording every time, and expect some clients to say no. For workshops, a quick summary of decisions and owners sent the same day is worth more than a full transcript.
7. Answer an RFP or write a proposal draft (Claude)
Load the request for proposal and your past proposals into a Claude project and ask for a compliance matrix of every requirement, then a first draft of each section. The matrix alone prevents the most common losing mistake: missing a requirement.
8. Red-team your recommendation (Claude or ChatGPT)
Before the steering committee, ask the model to play a sceptical CFO and list the five hardest questions about your recommendation, with the evidence you would need for each. It will not know the client's politics, but it is good at finding the weak link in an argument.
What consultants should keep human
- The recommendation. AI can list options and trade-offs. Choosing one, and standing behind it in the room, is the job.
- Every number and every citation in a deliverable. A person traces each one to its source before it reaches the client.
- Client relationships and sensitive conversations. Restructurings, leadership issues and bad news are delivered by people who understand the context.
- Scoping and pricing. What you promise, and for how much, depends on knowledge of the client the model does not have.
- Decisions about what may be shared. Whether client material can go into a tool at all is a contract question, not a convenience question.
A starter AI stack for consultants
| Job | Tool | Plan to use | Why |
|---|---|---|---|
| Research with sources | Perplexity | Pro, or Enterprise Pro for teams | Numbered citations; Enterprise data never used for training |
| Thinking, writing, proposals | Claude | Team or Enterprise | Long documents and careful reasoning; no training by default |
| Data and all-round help | ChatGPT | Business | Data analysis and Slides; no training on business data by default |
| Client documents | Gemini Notebook | Work account | Answers tied to the sources you load |
| Meetings | Fireflies.ai | A paid plan | Says it never trains on meeting content |
The competitor report stack on the stacks page chains Perplexity, Gemini Notebook, Claude and Canva into a one-page brief, which works just as well for a client engagement. For research-heavy work, read the best AI for research and our research prompts.
A 30-day AI rollout plan for a consulting team
- Days 1 to 7: check your contracts. Read the confidentiality clauses in your current engagements and your standard terms. Agree which tools are approved, on which plans, and whether clients must be told. Update your engagement letter template if needed.
- Days 8 to 14: internal work first. Use AI for research on public information, issue trees and proposal drafts, where no client data is involved. Keep a shared document of prompts that worked.
- Days 15 to 21: one engagement, with permission. With the client's agreement, use AI for interview synthesis and meeting notes on one project. Log every factual error you catch.
- Days 22 to 30: set the quality bar. Write a review rule for deliverables: every source opened, every number recomputed, AI use disclosed where your contract or policy requires it. Then decide which licences to buy.
The risks for consultants: confidentiality and invented facts
Your contract may forbid it. Many engagement letters and NDAs limit where client information may be stored or processed. Pasting a client's strategy into a personal AI account can breach that, even if the tool is secure. Use business plans: ChatGPT Business and Enterprise and Claude Team and Enterprise do not train on your content by default, Perplexity Enterprise data is never used for training, and personal plans of all three can train on your data unless you switch it off.
Invented facts end up in front of clients. In October 2025 the Associated Press reported that Deloitte Australia agreed to repay part of its fee for a 237-page government report that contained a fabricated quote from a judge and references to research that did not exist. The revised report disclosed that Azure OpenAI had been used, and the department said the recommendations were unchanged. The lesson for every consultant is simple: open every source an AI gives you, and never let a citation into a deliverable that a person has not read.
Confident synthesis can hide weak evidence. A theme "supported by most interviews" may rest on two quotes. Ask the model to count and quote, and check the count yourself.
Disclosure is becoming normal. Clients increasingly ask how AI was used on their work. Agree a short, honest line for your deliverables now rather than explaining afterwards. Our guide on checking whether AI is telling the truth covers the verification habits.
Prompts to copy
First-pass market scan
Give me an overview of the [market] in [region]: the main segments, the five largest players and what each has announced in the last 12 months, key regulation, and three open questions an investor would ask. Put a source link after every claim and say which points rest on a single source.
You get: A sourced market overview to check and build on
Interview synthesis
Here are notes from [number] anonymised interviews about [topic]. List the main themes. For each, give how many interviews support it, two short quotes, and any interviews that contradict it. Then list surprises: points raised by only one or two people that could matter. Do not invent quotes.
You get: Themes with counts, quotes and contradictions
Issue tree with hypotheses
Our client question is: [question]. Build a MECE issue tree three levels deep. For each branch, state a hypothesis and the specific analysis or data that would confirm or reject it. Then critique your own tree: which branches overlap and what is missing?
You get: A workplan skeleton and its weak spots
Analyse a dataset
This file contains [description, anonymised]. Check it for missing values and obvious errors first and tell me what you found. Then calculate [metrics] by [dimension], show the three biggest differences in a chart, and list the formulas or steps you used so I can recompute them.
You get: Checked analysis with steps you can verify
Storyline to draft slides
Turn this storyline into a [number]-slide draft. Use my action titles exactly. For each slide add three supporting bullets from the notes below and suggest the best chart type. Mark any bullet that needs a number I have not provided with [DATA NEEDED]. Storyline and notes: [paste]
You get: A draft deck with data gaps marked
RFP compliance matrix
Read this request for proposal and list every requirement in a table: requirement, section reference, whether it is mandatory, and where our response should address it. Then list the evaluation criteria and their weights if stated.
You get: A checklist so no requirement is missed
Workshop summary
From this workshop transcript, list the decisions made, the actions with owners and dates, the open questions, and any disagreement that was not resolved. Under 200 words, ready to send to attendees today.
You get: A same-day workshop summary
Red-team the recommendation
Act as a sceptical CFO at [client type]. Here is our recommendation and the evidence: [paste]. List the five hardest questions you would ask, why each matters, and what evidence would answer it. Be specific and unsentimental.
You get: The questions to prepare for before the meeting