Use of AI in Consulting: Where a Small Firm Should Start

Start with research on public information, move to proposals second, and leave client reporting until last. That order follows risk: research touches no client data, proposals reuse your own writing and are read before sending, and reporting involves client figures where a wrong number reaches the client.

The common mistake is to start with whichever task feels most impressive, usually reporting or a "client-facing assistant", and to meet confidentiality problems before having a routine for catching errors. A small firm has no spare partner hours to clean that up. The view here is practice-side: how a consultancy uses AI on its own work, not how it buys AI consulting.

Follow me on Instagram@sagnikteaches

Why the order is research, proposals, reporting

Three questions rank any candidate task for a first project. Does it involve client-confidential information? How bad is a wrong answer, and who sees it first? Can you tell quickly whether it worked? Scoring the three jobs in the title:

Connect on LinkedInSagnik Bhattacharya

Two further jobs sit outside the three but belong in the sequence: meeting notes (add after proposals, with consent) and internal admin such as time entries and follow-up emails (anytime, low stakes). Anything where the output is advice itself, such as recommendations, is left to the people for now: an assistant can pressure-test your reasoning, but the client is paying for your judgement.

Subscribe on YouTube@sagnikteaches

Stage 1: research that uses no client data

Use an assistant to build a reading list and a first-pass briefing from public sources: a sector's structure, a prospect's recent announcements, a regulation's main points. The deep-research modes in the main assistants can read the web and, in some plans, uploaded files and connected apps. ChatGPT's help page says eligible connected apps are available to deep research automatically when permissions and workspace settings allow, so for this stage disconnect anything that reaches client folders, or check the selected sources before you run it. Public in, public out.

I advise small manufacturers on operations. A prospect makes
commercial kitchen equipment with about 60 staff.
Build a briefing for a first call, in this order:
1. What the company says it sells and to whom (cite page).
2. Three industry pressures relevant to operations (cite sources).
3. Five questions I should ask about their production planning.
For every claim give a source link. Mark anything you inferred
rather than found as INFERRED. Do not guess figures.

A realistic result is a tidy two pages with eight or nine links. Now do the part the tool cannot: open every link. Typically two of the links will not support the sentence they are attached to, and one claim marked as found is actually inferred. Delete those. What remains is a briefing in half the time it would have taken to assemble from scratch, with an honest note of what was verified. Say a call-prep job takes 90 minutes by hand; with a tool and the verification pass it may take 50, and your own log, not a vendor's claim, tells you whether that is true for your firm.

The rule for this stage: a claim enters a client document only after a person has opened its source. For a longer treatment, How Consultants Use AI for Client Research Before Discovery Calls goes through the full routine.

Stage 2: proposals from your own material

A proposal combines a client brief, your standard approach, past similar work and a price. The assistant's job is assembly and tightening, not invention. Build a small library first: your last five winning proposals with client names removed, your standard scope paragraphs, your terms. Put them in a Project (ChatGPT Projects and Claude Projects both let you attach files and instructions to a workspace), so the assistant draws on your language and not on generic consultant prose.

Before and after, with an illustrative case. The brief says: "we want help reducing late deliveries; budget around $25,000". The old routine: a partner opens the last similar proposal, rewrites it over two evenings, and gets it out in four days. The new one: paste the brief into the Project with an instruction to draft in the firm's structure, flag missing information, and propose three clarifying questions. The first draft arrives in minutes and contains a confident sentence: "Our approach has reduced late deliveries by 30% in comparable engagements." That figure is not in your library. It is invented, and a proposal with an invented result is a liability. Delete it, and add a rule to the instructions: "Never state results, figures or client outcomes unless they appear in the attached files."

What to keep human: the diagnosis (what the client's problem really is), the price, the scope boundaries and the sign-off. What to hand over: formatting, structure, cutting length, drafting the "why us" paragraph from your own material, checking the proposal against the brief point by point. A useful final pass is a prompt asking the tool to list every requirement in the brief and say where the proposal addresses it, so that nothing the client asked for goes unanswered. How Consultants Use AI to Write Proposals in Under an Hour has the full workflow.

Track two numbers across your next ten proposals: partner hours per proposal and win rate. Ten is a small sample, so treat any change as a hint. If hours fall and win rate does not, the stage has paid for itself; if hours fall and win rate drops, the drafts are too generic.

Stage 3: client reporting, with the safeguards you now have

Recurring reports (monthly progress, a status pack, a findings summary) are the third stage because they combine client numbers, deadlines and a reader who notices mistakes. By now the firm has two habits worth reusing: verify before it goes in a document, and keep invented content out through instructions.

A reporting workflow that stays safe:

1. Numbers stay in the spreadsheet. Calculations are done in your own tools, not by asking a chat assistant to total a column.
2. The tool writes commentary from figures you supply. Paste the finished table and instruct it to describe movements only and never to explain causes you have not given.
3. A named person compares text to table. Every figure in the commentary is checked against the table, line by line.
4. Client data goes in only where the contract allows. Otherwise use anonymised extracts, or the AI built into a tool the client has already approved.

Illustration: an operations consultancy sends a monthly pack to a client with on-time delivery at 91.2% against a 94% target, down from 92.6%. The draft commentary says "performance is slipping due to supplier delays". The supplier delays are not in anything pasted; the tool guessed. The fix is the instruction "describe the change in the figures only", plus the partner adding the real cause from the client conversation. A reporting routine used at scale across a firm is covered in How Marketing Agencies Use AI to Automate Client Reporting, which transfers well to consultancies.

The confidentiality rules to write before you start

Write these down on one page and have every partner and associate read it. They are the real work of stage zero, and an afternoon is enough.

• Approved tools list. Name the plans the firm uses (business plans rather than personal accounts) and ban everything else for client work. Business plans from the major vendors do not train on business content by default, but check the plan page when you buy, because terms move. Note that some business plans have a two-seat minimum, which matters for a very small firm.
• What may go in. Public information: yes. Your own templates and past work with client names removed: yes. Client documents: only with the client's consent or a clause that allows it. Personal data about the client's staff or customers: no.
• Contract check. Read the confidentiality and subcontractor clauses in current engagements for language about third-party processing; an online AI service may count as one.
• Human sign-off. No AI-assisted deliverable leaves the firm without a named reviewer.
• Disclosure line. One sentence in each proposal stating which tasks use AI and that a person reviews all output.
• Retention. Decide how long chat histories and project files stay, and who deletes them when an engagement ends.

The checklist in AI Readiness Checklist for Accountants, Solicitors, Consultants expands on several of these for professional firms.

A 90-day plan for a six-person firm

An illustration, not a prescription. The firm has two partners, three consultants and an office manager.

The review at week 13 is the part that matters. If a stage did not pay in the firm's own numbers, drop it without embarrassment. A sensible outcome for many small firms is two stages kept and one deferred, and a fresh question about what to try next, informed by where the time really goes.

What small firms get wrong in the first month

• Using personal accounts for client work. The terms, retention and training settings differ from business plans, and the firm cannot see or delete the histories.
• Letting the tool produce findings. It can organise what you found; the finding is yours.
• Skipping the source check "because it looks right". A fluent paragraph with a wrong figure is the typical failure, and it reads as more convincing, not less.
• Training nobody. One partner who is good with the tool and four who are not gives uneven quality; run a one-hour session on the rules and a shared set of prompts.
• Measuring nothing. Without a before-and-after log, the firm ends up with a feeling, which is not enough to decide on a second tool.

Meeting notes: the useful in-between step

Between proposals and reporting, most consultancies find that call and workshop notes are where the hours go. Two decisions come before any tool: whether the client agrees to the recording, and where the transcript will live. Ask at the start of every call, record the answer in your notes, and switch the recorder off when the client says no. Then use the summary for what it is good at: a list of decisions, owners and dates, which a person reads against their own memory the same day.

Illustration: a 50-minute workshop with a client's operations team produces a summary with six actions. Two of them are wrong in a useful way. One attributes a decision to the finance manager who only asked about it; another records "reduce stock by 20%" when the client said "we are wondering if 20% is realistic". Those two slips are why the summary is a draft until a partner has initialled it, and why actions are sent to the client for confirmation before anyone acts on them. The pattern, tool drafts and person confirms, repeats at every stage.

For turning those actions into tracked work, How to Turn Meeting Notes Into Tasks Automatically With AI shows the hand-off to a task tool without the guesswork.

A worked hour-count for stages 1 and 2

The point of starting small is to measure. An illustration for the six-person firm above, over four weeks, with every figure invented. Six prospect calls were prepared, and three proposals written.

Run the sums: the recurring tasks save 3.5 + 4.5 - 0.5 = 7.5 hours over the four weeks, and the one-off set-up cost 6 hours, so the firm is 1.5 hours ahead after a month and would expect to be further ahead in month two if the volume holds. Look at the partner review line, though: it went up half an hour, because reviewers now read for invented claims, which is the correct trade. An honest log shows a modest gain and a real cost, and that is a better basis for the next decision than any claim of savings from a vendor or another firm.

If the numbers had gone the other way, with the tool adding hours, the diagnosis would have been one of three: prompts too vague (so drafts needed heavy rewriting), a library too thin (so the tool had nothing of yours to draw on), or too much checking on low-risk text. Each has a cheap fix. None calls for dropping the idea on a single bad month.

Partner objections, and honest answers

"Clients are paying for our judgement, not a machine." Right, and the sequence keeps judgement where it is. The tool handles reading, assembly and first drafts; the diagnosis, price and recommendation stay with partners.

"What if the client finds out and objects?" Then they should hear it from you first. The disclosure line in each proposal turns a possible discovery into a stated practice, and some clients will ask about it with interest rather than concern.

"Won't our work all start to sound the same?" It will if you accept first drafts. Feeding the Project your own best writing and cutting generic sentences in review keeps your voice, and the proposals' win rate will tell you if it slips.

"We are too small to need a policy." The smaller the firm, the more one slip matters, and a one-page rule takes an afternoon. A firm with six people cannot easily absorb a confidentiality incident.

How far to trust an assistant's recommendations is covered in How to Use ChatGPT as a Business Adviser Without Being Misled.

A knowledge library as the quiet fourth project

While the three stages run, collect material that makes every later stage better: past proposals with names removed, end-of-engagement lessons, your standard scope and risk wording, and short "how we do X" notes. A consultant who leaves takes experience with them; a library keeps some of it. An assistant working over a library of 40 or so well-named documents answers "how did we scope the last supply-chain review?" in seconds, and quotes the document so you can open it. Keep documents named by date and topic, one clean version of each, and delete drafts. A messy library produces confident answers from the wrong version, which is worse than none.

Where it can go next

Once the three stages are stable, the next candidates are usually inside the firm, not client-facing: a searchable library of past deliverables, templates for engagement letters, and a standard way to capture lessons at the end of each engagement. They reuse the same pieces, namely a controlled workspace, your own documents and a named reviewer. Hold off on anything where an AI tool speaks to clients directly until the firm has several months of reviewed output and a clear view of what it will not say. If you would rather map this with someone, that is what my AI implementation consultation covers; get in touch here.

Questions a consulting partner asks before starting

Should a consulting firm start with research, proposals or reporting?

Start with research on public information, then proposals, then client reporting. Research uses no client data, so confidentiality risk is lowest. Proposals reuse your own words and are reviewed before sending. Reporting touches client data and numbers, so it needs the rules and checks learnt on the earlier two.

Can we put client documents into ChatGPT or Claude?

Only where your client contracts allow it and the plan's data terms fit. Business plans from the main vendors do not train on business content by default, but contracts often restrict third-party processing. When unsure, ask the client in writing or work from anonymised extracts.

Do we have to tell clients we use AI?

Check each contract and your professional obligations, but telling clients is the safer habit. A line in your proposal saying which tasks use AI, that a named person reviews every deliverable, and that client documents are not uploaded without consent answers the question before it is asked.

How do we stop AI making up facts in research?

Require a source for every claim, open each source yourself, and delete claims you cannot verify. Treat AI research as a reading list with summaries, not as findings. Ask the tool to separate what the source says from what it infers.

How long before we see any benefit?

Judge it on your own time logs after a month, with checking time included. Some tasks get faster, some do not, and the point of starting small is to find out which, cheaply, before changing how the whole firm works.

Further reads

• How to Mention AI Use in Client Contracts and Proposals — Contract and proposal wording for mentioning AI use.
• AI Rollout Plan for a Ten-Person Marketing Agency — A rollout plan for a ten-person services firm.
• AI Literacy Requirements: What Your Staff Need to Know — What each staff member needs to know before using the tools.
• Hire an AI Specialist or Upskill Your Agency Team? — Hire a specialist or train the team you have.
• AI Content Workflow for Agencies: From Brief to Approved Draft — From brief to approved draft with checks at each gate.
• How to Update Your CRM Automatically After Every Sales Call — Keep the pipeline current after every prospect call.
• How to Do Competitor Research With AI in One Afternoon — Scale the stage-one research method to a market scan.
• Will AI Replace My Employees? An Honest Answer for Small Firms — The staffing question behind most partner objections.
• Best AI Tools for Management Consultants: Seven From Research to Decks — Seven AI tools for independent consultants from research to decks, with a confidentiality tier rule, costs and a worked two-week diagnostic.
• AI Consulting Report Writing: From Findings to a Client-Ready Draft — Turn a findings log into a client-ready report with evidence tags, a number check on every figure and an executive summary written last.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business data page, ChatGPT pricing page, Projects and deep research help articles; Anthropic commercial data-use article, Claude pricing page and projects help article (checked November 2026).