How to Use AI in a Small Consultancy: Proposals, Research and Delivery

Use AI in three stages, in this order: proposals (first drafts from your own past work), research (sourced summaries that a person verifies), then delivery (meeting notes, status updates and report drafting). Set the client-data rule before choosing any tool, because confidentiality decides which plan you may use.

Most small consultancies do it the other way round: someone pastes a client's board paper into a free chatbot on day one. That is the single step that can cost you an engagement. The order below is built so that the low-risk, high-hour tasks come first and the sensitive ones only arrive once you have a policy and a paid plan with the right terms.

Follow me on Instagram@sagnikteaches

A three-person firm and where its hours actually go

Take an illustrative consultancy: two partners and one analyst, selling operational reviews to mid-sized businesses. The partners log about 45 billable hours a week between them and the analyst around 30. Unbilled time is the problem. Over a typical month the firm spends roughly 25 hours on proposals, 30 hours on pre-engagement and background research, and 40 hours on admin around delivery: minutes, action lists, status emails, report formatting. None of those hours appear on an invoice.

Connect on LinkedInSagnik Bhattacharya

That 95 hours is the pool AI can shrink. It is not the billable thinking, which is the product. A fair target for a firm like this is to hand back about a third of the unbilled time, so around 30 hours a month, without touching the quality of the advice. Each stage below comes with its numbers, its prompts and its failure modes.

Subscribe on YouTube@sagnikteaches

Notice that the risk column rises as you go down. That is why the order matters, and why the first thing to settle is what counts as client data.

Decide what client confidentiality allows before you pick a plan

Write three lists on one page. Tell every client-facing person to keep it next to the keyboard.

• Green: public information about the client (website, published accounts, job adverts), your own past proposals with the client names removed, your templates, generic frameworks.
• Amber: anything the client gave you that is not marked confidential, such as an org chart or a process description. Allowed only on a business plan with no training on your content, and only when the engagement letter does not forbid it.
• Red: personal data about the client's staff or customers, unpublished financials, board papers, anything under a non-disclosure agreement that names AI tools or third-party processors, anything the client has told you not to share. Red stays out of every AI tool unless the client has agreed in writing.

Then choose the plan to match. Free and individual consumer accounts can have different data-use settings and retention, so they are green-only tools. A team or business plan from a major provider does not train on your content by default and gives shared projects; both Claude Team and ChatGPT Business need at least two seats, so a three-person firm pays for three. Check each vendor's current privacy page before you buy, because terms and plan names move. If a client contract forbids sub-processors you have not named, a business plan still does not make red material safe; the client's wording wins.

The client-data checklist for consultants goes deeper on what to ask each vendor, and the tutorial on rolling out an AI policy covers how to get a team to follow what you write. For this tutorial, the three lists are enough to start.

Stage one: proposals from your own past wins

A proposal is mostly reusable: problem framing, approach, team, timeline, fees, terms. The slow part is tailoring it to the buyer. Build a library first, and do it once.

Build a clean proposal library

Pick your three best past proposals. Strip client names, figures that identify them, and anything an outsider could trace. Save them as one document called "Proposal library" with a short header on each: sector, size of client, problem, outcome category (cost, speed, risk), fee band. Put that file into a shared project in your business-plan assistant, with a short instruction: "Use these as examples of our house style and structure. Never copy client-specific detail. Ask me for anything you are missing."

Draft from the public brief

Now the daily job. Paste only the request for proposal or the discovery notes that contain nothing red, and ask for a structured draft. A prompt that works:

You are drafting a proposal for our consultancy using the library in this project.
Client brief (public, pasted below): [paste]
Our call notes (no personal data): [paste]
Write: 1) the problem in the client's own words, 2) our proposed approach in three phases,
3) what we need from the client, 4) three risks and how we would handle each,
5) two fee options (fixed fee and time-and-materials) as a table with blanks for numbers.
Do not invent case studies or results. Mark anything you are unsure of with [CHECK].

Illustrative output, trimmed: "Phase 1, weeks 1-2: process mapping across the three fulfilment sites. Phase 2, weeks 3-5: bottleneck analysis and options. Phase 3, weeks 6-7: recommendations and a 90-day plan. Risk: [CHECK] the brief does not say whether site managers will be available for interviews." The [CHECK] flag is the point. You want the assistant to show you its gaps rather than fill them with confident filler.

What you edit every time

Partners should own four things the AI draft never gets right: the price, the scope boundary, the single sentence that shows you understood the buyer's real worry, and the exclusions. In an illustrative run, the draft proposed "weekly steering meetings", which would have added eight unpaid hours. The partner changed it to a fortnightly meeting and moved weekly updates to a short written note. That kind of edit is where a partner's hour is well spent; typing the rest was not.

Time check for the illustrative firm: a proposal that used to take a partner four hours (two of drafting, two of tailoring and formatting) takes about two, with a 15-minute analyst pass to check the [CHECK] items. Six proposals a month saves roughly twelve hours; allow some slippage and call it ten. If you want to benchmark against other practitioners, how consultants write proposals in under an hour shows a faster variant, and mentioning AI use in contracts and proposals gives wording for the disclosure paragraph.

A before and after of one paragraph

Before, written by a tired partner at 10pm: "We will review your operations and recommend improvements." After a prompt that asks for the buyer's own words and a measurable outcome: "You told us last month's late deliveries cost you three customers' repeat orders. In seven weeks we will map the order-to-dispatch process at each site, find where the delay is created, and give you a ranked list of fixes with an estimated effect on on-time delivery." Check that last clause with the client; the AI should never supply the estimate unprompted, and the partner decides how much of a promise it is.

Stage two: research a person can verify

Research tasks suit AI because they are mostly reading and structuring. They are also where fabricated detail hurts most, because a made-up statistic in a client deliverable damages the firm's name. The rule for this stage is simple: every claim that ends up in front of a client carries a source a person has opened.

Use AI to map the question, not to answer it

Start by asking for a research plan, not a result. Example prompt: "We are advising a regional distributor on whether to open a second warehouse. List ten questions an operations adviser should answer before recommending yes or no, grouped by cost, service level and risk. For each, name the kind of source that would settle it (company accounts, industry body data, supplier quotes, interviews)." This gives the analyst a checklist in two minutes. The analyst then goes and gets the sources.

Summarise sources you have already gathered

Once you hold real documents (public reports, regulator publications, vendor white papers), give them to the assistant one at a time and ask for a structured summary: claim, figure, date, page number. Page numbers matter because they let the reviewer verify in seconds. Illustrative output row: "Claim: average pick time fell after layout change. Figure: 14 per cent. Page 22, table 4." The reviewer opens page 22 and either confirms or deletes the row. If the model cannot give a page, treat the claim as unverified.

Many assistants can browse. Use that for discovery, not for facts. A live-search answer still needs the click-through, and the tutorial on researching a client before a discovery call covers a tidy way to do the pre-call version in about 20 minutes.

Competitor and market scans

For landscape work, ask the assistant to build a comparison table from sources you paste in, with an "evidence" column and a "confidence" column. Then run a second prompt: "Which rows in this table rely on a single source? Which could be out of date?" It will not be perfect, but it surfaces the weak rows for the human check, and the analyst then reopens only those sources.

A realistic mistake and how it showed up

Illustrative: an analyst asked an assistant for "typical warehouse labour cost per order in the sector" and pasted the answer into a draft with no source. The figure was plausible and wrong for the client's size band. It was caught only because the partner asked, in review, "Where does this come from?" and the analyst could not answer. The fix was a standing rule: no number goes into a client document unless the source file name and page sit beside it in the working notes. That rule costs a minute per figure and removes the whole class of mistake.

Time check for research

If the analyst spends about 30 hours a month on desk research, a plan-then-verify routine can cut about a third of that: ten hours saved, mainly in reading and note-taking. It will not cut verification time. If your numbers show it is cutting verification, you are saving the wrong hours.

Stage three: delivery, where the client's material lives

Delivery is where most of the value is and most of the risk. Treat it as an amber-to-red zone: use only tools on the plan you vetted, only with data the client has allowed, and always with a person reading the result.

Meeting notes and action lists

Meetings with client staff produce personal data (names, roles, sometimes opinions about colleagues). Before using any transcription or notes tool, tell attendees and get the client's agreement; a one-line statement at the start of the call and a clause in the engagement letter cover it. Then use the tool's summary as a draft. A prompt for the summary:

From this transcript produce: decisions made (with who decided), actions (owner, due date),
open questions, and anything the client asked us to keep confidential.
Quote the speaker's words for each decision in under 20 words. If an owner or date was not stated, write UNASSIGNED.
Do not infer owners.

The word UNASSIGNED does real work. In an illustrative case, an assistant that was not told to leave blanks assigned an action to the finance director who had only been mentioned in passing. The client then chased the wrong person for a week. Telling the model to leave unknowns blank removes most of that.

The weekly status digest

Each Friday the analyst pastes the week's action list and short notes from each workstream (no client personal data) and asks for a one-page status: done, in progress, at risk, needed from the client. Here is an illustrative filled-in output for a seven-week engagement in week four:

Done: process maps for sites A and B signed off.
In progress: site C interviews (3 of 5 held).
At risk: dispatch data extract promised for 3 Nov, not yet received. Impact: bottleneck analysis slips 2 days per day of delay.
Needed from client: named contact for the finance export by Monday.

The risk line is where the partner adds judgement. The AI states the slip; the partner decides whether to raise it as a formal issue. Run the same format for every client so partners learn to read them at a glance. If you juggle several engagements at once, managing projects across several clients with AI covers the tracking layer around this digest.

Drafting the report without losing your voice

Report drafting is the stage where a firm's voice dilutes quickest. The method that holds up: the partner writes the conclusions and recommendations in rough bullet form first, then asks the assistant to turn bullets into paragraphs in the house style from the library, section by section. The model never decides what the firm thinks; it only converts it. Then run a check prompt: "List every factual claim in this draft that is not supported by a source in the notes I pasted." Anything flagged goes back to the analyst.

A walk-through for an operations review

Week five of the illustrative engagement. The analyst has interview notes from five site managers (client staff names removed and replaced with roles). The partner's bullets say: "Site C loses about a day a week to manual re-keying; recommend one scheduled export; saves roughly six hours a week at a cost of one day's set-up." The assistant turns this into a two-paragraph finding. The partner changes "roughly six hours" to "an estimated six hours, based on two interviews" because two interviews are not a measurement, and that honesty is the difference between a finding the client trusts and one the client's finance director picks apart. Time saved on the report: about three hours of drafting on a 20-page document.

Check that it is working after one month

Do not judge by feel. Pick four numbers and look at them after four weeks.

1. Hours per proposal (partner time only). Target: down by a third or more.
2. Edits per proposal: count the number of factual corrections the partner makes. If it is rising, your library or prompt needs work.
3. Unverified claims reaching review: any number in a draft without a source note. Target: zero.
4. Client questions about AI: if clients ask, your disclosure wording needs to be clearer, or you have a client whose red list is stricter than you assumed.

The illustrative firm finds that proposal hours fall from four to just over two, research hours fall by about eight, and delivery admin by about nine, for a total near 28 hours a month. The honest part of that figure is the cost: about three hours a month maintaining the library and prompts, and a standing 20 minutes at each partner meeting to review corrections. Net saving, around 25 hours a month, or about two-thirds of a working week handed back to billable work or to a decent evening.

How it goes wrong in a consultancy specifically

• Confident, wrong benchmarks. Assistants produce industry averages with no source. Never let a benchmark in without the file and page.
• Cloned proposals that quote the wrong client. If a library file still contains a client's name, the model can carry it into the next draft. Search the output for every past client name before sending; a simple find-and-replace check takes a minute.
• Everyone uses a personal account. A partner pastes red material into a personal free plan "just this once". The policy has to name the accounts that are allowed, and the firm should pay for them so nobody has a reason to go round it.
• Over-trusting transcripts. Names, numbers and negatives ("we will not proceed") are where speech-to-text slips. Check those three types against the recording.
• A uniform voice. If every proposal starts to sound the same, buyers who see several consultancies' documents notice. Keep the partner's opening paragraph handwritten.

A small calendar to start next week

Week one: write the green, amber and red lists and choose the plan. Week two: build the proposal library from three past proposals and send one real proposal through the new routine. Week three: add the status digest to one live engagement and use the meeting-summary prompt on one internal meeting before using it with a client. Week four: run the four-number review above and decide what to keep. If you would rather map this against your own engagements with someone, that is what my AI implementation consultation covers.

Questions a consultancy partner usually asks next

Should we tell clients we use AI on their work?

Yes, in plain words, in your terms of engagement or proposal: which tasks use AI, that people review everything, and that client data is not used to train models. Clients with strict rules will often ask anyway, and a sentence written in advance is far easier than an awkward answer mid-project.

Which AI plan suits a three-person firm?

A team or business plan from a major assistant provider, because those plans do not train on your content by default and allow shared projects. Claude Team and ChatGPT Business both start at two seats, and three Standard seats on either cost $75 a month billed monthly or $60 billed annually, as of November 2026. Check the vendor's current terms before you sign.

Can AI replace a junior analyst?

It replaces some of the junior's first-draft hours, not the checking. Someone still has to verify figures, sources and whether the answer fits the client's situation. Firms that cut the checking step are the ones who send a client a confident, wrong slide.

Further reads

• AI Note Taker for Executive Coaches: Keeping Sessions Confidential — Note-taker consent and confidentiality for sensitive sessions.
• Perplexity vs ChatGPT for Business Research — Which assistant to trust for sourced research.
• How to Test a New Service Idea With AI Before You Launch It — Check a new consulting offer before you build it.
• ChatGPT Prompts for Financial Advisers: Follow-Ups and Review Prep — Follow-up and meeting-prep prompts from a neighbouring profession.
• How to Do Competitor Research With AI in One Afternoon — A one-afternoon method for competitor and market scans.
• Claude vs ChatGPT for Consultants: Which to Pay For and How to Test Both — One client deliverable run on both assistants: a 60-page report summary and an options analysis, scored by cost per finished deliverable.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Claude pricing page and Anthropic's privacy article on model training for commercial products; OpenAI's ChatGPT Business pricing page; all read on 4 November 2026. The rest is working method.