How Small Agencies Use AI to Write New-Business Pitches

Small agencies use AI at five points in a pitch: scoring whether the opportunity is worth chasing, researching the prospect, stress-testing the strategic idea, matching case studies from a credentials library, and drafting and rehearsing the document. The insight, the price and every claim about results stay human. Done well, a 20-hour pitch can come down to around 12.

The trap is thinking AI can supply the idea. Every agency on the shortlist has the same assistants, and the prospect has probably asked one about their own market already. A pitch built from AI research alone tells them what they know. The part that wins is the observation nobody else has: something you saw in their shop, heard from their customers or found in their own numbers. AI's real job is clearing the rest of the work so you have time to go and find that.

Follow me on Instagram@sagnikteaches

Decide whether to pitch at all: a go/no-go score

Unwinnable pitches are the most expensive thing a small agency does. Score every opportunity before anyone opens a deck. Paste the brief into a project and ask for a structured assessment against your own criteria:

Connect on LinkedInSagnik Bhattacharya

Score this opportunity for our agency from 1 to 5 on each criterion,
quoting the part of the brief that supports each score. Where the brief
doesn't say, score it "?" and write the question we should ask.
1. Fit with our strengths: [food and drink brands, e-commerce, email]
2. Budget stated and realistic for the scope
3. Decision-maker identified and reachable before the pitch
4. Number of agencies pitching (fewer than 4 is good)
5. Timeline we can meet without dropping client work
6. Chance of a long relationship, not a one-off project

An illustrative result for a farm shop inviting agencies to pitch for the launch of a home delivery service:

Subscribe on YouTube@codingliquids

The question marks are the useful output. One phone call answered both: the owners would attend, and three agencies were pitching. The agency went ahead. Set your own threshold, for instance no pitch below 18 out of 30 unless there's a strategic reason.

The threshold earns its keep through simple arithmetic. Take an illustrative three-person agency that values its own time at $75 an hour. A 20-hour pitch costs $1,500 of time. If it wins one pitch in four, each new client has cost $6,000 in pitching before any paid work starts. Cutting every pitch to 12 hours brings that to $3,600 per win, but turning down the two weakest invitations each quarter saves $3,000 of time outright, and often does more for the win rate than any drafting shortcut, because the hours go into pitches the agency can actually win.

One thing the score won't catch unless you ask: a brief that wants finished work up front, such as "present three campaign concepts with sample posts". Add a seventh line to the prompt ("Does the brief ask for unpaid creative or strategy work? Quote the sentence.") and settle your policy before you read the answer. Some agencies offer a paid discovery session instead; others present their thinking and approach but no finished creative. The model's job is only to find the sentence. Whether to walk away is the agency's call.

Research the prospect in two hours, not two days

Use a research tool that shows its sources, such as the deep-research modes in the main assistants or Perplexity (Pro lists at $20 a month), and ask for a brief with every claim cited. The differences between the main options are covered in Perplexity versus ChatGPT for business research.

Research [prospect] for a new-business pitch. Cover: what they sell and
to whom; how they describe themselves; recent news; reviews (themes, not
quotes out of context); their social channels and posting frequency;
three nearest competitors and how those competitors position themselves.
Cite a source for every point. Mark anything older than 12 months.
End with "Things to verify in person".

Then check the sources. Research tools confuse businesses with similar names, quote reviews from years ago as current, and summarise a competitor's old website. The routine in checking sources and citations in AI research takes 20 minutes and catches most of it. For the farm shop, the research brief claimed the shop "already offers delivery via a third-party app". The cited page was a two-year-old listing for a different farm shop with a similar name. Had it gone into the pitch, the first slide would have been wrong about the client.

Then do the part AI can't. Visit. Buy something. Read the notices by the till. At the farm shop, the most useful finding came from the counter staff: regular customers already phoned in orders for collection, and the phone line was jammed on Friday mornings. No research tool would have turned that up.

The prompt needs changing when the prospect barely exists online. Suppose the next invitation comes from a family-run joinery firm with a four-page website, 11 reviews and no social accounts. The standard brief comes back thin and padded with general remarks about the joinery trade, which is worse than useless because it reads as research. Point the tool at the market instead: "List the five joinery firms that appear first for 'bespoke fitted kitchens' in the prospect's service area. For each, quote the promise on their home page and the price signals they give." The gap in that list (nobody showing finished kitchens with prices, say) becomes the starting point, and the 11 reviews get read in full by a person, because with so few, each one matters.

Stress-test the idea before the prospect does

Once you have a strategic idea, use AI as a sparring partner. Two prompts do most of the work:

1. Here is our pitch idea: [idea]. Argue against it as the prospect's
   finance-minded owner would. Give the three strongest objections.

2. Three other agencies are pitching. Describe the most likely idea each
   would bring, based on the brief alone. How is ours different?

For the farm shop, the idea was to launch delivery first to the customers who already phone in orders, turning the jammed Friday phone line into the launch list. An illustrative first objection from the model: "Phone customers may be older and less likely to order online; the launch could underperform with its most loyal group." That is a fair point, and it changed the plan: phone orders stay, with staff taking them into the new system for the first month. The second prompt predicted the obvious rival idea, a social media campaign with a launch discount, which confirmed that leading with the phone-customer insight was distinctive.

Match proof from a tagged credentials library

Most small agencies rebuild credentials from old decks for every pitch. Instead, keep a library: one short entry per piece of past work, tagged so AI can select from it. A filled-in entry:

CASE: subscription launch, specialty coffee roaster (2025)
TAGS: food and drink; e-commerce; email; launch; subscriptions
PROBLEM: one-off online orders, no repeat purchase programme
WHAT WE DID: subscription offer, 4-email welcome series, launch social
RESULT (approved wording): "Subscriptions became 30% of online revenue
  within six months" (client approved, 12 Jan)
CAN WE NAME THE CLIENT? Yes, with logo
CONTACT FOR REFERENCE: yes, the co-owner

Building the library is a one-off job that AI speeds up. Upload three or four old pitch decks and ask for one entry per case in the format above, with "RESULT (approved wording)" left blank and a note of where each claim appeared. Then someone checks each result against the client's actual sign-off, an old email or a report, and fills in the approved line. Cases with no approved result stay in the library but can't be quoted with numbers. Expect a small agency's back catalogue to yield 15 to 25 usable entries in a day.

Ask the model to list every version of a claim it finds, not just one. An illustrative extract from that first pass:

CASE: loyalty email programme, independent bakery group (2024)
RESULT (approved wording): [blank - to check]
CLAIMS FOUND:
  "repeat orders up 40%" (2024 pitch deck, slide 9)
  "repeat orders up 25% in the first quarter" (2025 deck, slide 6)

Two decks, two numbers for the same work. When the account lead dug out the client's quarterly report, 40% turned out to be the target set at kick-off and 25% the result. The older deck had been quoting a forecast as an outcome for a year. Without the "every version" instruction, the model would have picked one figure at random and the conflict would never have surfaced.

Store the library in the pitch project and ask: "Choose the two cases most relevant to this brief and explain why, using only the approved result wording." The "approved wording" rule is essential. Models round up, merge results from two cases and turn "30% of online revenue" into "grew revenue by 30%", which is a different and false claim. Catching made-up figures in AI-drafted proposals covers the checking habit in detail.

Draft the document, then rehearse against a sceptical buyer

With the idea, research and proof agreed, drafting is quick. Keep the structure simple: what we heard, what we noticed, what we'd do, how we'd measure it, who'd do it, what it costs. Ask for a draft section by section rather than all at once, and write the opening page yourself; it's the page that decides whether the rest gets read. For slides, Copilot in PowerPoint (part of Microsoft 365 Copilot Business, $21 per user per month on annual billing) and Gemini in Slides (included in Google Workspace Business Standard and above) can turn an approved outline into a first deck, which a designer then fixes.

The opening page is where the difference between AI research and field research shows most. A before and after for the farm shop:

AI-drafted opening: "Home delivery is growing fast, and independent retailers must adapt to changing customer expectations. Our agency brings proven e-commerce expertise to help your farm shop compete and thrive."

Rewritten by the agency lead: "On Friday morning your phone rang 23 times in the hour we were in the shop, and most calls were regulars placing orders for collection. Your customers have already built a delivery service. They're just doing it by phone. Our plan starts with them."

The first version could open any agency's pitch to any retailer. The second could only be written by someone who stood by the till, and the owners will remember it. The count of calls is the agency's own observation from the visit, which is exactly the kind of fact a pitch should rest on.

Then rehearse. The prompt:

You are the owner of [prospect], sceptical of agencies after a bad
experience. I will present our pitch. Interrupt with the questions you
would really ask, one at a time. Be direct about cost and results.

An illustrative exchange from the farm shop rehearsal:

"Owner": You keep saying email. Half my customers are over 60. Why would they open your emails?

Agency: They already ring us on Fridays, so the first month runs through the phone as well as email.

"Owner": And who pays for the staff time on the phone while you're learning?

That second question wasn't in the deck. It became a slide on how phone orders would move online gradually, with the staff-time cost estimated. Rehearsing with the model is no substitute for rehearsing with a colleague, but it surfaces the awkward questions early.

After the meeting: follow up within a day

Pitches are often decided in the days after the meeting, and small agencies lose ground by following up slowly. Record the meeting if the prospect agrees, or take notes, and within a few hours ask the assistant to turn them into three things: a thank-you email that answers any question you couldn't answer in the room, a list of concerns the prospect raised, and your internal actions. Check the email against the notes before sending; models sometimes promise things in a follow-up that nobody offered in the meeting, such as a free first month. A short, specific email the next morning ("You asked how phone orders would move online; here's the three-step plan and the staff time we estimated") usually does more than a polished recap a week later.

A before and after from the farm shop follow-up shows how easily the extra promises creep in:

Illustrative AI draft: "Thank you for your time yesterday. As discussed, we'd be happy to include the first month of social media management at no extra cost, and we can have home delivery live within four weeks."

Sent version: "Thank you for yesterday. You asked whether social media is included: it isn't part of the launch fee, and a separate quote is attached. You also asked about staff time on the phones. We estimate about three hours a week in the first month, falling as regulars move to ordering online."

Neither promise in the draft was made in the room. The notes showed the owner asking whether social was included and the agency saying it would be quoted separately, and the deck's launch plan ran to eight weeks, not four. A quick check catches this every time: paste the draft and the notes together and ask, "List every commitment in this email and the line in the notes that supports it. Mark any with no support." Anything marked comes out before sending.

The farm shop pitch in hours

An illustrative three-person agency's time on the pitch, before and after building the library and prompts:

The savings come from research, credentials and drafting. The visit and the strategy barely shrank, and they shouldn't: that's where the pitch was won. Setting up the credentials library takes a one-off day, so the savings start from the second pitch.

What prospects notice in AI-written pitches

• Generic openings. "In a competitive market, standing out has never been more important" tells the owner nobody looked at their business. Open with something you saw.
• Facts about their business that are wrong. A wrong founding year or a product they stopped selling is fatal, because they know their own business better than anyone. Verify every statement about the prospect.
• Inflated or merged results. Use approved wording only, and keep the source approval in the library entry.
• Market statistics with no source. If you can't name where a figure came from, cut it. Prospects sometimes ask.
• A plan that fits any client. If the three-month plan would work unchanged for a bakery or a bike shop, it isn't a plan for this prospect.

That last test can be run literally. Paste the draft and ask: "Replace the farm shop with a bike shop throughout. List every sentence that still reads as true and sensible." In an illustrative run on the agency's first full draft, 14 of 22 sentences in the strategy section survived the swap, including every line under "Our approach". After the rewrite built around the Friday phone orders, four survived, and all four were in the team bios, which is the one place generic sentences belong.

Keep a pitch log to see whether the AI routine is helping

Hours saved mean little if the win rate drops, so log every pitch in a simple sheet: the go/no-go score, hours spent, the result and the reason the prospect gave. An illustrative log for the three-person agency's first six pitches after setting up the routine:

Two things are worth reading from a log like this. Hours are down and three of five pitches won, so the routine is working. And the one loss for being "general" was the only pitch where nobody visited or spoke to the prospect's customers, because the team was busy and the AI research looked thorough enough. That single row is the strongest argument for keeping the field research step, whatever else gets automated.

Proposals that follow a pitch have their own common errors; writing business proposals faster with AI covers the follow-up document.

Pitching questions from small agencies

Is it safe to upload a prospect's brief or RFP to an AI tool?

Check whether you signed a confidentiality agreement first; many briefs come with one. Use a business plan where content isn't used for training by default, keep the brief inside a project the pitch team can see, and delete it if you don't win and the agreement asks you to. Never paste a prospect's documents into a personal consumer account.

Should we tell a prospect we used AI to prepare the pitch?

You don't need to label every page, but be ready to answer honestly if asked, and make sure your agency's general approach to AI is easy to explain. Prospects increasingly ask how an agency uses AI on client work, so a short, clear answer in the pitch itself can count in your favour.

Can AI write our pricing section?

It can lay out a pricing page you've already decided, and it can check that the scope and the price match. It shouldn't set the price. Your costs, your capacity and how much you want the account are judgements only the agency can make, and a model will happily produce a plausible number with no basis at all.

How do we stop every pitch sounding the same?

Start each pitch from the prospect's material and your field research, not from last month's pitch document. Keep reusable parts, such as team bios and case studies, in a library, but write the opening, the insight and the plan fresh. A quick check: if the first page could be sent to a different prospect with the name changed, rewrite it.

Further reads

• Can AI Help a Small Business Respond to Tenders and RFPs? — Formal tenders need a stricter process than a creative pitch.
• How to Write Case Studies From Customer Interviews Using AI — Build the case studies your credentials library needs.
• Copilot in Word: Draft Proposals From Your Own Files — Draft the pitch document from your own files in Word.
• AI Proposal Mistakes That Cost You the Job — Errors that cost agencies the work, and how to avoid them.
• How Consultants Use AI to Write Proposals in Under an Hour — A faster proposal routine for smaller opportunities.
• AI Rollout Plan for a Ten-Person Marketing Agency — Fit pitching into a wider agency AI plan.
• AI for Recruitment Business Development: Winning New Clients — A weekly business-development routine for agencies: hiring signals, five-minute account briefs, outreach that names the vacancy, and consented spec profiles.
• AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals — Clause-by-clause wording for agencies using AI, with a worked campaign example and a prompt for spotting AI conflicts in a client's own contract.
• Which Agency Tasks AI Handles Well, and Which It Doesn't — Sort agency work by evidence, review effort and consequences, then test a reporting workflow with a clear human sign-off.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Perplexity, ChatGPT Business and Microsoft 365 Copilot Business pricing pages; Google Workspace plan pages (Gemini in Slides and Docs).