How do you build an AI proposal workflow for a marketing agency without overselling the retainer?

An agency proposal carries two commitments a one-off quote doesn't: months of delivery at a fixed monthly fee, and other clients' results used as proof of what you'll do for this one. So the workflow needs two things built before the prompt. A deliverable ledger with internal hours on every line, so you can see whether the scope fits the fee. And a proof file the draft may reproduce word for word and never paraphrase. The prompt itself is the easy part.

TL;DR

• Scope and price are one object in a retainer. Put your internal hours in the source file and make the draft print committed hours next to the fee.
• A deliverable line needs a number and a period to count as scope. "Ongoing SEO support" and "four articles a month" read the same to a model and cost you very different amounts.
• One added deliverable gets paid for every month of the term. In the illustration below, four extra hours a month runs to A$4,560 over a year, which is more than a month of the retainer.
• The case-study section carries legal exposure. The ACCC's guidance for business says a claim about future matters needs reasonable grounds at the time you make it, and that whether you meant to mislead makes no difference.
• Paraphrasing a result turns a record into a forecast. Reproduce proof lines from an approved file with the client, the metric and the measurement period attached to each.
• Speed is why most agencies look at this, and it's the weakest part of the case. The 98% time saving in the vendor comparison I read doesn't survive that vendor's own before-and-after figures.
• Decide the cut list before you're in a three-way pitch. The prose can go fast. The commercial table and the proof lines get a slow read every time.
• Two sign-offs on an agency proposal: whoever owns the price, and whoever has to deliver the scope. In an agency those are usually different people.

What does an agency proposal commit you to that a one-off quote doesn't?

Twelve months of delivery capacity, priced off one conversation.

A quote for a roof replacement names a bounded job, and when the job is done the commitment ends. A retainer proposal names a monthly fee and a list of things you'll do every month until someone gives notice. So the price is the capacity you just sold, and the scope is a claim about what fits inside it.

A language model has no way to know that. To a model, "Scope" and "Investment" are two sections of a document, written in sequence, each one sensible on its own. It'll give you a generous scope and a correct price and hand back a draft where both are wrong together.

The general rules here hold whatever you sell: name the files the drafting step may read, sort them by what the workflow is allowed to do with each, and require a visible marker when a fact is missing. I've set that out in full in how to keep an AI proposal draft inside your approved scope and price, so I'll take it as read and spend this page on the parts that only bite an agency.

Three kinds of fact appear in a retainer proposal and never in a trade quote. The term and the notice period. The deliverable counts per month. And the exclusions: ad spend, third-party licences, extra revision rounds, work outside the named channels. All three are commitments, and all three are easy for a draft to soften. The exclusions are the ones to watch, because a softened exclusion costs you every month instead of once. The ACCC's guidance for business names that directly, telling businesses not to make false claims about "any exclusions on the goods and services" 1. The same guidance holds that staying quiet can mislead on its own: "In some circumstances, failure to disclose information can be misleading." 2 Leave an exclusion out of the draft and you've agreed to the work.

How do you stop the draft adding a deliverable?

Give it a closed ledger to choose from, and make selection the only thing it can do with scope.

The source file most agencies point a drafting step at is a services page or last quarter's proposal. Both are prose, and prose invites composition. A model reading "we offer SEO, content and paid media" will compose a scope from those words, sized to whatever the discovery notes suggest the prospect wants. You get a proposal that sounds exactly like your agency and commits you to a volume nobody costed.

Replace the prose with a ledger. One row per deliverable, five columns:

Then two instructions. The draft may only use deliverable rows from the ledger, reproducing the wording exactly. And the output must carry a total of committed internal hours for the month, printed next to the fee. The prompt frame around those instructions is settled, and the five inputs it needs are Context, Task, Source, Rules and Output, so this is one line in Output and two in Rules.

That hours total is the part almost nobody builds, and it's the check that does the work. The figures below are mine, chosen to show the shape.

Say the retainer is A$4,000 a month excluding GST and your blended internal delivery cost is A$95 an hour. The scope you costed is four articles at three hours each, three hours of reporting, a full day of technical work and two hours of calls. That's 21 hours, or A$1,995 of delivery cost, which leaves a delivery margin just over 50%.

Now the draft adds a helpful line about social content at a full day a month. Committed hours go to 25 and delivery cost to A$2,375, so the margin falls to about 41%. In one month that reads as survivable. Across a twelve-month term it's 48 hours and A$4,560, more than a month of the retainer, on a sentence nobody priced and nobody noticed.

With the total printed beside the fee you see 25 hours against a scope you costed at 21, and you catch it in a second. Without it you're reading five pages of deliverables hunting for a line you don't remember approving.

Test it on a retainer you already sold, because you know the approved scope and the fee, which makes the hours number a marking key. One of the vendor guides I read gives the same advice about trialling software, which is to bring your real proposal template instead of a demo file 3, and it applies just as well to trialling your own prompt. A workflow proved on a tidy example meets your real notes for the first time on a live pitch.

Reusing another client's results is the riskiest section in the document, and the risk is legal

Every agency proposal borrows proof. Someone else's traffic graph, someone else's cost per lead, your own team size and years trading. Hand that library to a model and it will do four things with it. Each one reads exactly like the sentences that are true, which is why a proofread catches none of them.

It transfers. A result measured for one client gets attached to a prospect in the same vertical, or a specific outcome becomes a general capability. Say your file holds "we grew organic sessions 40% for a Brisbane equipment hire business". That turns into "we grow organic sessions around 40% for equipment hire businesses". The first is a record. The second is a forecast, and the ACCC is explicit about forecasts: "A business that makes a claim about future matters (including predictions or projections) must have reasonable grounds for making the claim at the time of making the claim. The business is responsible for showing that it had reasonable grounds to make the claim." 4

It blends. Two case studies merge into one composite client carrying the better number from each. That client doesn't exist, so when the prospect asks for a reference you have nobody to offer.

It identifies. Your anonymised case study comes back as something like "a Sunshine Coast dental practice turning over about $2 million", which is enough for anyone in that market to name them. Apply a guess test to every anonymised line: if someone inside the vertical could work out who it is, it isn't anonymous.

It rounds. Sixty-three clients becomes "over 100 businesses". Four years becomes "a decade of experience". A lapsed platform badge comes back as a current partnership.

None of that requires bad intent, and intent is where agency owners misread the exposure. The ACCC's position runs one sentence: "It makes no difference whether a business intends to mislead or not." 5 The same guidance says any claim a business makes about its services must be "accurate, truthful and based on reasonable grounds" 6, that the business has to be able to prove it 7, and that the rule covers a quotation and a testimonial alongside an advertisement 8. A proposal is a communication from your business. A model drafting the sentence changes nothing, because your name is on the document. The ACCC publishes guidance and says plainly that it doesn't give legal advice 9, so get your own on any wording you're unsure about.

Build the proof file the way you built the ledger. One row per case study, carrying the client, the metric, the number, the measurement period, and a permission level: named with written permission, anonymous by vertical and size, or internal only. Then the tightest rule in the whole workflow. The draft may select a proof row and reproduce it, and it may not paraphrase it, combine two of them, or restate a number in a sentence of its own. Paraphrase is the step where a record becomes a claim.

Credentials get their own file with a date on it: client count, team size, years trading, current certifications, one owner, reviewed quarterly. These are the numbers your team quotes from memory, and memory rounds up.

While you're in there, settle what your people may paste into a tool in the first place. Client material in a proposal draft is client material leaving your building, and a written policy for what your team puts into an AI tool answers that before someone asks about it in a group chat.

What should the workflow do when three agencies are pitching and the deadline is Friday?

Decide the cut list now, because Thursday night is a bad time to be making that call.

Speed is why most agencies look at this at all. LLeMental, which sells proposal software, puts it plainly in its own agency comparison: "the difference between winning and losing clients is often how fast and how well you can respond" 10, and it observes that agencies often create multiple proposals weekly 11. Agenticise, a UK automation consultancy, ranks proposal generation second of eight agency workflows, claiming 15 to 20 hours saved per pitch against 8 to 15 days of setup, and noting that voice and strategy "must stay senior-led" 12.

Read those numbers as what they are. Both come from businesses selling the capability, and one of them falls over on inspection. The LLeMental page claims a "98% reduction in creation time (from average 74 minutes to under 5 minutes)" 13 and then, in its own FAQ, "up to 98% time reduction on template-based proposal workflows, bringing a 90-minute proposal down to under 5 minutes" 14. Do the arithmetic. Seventy-four minutes down to five is a 93% reduction. Ninety minutes down to five is 94%. Reaching 98% from 74 minutes means finishing in about a minute and a half. Two different baselines, one headline percentage, and neither pair of figures produces it.

I use a proposal workflow in my own business, so I'm not dismissing the saving. The figure you should care about is your own, measured on your own pitches, which is its own small job: measuring what an AI workflow actually saves you stops you defending a vendor's percentage in your own management meeting.

The harder question is what the saved time buys. If three agencies are pitching, all three can now produce a well-formatted document by Friday. Speed stops separating anyone the moment everyone has it. What still separates the three documents is whether the diagnosis reflects something the prospect said out loud on the call. A model can restate what's in your notes. It can't hear what nobody wrote down.

So agree the list before you need it.

• Goes fast, every time: section order, headings, the summary, the narrative of how you work, the selection of which case studies to include.
• Gets a slow read from a person: the deliverable rows and their hours, the fee, the term and notice period, the exclusions, and any sentence carrying another client's number.
• Gets the time you saved: one paragraph of the diagnosis that only you could write, because you were on the call.

The diagnosis is where an agency pitch is won, and it runs on evidence you gather before the prompt

The section a prospect reads closest is the one telling them what's wrong with their marketing, and it's the section AI drafting helps with least.

Give a model the prospect's industry and a page of call notes and you'll get a competent diagnosis any agency in the country could have sent. A diagnosis is a decision made from evidence, and with no evidence in front of it there's nothing to decide from. The constraint was always the evidence.

In my own business, Searchmaxxed, we built AI workflows for creating blog posts, doing SEO and building custom proposals. The proposal one sits downstream of something I built first. I described it at the time as a system "that connects all of your data about your search presence online into one process that is then able to make recommendations on what we should do and improve it". Competitor keywords, rankings, AI mentions, backlinks and conversion data, pulled into one place so the next action comes out of what the market shows. The same evidence points outside the website too, which is how you work out "what the best places are to get placements, maybe press releases, maybe Reddit comments".

Once that assembly exists, the diagnosis section of a proposal takes twenty minutes, because the thinking already happened and the draft is reporting a decision somebody made. Without it you're asking a model to invent a position. It will, and it'll read like everyone else's.

What transfers is the ordering. Your evidence assembly will look nothing like mine, because you sell something else, and the point is that it exists before the prompt does. If search is what you sell, the AI workflows behind the search work itself come before the proposal workflow, because they produce what the proposal has to report.

Omid Ghiam at Marketer Milk, who spent three years working this out, puts the underlying rule better than I can: "you can't automate what you can't teach someone else to do" 15. If your diagnosis isn't a repeatable decision from named inputs, the drafting step has nothing to draft from. Agenticise reaches the same conclusion about process generally: "Automating a broken process makes the broken bit faster, not better." 16

Who has to sign a retainer proposal before it goes out?

Two people: whoever owns the price, and whoever has to deliver the scope.

In most agencies those are different people, and the second signature is the one that gets skipped. The owner or strategy lead runs the pitch, writes the diagnosis, sets the fee and sends it. The delivery team meets the scope for the first time at kickoff, by which point the deliverable count is contractual. Agenticise keeps the review with the pitch author, where "the strategy lead owns the narrative, the rates, and the personalisation paragraphs" 17. On a retainer I'd add the delivery owner on one question only: does the committed-hours total fit what we sold? They're the only person who knows.

The quieter failure is drift. You write a good prompt, share it with the team, and a month later four versions of your own commercial position are circulating, because someone added a helpful line about payment terms and someone else pointed the source at a folder instead of a file. Fix that with ownership. One ledger file, one proof file, one credentials file, each with a named owner and a review date. Agenticise makes the point about tone of voice, that it "is a system, not a single prompt" 18, and your commercial terms deserve the same treatment. If you want the sign-off to survive contact with a deadline, build the approval step into the workflow so a draft can't reach a prospect without passing it.

Building that with your own material on the screen is what a session does. Private onsite training happens at your workplace, scoped around your real ledger, your real rate card and the proposals you're actually sending, and private AI training for your team suits an agency where the rules have to hold across several people. It's enquiry-only with a fixed quote after we scope it, usually somewhere between 4 and 20 people, at your own office.

The public workshop is the other route. A full day in person at Avani Mooloolaba Beach Hotel, The Cove room, in Mooloolaba, at A$495 excluding GST per person, capped at 15 paid attendees, with the Wednesday 7 October 2026. Bring a laptop, a charger and the jobs your business repeats with a safe example. Your proposal drafting step is a good candidate, and you leave with the workflow you built and tested during the session. Both routes teach the same build, and AI workshops for marketing agencies sets out how we scope it for an agency team.

One distinction decides which conversation to have. Training and implementation are separate engagements with different scope and price. Training teaches your team to build the workflow and run it themselves, which is what a session does. Implementation means the system gets built and connected for you. Both are available from us, so name the outcome you want in your enquiry and we'll scope from there.

Frequently asked questions