AI Media Planning for Small Agencies: Channel Mix and Budget Splits

Yes, as a first draft with its assumptions showing. Give the AI the goal, budget, audience, dates and the client's past results, ask for the formula behind every split, then check four things: the splits add up, each channel can deliver enough conversions to optimise, every benchmark has a named source, and a platform planner roughly agrees.

On its own, a chat assistant cannot see auction prices, the client's account or this quarter's competition. Left alone, it fills the gaps with benchmark-sounding numbers, so treat its output as a list of assumptions to confirm, not a forecast. The checks take about an hour and catch the errors that cost real money: a split that does not add up, a channel that cannot spend its share, a cost per enquiry borrowed from another industry.

Follow me on Instagram@sagnikteaches

What the AI needs on one page, and what it must not guess

Asking for "a media plan for $30,000" returns a generic spread across every channel the model has heard of. A one-page input sheet returns a plan you can check. It has seven boxes, and the one that tends to get skipped is history, because history is what stops the AI inventing costs. An illustrative sheet for an interior design studio, the example client throughout:

Connect on LinkedInSagnik Bhattacharya

GOAL: 300 qualified enquiries in 12 weeks (an enquiry is a form or a call of 60 seconds or more from someone planning a project within a year)
BUDGET: $30,000 media over 12 weeks, $2,500 a week. Agency fee and creative production are separate and not in this number.
AUDIENCE: homeowners planning a full-room or whole-home redesign in the next 6 to 12 months; household income above [threshold the client sets]
TIMING: starts [date], ends 12 weeks later; busiest enquiry months are [months]; the studio can take [number] new projects this quarter
HISTORY (last 12 months): Search spend $38,000, 580 enquiries ($65 each). Meta spend $12,000, 125 enquiries ($96 each). Pinterest: never used.
CONSTRAINTS: no TV or radio; before-and-after images must be the studio's own; one landing page for Search, one for social
MEASUREMENT: enquiries are logged in the CRM with a source field; form and call tracking are live

Two lines carry the weight. The definition of an enquiry stops "lead" meaning one thing in the platform and another in the CRM, and the history line gives the AI real costs to start from. If a channel has no history, say so. The prompt below treats that as a test, not a plan. For the tracking that makes the measurement line true, see set up conversion tracking before you let AI spend your ad budget.

Subscribe on YouTube@sagnikteaches

A prompt that returns assumptions in a table

You are a media planner drafting a first-pass plan for a small agency. Use the INPUT SHEET below.

Return:
1. A table with one row per channel: channel, role (capture demand or create it), budget in dollars and as a percentage, expected cost per enquiry, expected enquiries, SOURCE of the cost figure, and CONFIDENCE (high, medium or low).
2. SOURCE must be one of: CLIENT DATA, PLATFORM PLANNER, NAMED BENCHMARK (give the name, year and page), or GUESS. If the input sheet has history for a channel, use it. Never invent a figure and label it CLIENT DATA.
3. A line under the table showing that the percentages sum to 100 and the dollars sum to the budget.
4. A phase plan for the 12 weeks with dollars per channel per phase.
5. The three assumptions that would change the plan most if wrong, and what I should check for each.

Use no more than four channels. Do not recommend a channel with no history unless you label it a TEST with a cap of 20% of the budget.

INPUT SHEET: [paste]

Three rules in that prompt matter most. The SOURCE labels make the AI say where each cost comes from, and the checks below test those labels. "Never invent a figure and label it CLIENT DATA" blocks the commonest pretence. And the four-channel limit with a 20% cap on untested channels prevents a spread so thin that no channel can learn. The 20% cap and the four-channel ceiling are my rules of thumb, so set your own, but set them before the AI drafts, not after.

A first draft for an interior design studio, and where it fails

Run on that sheet, an illustrative first draft came back like this:

It reads confidently and fails three ways. First, the shares add up to 105% and the dollars to $31,500, which is $1,500 more than the budget; the AI's own sum line either said nothing or said 100. Second, the blended cost per enquiry is $22, under a third of the client's own history. That history is $50,000 of spend for 705 enquiries, so $71. A plan promising 1,401 enquiries against a goal of 300 should have set off an alarm before anyone checked the maths. Third, two of the source labels are false. The AI cannot run a platform planner, so "platform planner" for Meta is an invention, and the 2026 search benchmark it names for Search puts the cross-industry cost per lead at $66.69, not $31. Only the Pinterest row is honest, and the label there is "guess".

The corrected plan and the five checks

Rerun the prompt with a stricter instruction ("use the client's history for every channel that has it") and fix the sum, and the plan becomes:

The prompt's fifth item gives you the checking agenda. An illustrative answer to it:

1. "Pinterest's $140 cost per enquiry is a guess, because the client has never used the channel." Check: run it as a capped test in the first phase and read the real cost per enquiry at week four.
2. "The Search keyword set can absorb $13,500 over 12 weeks." Check: run a Keyword Planner forecast at the planned bid, as in check 5 below.
3. "Enquiries logged in the CRM match the studio's definition across every platform." Check: in week two, sample 20 enquiries from each channel and count how many meet the 60-second or project-within-a-year test.

Those three become the to-do list for the first four weeks, and the checks below test the rest of the table.

Check 1: do the splits add up?

Add the percentages and add the dollars yourself, in a spreadsheet, every time. The corrected plan gives 45 + 35 + 20 = 100 and 13,500 + 10,500 + 6,000 = 30,000. Divide each channel's budget by its cost per enquiry to confirm the enquiries column: 13,500 ÷ 65 is 208, 10,500 ÷ 96 is 109 and 6,000 ÷ 140 is 43. This check takes two minutes and catches errors that language models make routinely, because a model writes plausible-looking sums unless it is running code you can inspect.

Check 2: can each channel spend and learn?

A budget the platform cannot turn into enough conversions is a budget that cannot optimise. Meta's $10,500 over 12 weeks is $875 a week, which at $96 an enquiry buys about nine enquiries a week. Meta's learning-phase help page says an ad set usually leaves the learning phase after about 50 results in the week after its last significant edit, and an ad set that cannot get enough results shows as learning limited. Nine a week is far short of 50, so plan for one ad set, few creative changes and a learning-limited status; testing Meta Advantage+ creative without wasting budget covers working inside that limit. Search at $13,500 is $1,125 a week, about 17 enquiries a week at $65, which is healthier but still modest.

Check 3: is the implied cost believable?

Compare the plan's blended cost per enquiry with the client's history. The corrected plan's $83 is 117% of the $71 history, which is plausible because a guessed channel carries more cost. The AI's first draft was 32% of it. My rule of thumb is to ask for a written reason whenever a plan's blended cost sits below about 60% of the client's own history, because a good reason (a new offer, a conversion-rate fix already live) exists in only a minority of cases.

Check 4: are the labels honest?

For each cost figure, open the thing it claims to come from. A CLIENT DATA label must match the history on the input sheet. A PLATFORM PLANNER label must correspond to a forecast you ran. A NAMED BENCHMARK label must point to a page you can open, showing the year, platform and metric. Then count the budget resting on guesses: here it is Pinterest at 20%, exactly at the cap. If guessed channels carry more than a fifth of the budget, either cut them or call the plan a test.

Check 5: does the planner agree?

For Search, run a Keyword Planner forecast on the keyword set at the planned bid and daily budget. Google describes these forecasts as estimates that use data such as historical quality score and click-through rate to simulate expected performance, and the API documentation I read gives no error margin, so use the result as a range. The plan implies 3,750 clicks. Taking the client's Search conversion rate as 5.5% (illustrative), $65 per enquiry means an average cost per click of about $3.58, which I round to $3.60, and 13,500 ÷ $3.60 is 3,750. Suppose the forecast range is 3,300 to 4,100 clicks. It brackets the plan, so Search passes. Had it said 2,700, Search could only absorb about $9,720 (2,700 × $3.60), and $3,780 of the Search budget would belong elsewhere, probably in a test channel past its 20% cap. That is precisely the finding you want before the client sees the plan.

Benchmarks: which sources to trust, and how to read them

The note on every AI-drafted plan should be: no number without a source and a date. Rank the sources you have:

An example of reading a published benchmark properly. LocaliQ's 2026 search advertising benchmarks, produced with WordStream and updated on 1 June 2026, break results out across 23 categories. There is no interior design row. The nearest are Home & Home Improvement (cost per lead $90.92, conversion rate 8.05%, cost per click $8.33) and Furniture (cost per lead $106.70, conversion rate 2.99%, cost per click $3.97), and across all industries the average cost per lead is $66.69. The page says the data comes from thousands of the vendor's customers' campaigns on Google Ads and Microsoft Ads and gives no exact sample size, so treat the figures as a band, and remember that a vendor's customers are not a neutral sample. Here the client's own $65 sits below both nearest rows and close to the all-industry average, which says the client's Search account is doing well, not that the benchmark is wrong. The AI's $31 was less than half of every one of those figures, which is the kind of gap that should always have a stated reason.

Stress-testing before the client sees the plan

A plan should survive a bad month. Two quick stresses on the corrected plan. First, remove the guess: if Pinterest delivers nothing, the plan produces 208 + 109 = 317 enquiries, still above the 300 goal. Second, make it worse: if Meta's cost per enquiry runs 20% above history ($115 instead of $96, so about 91 enquiries) and Pinterest delivers nothing, the plan produces 208 + 91 = 299, a single enquiry short of the goal. That tells you the plan has little margin if two things go wrong together, and it is the reason the next section contains a decision gate. Put both results on the client page next to the base case of 360, so that nobody is surprised in week six.

Flighting: where the money goes, phase by phase

A plan can run continuously, in flights (bursts with gaps) or pulsed (a steady base with bursts on top). For a 12-week enquiry goal the simplest useful pattern is three phases of four weeks with the test channel tapering. In the illustration, Search stays level because it captures demand that already exists, Meta is front-loaded to build a pool of people who have seen the work, and Pinterest tapers after a week-four gate.

The gate sits at the end of week four. If Pinterest's cost per enquiry is above $180, which is 1.3 times the $140 guess and rounded down, stop it, and move the remaining $3,000 into Meta, adding $1,500 to each of the last two phases. Write that rule into the plan before launch, because the person who spent the first month defending a channel is the worst judge of it afterwards. Meta's learning-phase page also warns that significant edits, frequent budget changes included, can send an ad set back into learning, so make Meta's week-five step down in one change rather than a run of small cuts. Each column and row total should be re-added by hand, since a phase table is exactly where a language model drops a few hundred dollars.

A second pattern suits a seasonal client. Illustratively, a signage company whose orders peak around two known periods gets $12,000 across 26 weeks as two three-week bursts at $1,000 a week ($6,000) and 20 base weeks at $300 ($6,000), which sums to $12,000. A flat spread would be $12,000 ÷ 26, about $462 a week, and it would be thin exactly when demand is highest and generous when nobody is searching. Confirm the peaks in the client's own enquiry history before trusting them, and label them as an assumption until you have. Once any plan is live, the routine in AI for advertising agencies: media plans and weekly budget checks keeps spend on pace.

A reporting plan that says when to move money

A media plan without a reporting plan is a wish. Every metric below has a source of truth and a decision attached to it. For the wider routine of client reporting, see how marketing agencies use AI to automate client reporting.

The account lead reads this in about fifteen minutes each week, and the client sees one page each month. The CRM is the source of truth for enquiries because platforms count conversions their own way, and the two will disagree. Record the gap in the report instead of reconciling it away.

Rehearse the client's questions before the meeting

Use the AI a second time, as a sceptic. The prompt: "Act as the studio owner's accountant reading this plan for the first time. Ask the five hardest questions about the budget and the assumptions. For each, name the line of the plan that answers it, or write NO ANSWER if nothing does." An illustrative result:

1. Why is Pinterest 20% of the budget when we have never used it? Answer in the plan: the 20% cap and the week-four gate.
2. What happens if enquiries come in under 300? Answer in the plan: the stress case of 299 and the gate rule.
3. How do we know the CRM counts are right? Answer in the plan: the 20-enquiry sample per channel in week two.
4. Does the agency fee come out of the $30,000? NO ANSWER. The input sheet says it is separate, but the plan never says so.
5. What if the studio reaches capacity before week twelve? NO ANSWER. The timing line mentions capacity, but nothing says what happens next.

The two NO ANSWER lines are the value of the exercise. Add one sentence to the plan for each: the fee is excluded from the media budget, and spend pauses when the studio tells you it is full. A sceptic found them in a minute, and a client would have found them in the meeting.

The one-page version for the client

Clients read one page, so give them one, with the working underneath it. It carries six lines: the goal and the budget; the plan table with its source column; the base case and the two stress cases; the gates and the date of each; the weekly metrics and who reads them; and a short list headed "what we do not yet know". That last list is not a weakness. A plan that names its guesses invites the client to help test them, and a plan that hides them loses the argument in week five.

When the goal is reach, not enquiries

If the client wants awareness, the same discipline applies with different arithmetic. Say $10,000 goes on a reach campaign at an assumed cost of $12 per thousand impressions. That buys 10,000 ÷ 12 × 1,000, or about 833,000 impressions. At a target of four exposures per person, it reaches about 208,000 people. Now check the audience: if the platform's audience estimate for the targeting is 150,000 people, the same impressions give an average frequency of 5.6, not 4, and the plan is overspending on the people it can reach. Either widen the audience, cut the budget or accept the higher frequency knowingly. The cost per thousand is an assumption to label like any other, so source it before it goes in the plan.

Why four channels is the ceiling at $30,000

The commonest AI draft for a small budget spreads it across eight channels, because the model has seen long channel lists in plans for large advertisers. At $30,000 over 12 weeks, eight channels leave each with $3,750, or about $312 a week. At the client's $96 for a social enquiry, that is roughly three enquiries a week per channel, too few for any platform to learn from and too few for you to judge. Four channels with a capped test is the plan you can learn from. If the client insists on more reach, the honest conversation is about the budget, not the channel count. And if one of your four is Google's own, the choice between its campaign types on a small budget is covered in Performance Max vs Demand Gen.

Questions agencies ask about AI-drafted media plans

Which AI tool is best for drafting a media plan?

ChatGPT, Claude and Gemini can all produce a usable first draft. The structure of your prompt and the checks you run afterwards matter far more than the model. Give it a one-page input sheet, make it label the source of every cost figure, and keep client account data to what the task needs.

Can AI pull benchmarks for my client's industry?

It can suggest figures, but it cannot show you the data behind them, and it may attach a real-sounding source to an invented number. Treat any benchmark as a guess until you have opened the page and can see the year, platform and metric definition. Use the client's own history first and published benchmarks only as a sanity band.

How much budget is too little for one channel?

Enough to produce the conversion events the platform needs to optimise. Meta's help page says an ad set usually leaves the learning phase after about 50 results in the week after its last significant edit. If a channel's weekly budget cannot reach that, use fewer ad sets, a higher-volume conversion event, or treat the channel as a capped test.

How long should a channel run before I move its money?

Give it at least four weeks and enough conversions to compare, then use a pre-agreed gate rather than a hunch. A gate such as 'stop if cost per enquiry is above 1.3 times the assumption at week four' catches clear failures without judging a channel on a handful of results.

Should the plan keep some budget unallocated?

Many agencies hold back around 10 percent until the first gate so there is money to move without cutting anything. The phase plan in this tutorial gets a similar effect by tapering the test channel and keeping a stated rule for where its remaining budget goes if it fails.

Further reads

• ChatGPT Prompts for a Marketing Agency: Briefs, QA and Client Updates — Prompts for briefs, QA and client updates.
• How to Measure Whether AI Is Improving Your Marketing Results — How to tell whether AI is improving results.
• Best AI Tools for Advertising Agencies: Copy Variants, Mock-Ups, Reporting — Tools for copy variants, mock-ups and reporting.
• Performance Max Reporting: Explain Asset Results to Clients With AI — Explaining Performance Max asset results to clients.
• Marketing Agency Pricing After AI: Retainers, Packages and Deliverables — Pricing retainers when AI speeds up the work.
• AI Ad Copy Generator for PPC Agencies: Responsive Search Ads at Scale — Responsive search ads at scale.
• AI Tools for Agencies: What a Five-Person Team Pays Each Month — Seven layers priced per seat for a five-person agency, with annual and monthly totals, a leaner option and a skip list.
• PPC Audit Template: Run Google Ads Audits for New Clients With AI — A five-section Google Ads audit template, the AI prompts that summarise the exports, and a filled sample with a one-page prospect summary.
• PR Coverage Report: Build One for Each Client With AI — A coverage report structure, the clippings log columns, AI prompts for the summary, and a sample monthly report with the checks a person must do.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: LocaliQ search advertising benchmarks 2026 (updated 1 June 2026), Google Ads API documentation on Keyword Planner forecasts, and Meta's learning-phase help page, as checked in November 2026.