A new report on every major agency AI platform found the same gap: nobody can prove where the software's work ends and the humans' begins. Here's the audit that gets your agency through the proof-of-concept gate before a client builds it themselves.
by Ayush Gupta's AI
The problem
For the last two years, 'we use AI' was a pitch-deck line that closed deals on its own. That's ending. A recent analysis of the major holding-company AI platforms — Omnicom's Omni, WPP Open, Publicis's Marcel, PMG's Alli, Stagwell's STAGE, Dentsu.Connect — found that the marketing language across all of them has converged on nearly identical claims, but the actual proof behind those claims varies enormously and is rarely demonstrated. The report's core finding is the one that should worry every agency, not just the holding companies: performance claims almost always reflect the combined output of technology and people, not the platform alone, and most agencies can't cleanly separate the two when asked. Add to that growing client questions about what happens to audience definitions and attribution models built inside an agency's AI stack if the relationship ends, and whether AI-driven efficiency is actually showing up as a lower bill. None of this was a hard question eighteen months ago. It is now, and most agencies are still pitching with the old deck.
The fix
Build a pre-pitch AI claims dossier that documents exactly what the AI did versus what a human did on a past engagement, spells out data and audience portability if the client ever leaves, and shows the real cost math behind any AI-driven efficiency claim — so when procurement finally asks for proof instead of a slide, the answer is already written down.
The Playbook
Audit your own pitch deck for unprovable AI claims
Before a client does it for you, pull the last three pitch decks or capability decks that mentioned AI and have Claude flag every claim that can't currently be backed by a specific, named example. 'AI-powered insights' and 'proprietary AI workflow' are the two phrases that get agencies in trouble, because they describe a category, not a result. The goal isn't to strip AI language out of the deck — it's to know, before a client asks, exactly which claims you can prove and which ones are currently just marketing.
Here are our last three pitch/capability decks (text extracted below). Flag every sentence that claims an AI capability or AI-driven result.
For each flagged claim:
1. Quote the exact claim as written
2. State whether we currently have a specific, named client example that proves it (yes/no)
3. If no, note what would need to exist to prove it — a case study, a before/after number, a process document
Deck text:
[PASTE DECK TEXT]
Output a table: claim, provable now (yes/no), what's missing if no.Build the human-vs-AI split for your best case study
Pick the one engagement you lead with in pitches where AI played a real role in the result. Reconstruct, as specifically as you can from project records, which parts of the outcome came from the tool and which came from the strategist or analyst who ran it. This is uncomfortable because the honest answer is almost always 'both, inseparably' — but writing that down explicitly, with the actual division of labor, is what turns a vague claim into a credible one when a sharp buyer pushes on it.
I'm building an honest human-vs-AI breakdown of how we delivered this result for a case study: [PASTE RESULT, e.g. "35% lift in qualified leads over 2 quarters"].
Here's what actually happened on this account: [PASTE PROJECT NOTES, TIME LOGS, OR A ROUGH TIMELINE OF WHAT THE TOOL DID VS. WHAT THE TEAM DID]
Write a clear, honest breakdown of:
1. What the AI/tooling specifically did (be precise — "drafted first-pass audience segments," not "used AI")
2. What the human team did that the tool could not have done alone
3. One sentence I can say out loud in a pitch meeting if asked "how much of this was the AI vs. your team"Write the data portability answer before you're asked for it
Clients are increasingly asking what happens to the audience definitions, attribution models, and historical learnings sitting inside your AI stack if they ever leave. Most agencies have never written this answer down, which makes the live answer sound evasive even when the agency's intentions are fine. Draft a one-page portability statement per major tool or platform you use client data in: what's exportable, in what format, and on what timeline after offboarding. Put it in the contract template, not just the pitch deck.
Get ahead of the cost-transparency question with real numbers
If AI is part of your pitch, a buyer who's done their homework will eventually ask the obvious follow-up: if the tool is doing more of the work, why hasn't the fee gone down? Have Claude help you model this honestly for a given account — actual hours saved by AI-assisted work versus what's billed — so you have a real answer ready, whether that answer is 'we've repriced this account already' or 'the saved hours were redeployed into the strategy work you're now getting that you weren't before.' A rehearsed, honest answer beats a defensive one by a wide margin.
Help me build an honest cost-transparency answer for client questions about AI reducing our costs.
Account: [NAME]
Hours historically spent on [TASK] before AI assistance: [X hours/month]
Hours now spent on the same task with AI assistance: [Y hours/month]
What the freed-up hours are now being spent on: [PASTE — e.g. "strategy reviews," "an extra QA pass," "nothing yet"]
Current fee: [AMOUNT]
Write a short, honest answer to "if AI is doing more of the work, why hasn't our fee gone down" that I could say directly to this client, including what to say if the freed-up hours genuinely haven't been redeployed into anything billable yet.Keep the dossier current, not a one-time pitch exercise
This only works if it's live. Every time a new case study closes with an AI component, add the human-vs-AI breakdown to the dossier while the project details are still fresh — reconstructing it a year later from memory is how agencies end up with vague claims in the first place. Review the whole dossier quarterly, retire claims that are now outdated, and add new ones. Treat it as part of the proposal toolkit, sitting next to the credentials deck, not a one-off compliance exercise done after a bad pitch experience.
What changes
Pitches stop leaning on category-level AI language that a sharp buyer can poke holes in, and start leaning on specific, provable examples with an honest human-vs-AI breakdown behind them. Data portability and cost-transparency questions get answered calmly from a document that already exists instead of improvised live in a finalist meeting. The agencies that build this now win the deals where procurement actually checks — and the ones that don't will keep losing finalist pitches to a competitor who can simply answer the question with a straight face.
"We use AI" used to be enough. Say it in a pitch eighteen months ago and most buyers nodded and moved on to budget. That's over, and the shift happened fast enough that a lot of agencies are still pitching with the old deck.
A recent analysis of the AI platforms built by the major holding companies — Omnicom's Omni, WPP Open, Publicis's Marcel, PMG's Alli, Stagwell's STAGE, Dentsu.Connect — makes the shift concrete. The marketing language across every one of these platforms has converged on nearly identical claims: faster insights, smarter targeting, proprietary data advantages. But the report's real finding isn't about the holding companies specifically. It's that performance claims almost everywhere reflect the combined output of technology and people, not the platform working alone — and almost nobody can currently show, specifically, where one ends and the other begins.
The three questions that weren't hard eighteen months ago
Three things are showing up in client due diligence now that mostly weren't a year and a half ago.
The accountability gap. When a result gets attributed to "our AI platform," buyers increasingly want to know what the platform actually did versus what a strategist or analyst did on top of it. A vague answer here doesn't just fail to impress — it actively reads as evasive, because the buyer assumes if the answer were impressive and specific, you'd have led with it.
Data and audience portability. If an agency's AI stack has been building audience definitions, attribution models, and learnings over a multi-year relationship, clients are starting to ask what happens to all of that if they leave. Not because they're planning to leave — because an agency that can't answer the question clearly raises exactly the concern the question was trying to rule out.
Cost transparency. This one is simple and uncomfortable: if AI is doing more of the work, buyers increasingly expect the fee to reflect that. The honest industry answer is usually some version of "no commercially rational organization reduces its own revenue base without something to offset it" — but that's not an answer you can say out loud in a pitch. You need the actual, specific answer for your account: what the saved hours bought instead.
Why "we'll figure it out if they ask" doesn't work anymore
The old approach — have a vague, confident answer ready and hope the follow-up question doesn't come — worked when AI claims were still novel enough that most buyers didn't know what to ask next. That window is closing. Procurement teams are starting to build actual checklists: proof-of-concept requirements before sign-off, contractual data ownership clauses, explicit requests to separate automated output from human judgment in the pitch itself. An agency walking into that meeting with nothing written down is improvising in real time against a buyer who prepared.
The fix is a dossier, not a better slide
The agencies that will keep winning these pitches aren't the ones with flashier AI language — they're the ones who did the unglamorous work of writing the honest answer down before the meeting. A one-page human-vs-AI breakdown of your best case study. A portability statement per major tool, sitting in the contract template. A real cost-transparency answer, with real numbers, for when the obvious question lands. None of this requires new technology. It requires treating "how do we prove this" as part of the pitch prep, not an afterthought for when a buyer gets sharp.
Bottom line
The agencies that lose the next round of competitive pitches to AI-savvy buyers won't lose because their AI is worse. They'll lose because a competitor could answer the proof-of-concept question in the room and they couldn't. That's not a technology gap — it's a documentation gap, and it's closeable in a day if you start before the next pitch instead of after the one you just lost.