Procter & Gamble AI Case Study: Ad Tests $30,000 → $1,000, Launches 5x Faster, Model Rollouts 90% Faster

In 2023, P&G’s chief brand officer put a number on a Pampers ad test: one day and $1,000. The traditional version took 30 days and $30,000.

A year later, 791 P&G professionals took part in a randomized experiment. One person working with AI matched the output quality of a two-person team working without it.

Those two facts come from different corners of the company. What connects them is a product funnel: insight, concept, claim, formula, creative, production line, shelf. P&G has published AI results at every one of those steps.

The measured ones are narrow:

• A Pantene launch workflow ran five times faster than the usual initiative timeline
• AI-generated product claims went from 33% to 100% “highly appealing” across two feedback rounds
• Deploying a new model version to a production line takes up to 90% less time
• Scrap rates fall 10 to 20% on lines with AI visual inspection

Each number belongs to one workflow or one experiment. P&G reports them that way. This case study keeps them that way.

The Problem: A Product Funnel Where Judgment Was Slower Than Ideas

P&G sells in 180 countries, runs plants in more than 35, and collects over two million consumer research touchpoints a year. Ideas were never scarce. Deciding which ones were right was slow, expensive, and siloed.

The friction showed up at four points in the funnel:

• Concepts carried their author’s bias: R&D professionals proposed technical ideas, commercial professionals proposed commercial ones, and nobody could tell early which were best.
• Claims took months: writing a product claim that was true, legally clean, and appealing meant weeks or months of draft-and-test cycles.
• Ad tests took a month: at 30 days and $30,000 per test, a creative team could not afford to test a revision before it shipped.
• Models did not travel between plants: equipment varied by country and by line, so a quality model that worked in one pilot needed rework in the next.

Underneath sat a data problem. In Personal Health Care alone, 5,000 products built from 22,000 components lived in systems that took hundreds of expert hours to join.

A routine supply question took two or more hours a day to answer.

Generative AI could make the ideas cheaper still. That was not the bottleneck.

P&G’s work went into the other half: making the verdict on an idea, a claim, an ad, or a unit on the belt as fast as the idea itself.

What Procter & Gamble Delivered

P&G’s investor decks, R&D publications, partner case studies, and one peer-reviewed experiment make the sequence reconstructible. The dates are publication dates. They do not describe one program.

Laid along the funnel, those deliveries form seven capabilities. Each one puts a faster verdict next to a faster generator.

How Procter & Gamble Built It

The funnel below is a synthesis. P&G has never published it as one diagram. The stages are as the company describes them; the connections between them are editorial.

Insight: make the data reachable first

P&G’s 2026 investor deck shows its AI stack as three blocks. A core data lake fed by consumer research and operational data. An “AI Factory” of infrastructure, models, agents, evaluation, interfaces, and governance.

And users, from data scientists to all employees.

The unglamorous version of that lake sits in Personal Health Care supply chain. P&G automated the joins with KNIME workflows that merge bills of materials, supplier, inventory, and logistics records into shared dashboards.

None of this is an AI result. It is what the next six stages stand on.

Concept: run the experiment before the rollout

Before scaling generative AI into product development, P&G tested it. Between May and July 2024, 791 professionals from four business units worked on live innovation problems. Harvard and Wharton researchers assigned them at random to four conditions.

The study appeared in Organization Science in June 2026.

Three findings shaped what came after:

1. A person with AI equals a team without it. The two conditions were statistically indistinguishable on quality.
2. AI erased the expertise silo. With AI, R&D and commercial people both produced balanced concepts.
3. AI did not improve selection. Unaided teams picked their top-rated idea about half the time. AI-assisted participants did so about 37% of the time.

Two caveats travel with it. Quality was scored by 19 trained evaluators on P&G’s criteria, not by market results.

And the time savings often quoted, 16.4% for individuals and 12.7% for teams, come from the working-paper version. They are not in the journal text.

For Pantene Sunkiss Glow, teams later built “twins” of the target consumer from the research database and brainstormed against them in real time. A package mock-up existed within a day.

Claim: generate with GPT-4o, rank with a model that thinks like a consumer

Claim Advisor splits claim writing in two. GPT-4o generates candidates from approved claims and images. A separate Phi-3 model, fine-tuned on 100,316 examples from past MaxDiff consumer studies, ranks them.

The ranker is the point. It imitates the best-worst choice consumers make in a MaxDiff survey, so it tracks what consumers would pick, not what reads well.

This is iteration, not a controlled comparison. Each round learned from the last. Legal review still follows, by people.

Formula: proprietary libraries behind the models

P&G’s chief R&D officer reports that a Perfume Development Digital Suite lets the company develop fragrances five times faster. It pairs a scent and molecular library with consumer-acceptance models.

The 2026 deck adds a Molecular Discovery Suite behind the Unstopables Luxe fabric enhancer. The CIO said molecule identification that typically takes six to eight years can happen in as little as six months.

That is a best case for one step, not a finished product.

The older, more measurable work is in formulation testing. P&G researchers published more than 47,000 labeled images of fabric softener vials so that convolutional networks could flag instability without destructive testing.

• Model does: shortlists molecules and scents; flags unstable formulations from images.
• Person does: safety, stability, and regulatory work, unchanged.
• Published figure: fragrance development 5x faster; molecule timing a transcript best case.

Creative: a one-day ad test

In Pampers AI Studios, open-ended consumer reactions to a draft ad feed a model trained on decades of historical ad responses. The score guides revision. It is the clearest before-and-after in the P&G record.

A test that takes a day can change the ad before it ships. A test that takes a month cannot. This is the experiment’s selection problem, solved with tooling.

Pampers’ 2022 U.S. results, 10% sales growth, 17% higher ROI, and 15% media savings, came from a bundle of targeting, media, and creative changes. The one-day test is one input.

Line: build in the cloud, run at the edge, reject on the belt

P&G’s manufacturing problem was variation, not modeling. Its VP of IT has said equipment differences across countries “made things difficult when we went to scale.”

The fix, built with Microsoft since 2022, separates where a model is trained from where it runs:

1. Line data flows up: sensor and quality data lands in an Azure data lake.
2. Models are built once: data scientists train and retrain in the cloud.
3. Models ship down: Kubernetes-packaged models deploy to plant edge devices through Azure Arc and MQTT.
4. Operators keep control: they tune tolerances and schedule version upgrades around production.
5. Inference becomes a task: a flagged anomaly triggers an operator inspection.

The published result: deploying a new model version takes up to 90% less time. That is a release figure, not a build figure.

Inspection goes a step further. The Visual Inspection Cockpit, built with Siemens, was expanded worldwide in September 2026.

It runs P&G’s deep learning models on GPU industrial PCs at full line speed. Inference is wired directly to alerts and rejection actions.

Shelf: availability as the last metric

P&G reports that AI-driven insights reduced out-of-stock rates by 15%, without disclosing geography or period.

The Guangzhou “lighthouse” plant offers a fuller picture. It faced a 45% rise in e-commerce demand.

A three-year program of AI, automation, and digital twins delivered 30% lower inventory, 15% lower logistics cost, and 99.9% on-time delivery, per the World Economic Forum.

That is a program result, not an isolated AI effect.

Where the model is allowed to act alone

Lining the stages up exposes a rule P&G never states but follows. The model acts on its own only where a wrong call is cheap and reversible, or where a physical check happens anyway.

Adoption followed the same logic. By January 2025, ChatPG, P&G’s internal assistant, had more than 30,000 users. One employee had trained over 1,600 colleagues.

The Harvard researcher behind the teaching case on P&G’s AI Factory credits uptake to three things: close work with intended users, experimentation days, and business teams owning the problem.

Business Outcomes in Numbers

P&G’s figures do not share a method. Some come from a randomized trial, some from the company, some from vendors, some from conference transcripts. The tables below grade them rather than add them.

Experimental

Company-reported with a stated baseline

Partner-reported or program-level

Not used as evidence here

Conclusion

P&G’s AI record is often told as a list of tools. Walked as a funnel, it is a list of verdicts made faster.

The experiment set the terms. Generative AI raised the quality of what people produced and flattened the gap between specialists. It did nothing for their ability to choose.

Everything P&G built afterward closes that gap at one stage. A ranker trained on 100,000 consumer choices. An ad test that returns in a day. A camera that rejects at line speed. A deployment path that ships a model in a tenth of the time.

For enterprises mapping their own funnel, the question is not which model to buy. It is which stage has the slowest judgment relative to its generation.

That is where production AI engineering, data platforms built for evaluation, and cloud-to-edge delivery earn their keep.

The model is rarely the bottleneck. The verdict is.

Disclaimer: This is an independent case study review, not a Zoolatech project. It is based on publicly available sources, including P&G’s investor presentations and R&D publications, peer-reviewed research, partner case studies, and press coverage.