second open

Your AI feature earns 41% more per user. It also churns 36% faster.

We build products and AI systems from zero that run in production, and we fix the ones that leak users. On a 30M-user AI learning app, monetization features we built helped cut churn by 35%.

Book two weeksSee the four tests

€6,000 for two weeks. Stop after day five and pay half. Figures from RevenueCat, State of Subscription Apps 2026.

The market's own numbers

Consumers pay more for AI features and leave them faster. And almost none of the people using the feature are paying for it.

+41%revenue per user in apps with AIRevenueCat, 2026
36%faster churn than apps without itRevenueCat, 2026
~3%of consumer AI users ever payInworld, 2026
33free users' AI bills carried by every payerInworld, 2026

Adding AI is the easy part. Keeping it, and making it pay, is not.

We build the whole system behind consumer AI, from the model router to the App Store build.

Day five, in writing

Week one ends with written findings against the four tests, each marked fix now or fix next, and the top fix costed at your volume.

Day-five findings
An AI tutor in a study app
Example
  • Does it feel right?Nothing on screen for six seconds while the answer loads.Fix next
  • Does it hold up?The answer is lost if the phone drops signal halfway through.Fix next
  • Is it right?Answers are grounded in the course, but one student's notes can reach another student's answer.Fix now
  • Does it pay?A free user costs €0.41 a month, and 2% of users drive 48% of the bill.Fix now
Top fix: cap free usage before the AI callEstimated saving at their volume€9,800 / mo

Every consumer AI feature has to pass four tests

Most agencies answer one. A design studio answers the first, a dev shop the second, an AI boutique the third, a growth agency the fourth. We answer all four, because a feature that fails any one of them does not survive.

01 · Design
Does it feel right?

The feature lives inside the product, not in a chatbot bolted onto the side of it.

  • Streaming, waiting and empty states that are designed, not left to the model
  • Failure that looks deliberate: offline, rate limited, refused
  • A designer on the team, so it ships looking like your product
02 · Mobile and backend
Does it hold up?

Consumer traffic, consumer networks, consumer patience.

  • Streaming that survives a train tunnel: resumable, cancellable, cheap to retry
  • On-device against server, decided on latency and unit cost rather than fashion
  • A gateway chosen for your latency budget, because every hop shows on a phone
  • App Store and Play review for AI content, which fails for undocumented reasons
03 · AI, retrieval and evals
Is it right?

Grounded answers, and a way to know the moment they get worse.

  • Chunking, embeddings, reranking and grounding, tuned on your content
  • Retrieval scoped per user, so private data never reaches someone else's answer
  • An eval set built from your own failures, in Langfuse or Braintrust
  • Evals in CI, so a regression fails the build rather than earning a one-star review
04 · Metering and pricing
Does it pay?

Around 3% of consumer AI users ever pay. We work both sides of the margin.

  • Free-user spend capped before the request, not discovered on the invoice
  • Cheaper models wherever the evals say they are good enough
  • Credits and hybrid pricing on RevenueCat, with every pack priced against real inference cost
  • Store-fee strategy per region: in-app, web-to-app, and US external purchase links
We have already built this once
A multi-agent system built from zero, with our own router across four model providers, retrieval on pgvector, ingestion in terminable workers behind kill switches, an embeddable chat widget, an in-app copilot, and a token wallet with a usage ledger, all before the tools for any of it were mature. Paying customers, not a pilot. Today we would buy half of it, and that is exactly why we know which half.
In production
4model providers behind one production router
8services in production from a single monorepo
292tables across 3 Postgres instances
30M+users on a consumer app we engineered for

Cost per 1,000 requests

routing change, W5

Same feature, same quality bar. Small models for classification and reranking, the expensive one only where it earns its place.

Work, not logos

We describe every project by what it does, not by name: two companies we worked inside, and five products our team built and still runs.

Previous work
AI learning app at 30M users
Product engineering on a European AI learning companion used in 20 countries. We shipped more than ten features to students worldwide, helped open new markets, and built monetization features that helped cut churn by 35%.

Churn

−35%

Features shipped

10+

Surface

iOS + Android
Read the case study
Previous work
Multi-agent AI platform
A multi-agent system built from zero, with multi-LLM orchestration tuned for commercial viability. Retrieval, ingestion workers, an embeddable widget and an in-app copilot, all metered per customer.

Router

4 model providers

Retrieval

pgvector, 1536 dims

Metering

Wallet + usage ledger
Read the case study
Built by our team
Five products we built from zero
The apps are published by the product studio our team runs; the analytics platform is in early access. Running them ourselves means we handle the unglamorous parts: in-app purchases, releases, store review, paywalls and churn.
All five case studies

You are talking to other agencies too

Good, you should be. Almost everything said in those conversations is unfalsifiable, and everyone has a case study. These questions separate people who have run consumer AI in production from people who have read about it. Ask us. Ask them. If someone answers them better than we do, hire them.

Ask every agency on your list
None of these are trick questions. All of them broke on us first.
  1. 01What does one free user cost you per month, and where is that limit enforced?
  2. 02What happens to the feature on a phone that loses signal halfway through an answer?
  3. 03Is retrieval scoped per user, or can one person's notes reach another person's answer?
  4. 04What share of your AI revenue goes to Apple and Google, and have you looked at web checkout?
  5. 05When a prompt change makes answers worse, what fails first: a test, or a review?
  6. 06Do your AI users churn faster than everyone else, and would you know if they did?

We publish these because the answers are the whole difference, and because a vendor who cannot answer them will cost you a year either way.

What is actually different here
None of which is "we are senior and we care".
  • We built our own router, token wallet and usage ledger before the tools were mature, so we know where each tool breaks
  • We do not build what you can buy. You get the right tools wired together, not a bespoke stack to maintain
  • A number on day five, not a demo: what a free user costs you, and which of the four tests fail
  • We pay our own inference and store bills on the products our team runs, so we optimise like it is our money
  • We have a designer, so the feature ships looking like your product rather than an internal tool
  • One engineer runs your engagement end to end. No project manager, and no pyramid of juniors underneath

Start with two weeks

No six-month contract up front. You start with two weeks and decide after that. If it is live, week one measures it. If you are starting from zero, week one scopes it against your real data and works out what it will cost. Either way you get written findings on day five, and week two is building.

01 · From zero
Build it from zero
A new AI product, or a new AI feature in the one you have. In two weeks: a first working version on your real data, what it will cost per user at launch, and the architecture to take it live. Built to pass the four tests from day one, not patched afterwards.
02 · Already live
Stop the churn
It shipped and people try it once. We find which of the four tests it fails, from the waiting screen to the answers, and fix the one that is costing you the most users.
03 · Already live
Fix the margin
For AI features users love and finance does not. Both sides: cap free-user spend and route cheaper models on cost; price credits against real inference and choose the right store and web checkout on revenue.
04 · Already live
Fix retrieval, add evals
For an assistant that is confidently wrong and nobody can say why. We instrument retrieval, build an eval set from your own failures, and wire it into CI so the next change is measurable.

What you get

01
Findings on day five.

What a free user costs you, measured if it is live and projected if it is new, and which of the four tests are at risk. Continue, or stop there and pay half.

02
Working code in a branch.

In your repo, or a new one you own, against your real data. Reviewable by your engineers, not a demo on someone else's content.

03
An eval set.

Running in the tool you already use, or the one we recommend, so the next change is a measurement instead of a vibe.

04
A cost model.

Per request and per user per month at your real volume, with both levers: what it costs, and what it earns.

05
An architecture note.

What to buy, what not to build, and why. Two pages, written for engineers.

Fixed price, two weeks

€6,000

Half at the start. Half after the day-five findings, and only if you continue.

Whatever you decide afterwards, you keep everything: the findings, the branch, the eval set and the cost model. Build it with us, with your own team, or not at all. If the honest answer is not at all, you get that in writing.

Book two weeks

Who this is for

Good fit

A consumer product with real users, adding AI or struggling with the AI it already has

A funded team building a consumer AI product from zero, that has to work past launch day

A model bill growing faster than the revenue it is supposed to drive

AI users who churn faster than everyone else, or a feature nobody opens twice

Funded or already earning, in Western Europe, the UK, Ireland or the US

Not a fit

You want a chatbot because a competitor shipped one

An idea looking for a technical co-founder, paid in equity

A web front-end rebuild with no AI or backend work in it

Staff augmentation by the seat, billed hourly, managed by someone else

Book a call.

Thirty minutes, no deck. Tell us what the feature is, and which of the four tests it is failing. You get a straight answer about whether we are the right people, and we say no when we are not. Rather write? Leave a note below and you usually hear back within a day.

Book a 30-minute callpoppaultudor@gmail.com