The Markin ARPU report for B2C enterprisesRead now
MARKIN

COMPARE/Markin vs an LLM connected to your data

vs an LLM with MCP: asking questions is not running growth

+17–35% ARPU against holdoutObserved range across Markin deployments, measured on treated cohorts.

An LLM with MCP access reads your warehouse and answers questions about it, brilliantly and cheaply. Markin runs the loop that comes after the answer: it sizes each hypothesis in money, tests it against a randomised holdout, and routes the work across cheap models, frontier models and classical ML so that reasoning about millions of customers every week is economically possible.

The short answerLast updated: August 2026

This is the real alternative most teams weigh, and for exploration the LLM usually wins. The difference is what happens next. A chat answer is plausible and unfalsifiable; Markin turns a candidate into a sized hypothesis, an experiment with a control group, and a number finance accepts. Cost is the reason the architecture underneath looks nothing like a prompt.

01

An LLM with MCP access

A frontier model, ChatGPT, Claude or Grok, given tools over your warehouse, BI and product systems through the Model Context Protocol. It reads, reasons and explains on demand.

Choose it when the job is to understand what happened, draft an analysis or unblock an analyst in minutes.

  • Answers on demand
  • Priced per token
  • No control group

02

Markin

An autonomous growth-science team. Versioned skills generate, size, design, launch and read hypotheses continuously, using the cheapest model class that can do each step correctly.

Choose it when the job is to move ARPU across millions of customers and prove the movement was incremental.

  • Decisions, not answers
  • Holdout on everything
  • Routed model stack

Line by line

The same ten questions, answered for both.

Markin compared with An LLM with MCP access across ten dimensions
DimensionMarkinAn LLM with MCP access
What it isAn autonomous growth-science team: it investigates why revenue per customer is stuck and acts on what it finds.A general reasoning model with read access to your systems through MCP servers.
What it decidesWhich commercial opportunity deserves to exist for each customer this week, what it is worth, and when the right answer is to do nothing.Nothing on its own. It proposes; a human decides, briefs and builds.
Where hypotheses come fromGenerated by Markin from customer, product, pricing and technical-health data, then sized before anyone builds anything.Generated on request, in the direction the prompt points, with no memory of what has already been tried.
How a hypothesis is evaluatedSized in money on the eligible population, filtered by statistical power, then killed or kept by a randomised holdout.Judged by how convincing the explanation reads. Nothing sizes the idea in money or checks whether it could ever reach significance.
Scope of actionMarketing, product, pricing and technical-health hypotheses, arbitrated against each other in one queue.As wide as the tools it is given, but always ending at a recommendation in a chat window.
Which models do the workA routed mix: frontier language models to write and explain, small cheap models for volume classification, and classical ML and deep learning, uplift, survival, time series, embeddings, for the numbers.One frontier model for every step, whether the step needs reasoning or not.
What the cost scales withDecisions taken and revenue proven, not tokens burned. Per-customer reasoning is handled by the cheap layers by design.Tokens consumed. Reasoning over each customer weekly on a base of tens of millions is not a pricing problem, it is an impossibility.
How the work reaches the customerWritten back into the systems you already run, as attributes, events or API calls. Markin does not add a new customer-facing surface.A human reads the answer and does the work in another system.
How impact is provenA randomised holdout on every decision. The reported number is incremental revenue and ARPU, not attributed conversions.Whatever the analyst sets up afterwards, usually attribution rather than a holdout.
Where the data sitsReads context where it already lives, warehouse, CDP, product and billing systems. No new system of record.Your warehouse and tools, accessed live through MCP servers you host and secure.
Governance and controlEvery action carries its hypothesis, its expected value, its guardrails and its control group, reviewable before launch.Prompt and tool permissions. No record of why a given recommendation was made or what it was worth.
Time to a verified numberOne revenue theme, one channel, one holdout: a defensible incremental number inside 90 days.Minutes to an answer. The clock to a verified number has not started.
Best fitLarge B2C bases where the constraint is how many good hypotheses get tested, not how many messages get sent.Analysts and operators who need to understand the business faster.

What is at stake

A decision layer is not a tool line item. It moves ARPU on the whole base, every month.

Installed base

2.0M

customers at $24 ARPU / month

Addressable revenue

$259.2M

per year, reachable base

Verified ARPU uplift

+17% to +35% ARPU

on treated cohorts, against holdout

What that is worth

$44.1M – $90.7M

incremental revenue per year

Measured on treated cohorts against a randomised holdout, read over a full measurement window rather than the first weeks. Anonymised range across Markin deployments in large B2C bases; your own holdout is the number that decides. The figures above apply that range to the reachable share of the base on this page's assumptions; they are arithmetic, not a forecast for your business.

Run it on your own numbers

The unsolved part

What is still missing when the model can read everything

Access was never the hard part. The hard part is turning a plausible statement into a decision that is affordable to take a hundred million times, and into a number that survives being questioned.

  • Plausible is not falsifiable: nothing in a chat answer estimates the value at risk or the power needed to detect it.
  • No memory of method: the same question asked twice can produce two incompatible recommendations.
  • Cost scales with curiosity, not with revenue, so per-customer reasoning stays out of reach.
  • No control group, so every reported win inherits the attribution problem you already have.

Hypothesis space

Everything a human growth scientist would look at.

Most growth problems are not message problems. Markin is not restricted to the campaign surface: if something is holding ARPU back, it is in scope, and it gets tested the same way.

Marketing

The classic surface, but chosen per customer rather than per segment, and always against a holdout.

  • Which offer this specific customer is worth making
  • Channel and timing chosen per person, not per campaign
  • Contact pressure and fatigue arbitrated across every programme
  • Win-back economics: who is worth a discount and who is not

Product

Where the customer actually experiences the value, and where most silent revenue loss happens.

  • Onboarding steps that lose customers before first value
  • A feature with high retention correlation that half the base never discovers
  • Paywall and upgrade prompt placement
  • In-product surfaces used as a treatment arm, not just email and push

Commercial

Pricing, packaging and the shape of the offer itself, tested rather than argued about.

  • Plan and bundle structure by cohort
  • Discount depth against margin, not against conversion alone
  • Annual versus monthly framing per customer
  • Dunning and involuntary churn recovery sequences

Technical health

Anomalies nobody asked it to look for. This is the category no decisioning engine covers.

  • A checkout error rate that rose on one device and one region
  • Payment failures concentrated in a single issuer or method
  • A broken deeplink quietly killing a high-value journey
  • Latency or delivery degradation eating conversion before any message does

Think of Markin as a data science and growth team that never sleeps: it investigates, forms hypotheses, ships them into your own stack and proves each one against a control group, at a volume no human team can reach.

Job to be done

The same work, at a different throughput.

Nothing below needs a tool that does not exist. It needs the work to happen continuously instead of once a quarter, and to be proven against a holdout instead of argued about.

Job to be done, compared between With An LLM with MCP access alone and With Markin
Job to be doneWith An LLM with MCP access aloneWith Markin
Notice that revenue per customer is drifting in a segmentSomeone spots it in a dashboard review, weeks after it started.Detected as a signal the day the drift clears noise, with the segment already sized.
Explain why it is happeningAn analyst is pulled off the roadmap for a two-week investigation.An investigation runs automatically and returns the drivers with their evidence.
Come up with hypotheses worth testingA workshop produces the handful of ideas the room happened to think of.Hypotheses are written continuously across marketing, product, pricing and technical health.
Decide which hypotheses deserve budgetPrioritised by seniority and gut feel, with no size attached.Each one is sized in revenue and ranked before anything is built.
Choose the next best action for one customerSegment rules and campaign calendars decide, refreshed when someone has time.Chosen per customer, per moment, against everything else competing for that customer.
Actually launch itA ticket to the lifecycle team, then a slot in next month's calendar.Executed inside the systems you already run, with no new channel to adopt.
Prove it caused the revenueReported against non-qualifiers or a global holdout, if at all.Every decision carries a randomised control group; uplift is read against it.
Kill what does not workProgrammes survive because nobody owns retiring them.Failing to beat control retires the programme automatically.
Do all of it again next weekCapacity-bound: four to eight tests a quarter.Hundreds of hypotheses in flight in parallel, continuously.

Honest take

What An LLM with MCP access does better.

A comparison that only flatters one side is not worth reading. These are the cases where we would tell you to stay where you are.

  • For exploration, the LLM is better and far cheaper

    If the question is what happened last week, why a cohort moved, or how to write a query, an LLM with MCP beats any platform on speed and cost. We use exactly this internally, and no one should buy software to replace it.

  • The MCP ecosystem is moving fast

    Some of the plumbing we build today, tool access, schema awareness, safe read paths, will be commodity. The part that will not commoditise is the experimental discipline: sizing, power, holdouts and retirement.

  • A small base does not need any of this

    Below a few hundred thousand customers, one good analyst with a frontier model and warehouse access will out-produce an autonomous system, because there is not enough population to power the experiments Markin depends on.

  • General models reason better in the open

    Faced with a novel, unstructured question, a frontier model with tools is more flexible than any skill library. Markin is narrower on purpose: it trades breadth for reproducibility.

Where Markin fits

Not a replacement. A growth-science team on top.

Markin is what sits between the answer and the money: a skill library that sizes, designs, launches and reads, with the model class chosen per step so the economics work at population scale.

Skills, not prompts

Anomaly detection, sizing, experiment design, holdout reading and arbitration are versioned components with their own evaluations.

Model routing is a cost decision

Frontier models where language adds value; small models for volume; classical ML and deep learning where the number has to be defensible.

Every claim carries a control group

The output is incremental ARPU, not a convincing paragraph.

Evidence standard

Most of this category reports its own lift.

None of the major engagement, CDP or personalisation vendors publishes an independently verified uplift figure for its decisioning product. Where numbers exist, they come from vendor-commissioned studies or single-customer case studies with no disclosed holdout methodology. The most rigorous public research in the category is not flattering to anyone, including us, which is exactly why we build against it.

How Markin holds itself to it

  • Every decision Markin makes carries a control group. Uplift is reported against that holdout, not against the customers who did not qualify.
  • Results are read over a full measurement window rather than in the first weeks, so novelty is not mistaken for effect.
  • Programmes that fail to beat control are retired automatically. Killing decisions that do not pay is part of the loop, not an annual review.
  • The one figure we quote about ourselves is a range, not an average: +17% to +35% ARPU on treated cohorts against a randomised holdout, across Markin deployments in large B2C bases. We publish no industry benchmark, because we could not source one we would be willing to defend. Your holdout is the number that matters.

Time to value

90 days to a number that survived a holdout.

No replatform, no data migration, no rebuild of the channels you already run. If the first cohorts do not beat control, nothing scales and you have lost a quarter, not a roadmap.

  1. Weeks 0–2

    Read the context you already have

    Markin connects to the data and the channels you run today, An LLM with MCP access included. No migration, no replatform, no new source of truth.

  2. Weeks 3–6

    First sized opportunities in test

    Opportunities are ranked by expected value, treatments are chosen per customer, and the first cohorts go live with a randomised holdout attached.

  3. Weeks 7–12

    First verified incremental revenue

    Results are read over a full measurement window. What beats control scales; what does not is retired. Nothing scales on a number that has not survived a holdout.

Which one you should pick.

Choose Markin if

  • You need a decision per customer per week, not an answer per question.
  • The number has to survive a randomised holdout and a finance review.
  • The base is large enough that untested ideas cost real money.
  • Per-customer reasoning has to be affordable at tens of millions of records.
  • You want the method to be versioned and repeatable, not re-prompted each time.

Choose an LLM with MCP if

  • The job is exploration, diagnosis or drafting.
  • Your analysts are the bottleneck and want leverage, not autonomy.
  • You are not ready to run controlled experiments on the base.
  • The base is small enough that judgement beats testing.
  • You want to start this week with tools you already pay for.

When you don’t need Markin.

  • You want a chat interface over your warehouse: use an LLM with MCP, it is the right tool.
  • The base is too small for controlled measurement.
  • There is no appetite to act automatically on anything, even under guardrails.

Questions buyers ask.

Why can't we just connect ChatGPT or Claude to our warehouse with MCP?

You can, and you should, for analysis. What it will not do is decide what each of ten million customers should get this week, size those decisions in money, and prove the result against a control group. Those steps need a method that is versioned and an execution cost that does not scale with tokens.

Isn't Markin just an LLM with better prompts?

No. Language models are one layer of several, used where language actually helps: writing hypotheses, explaining findings, drafting briefs. The scoring, uplift modelling, survival analysis and anomaly detection that produce the numbers are classical ML and deep learning, and most per-customer work never touches a frontier model at all.

How does the cost compare?

Different unit entirely. An LLM with MCP is priced per question asked. Markin's economics are built around decisions taken, which is why cheap models and trained models do the volume work and expensive reasoning is reserved for the few steps where it changes the outcome.

What stops an LLM from producing good hypotheses?

Nothing. It produces plenty, and some are good. The problem is that it cannot tell you which ones are worth the opportunity cost, whether the eligible population is large enough to detect the effect, or whether the same idea failed two quarters ago.

Do you use MCP internally?

Yes. Tool access to systems is useful plumbing and we treat it as such. It is a transport layer, not a growth programme.

How is Markin different from the decisioning or AI already inside An LLM with MCP access?

A decisioning engine ranks actions a human already defined, inside the campaign surface it was given. Markin forms the hypotheses itself, marketing, product, pricing or a technical anomaly holding growth back, sizes them, executes them inside An LLM with MCP access and your product surfaces, and reads each one against a randomised holdout. It behaves like a data science and growth team, not like an optimiser.

Does Markin only test messages and offers?

No. Anything a human growth scientist would investigate is in scope: onboarding friction, feature adoption, pricing and packaging, dunning, and technical health issues such as a checkout error rate or a broken deeplink quietly killing conversion. Marketing is one of four hypothesis domains, not the boundary.

What is the business case for adding Markin on top of An LLM with MCP access?

On a large B2C base, a small move in ARPU is a large number in absolute terms, because it applies to the whole installed base every month rather than to a campaign. Across Markin deployments the verified range on treated cohorts is +17% to +35% ARPU against a randomised holdout. The point is not more messages: it is finding the highest-value action per customer, launching it, and proving it against control before it scales.

How long before it pays for itself?

First sized opportunities are in test within six weeks and the first holdout-verified result lands inside 90 days. Payback depends on your base, margin and programme cost, the calculator on this page computes it from your own numbers, after applying the 20% to 40% haircut BCG finds when next-best-action programmes are incrementality-tested.