Machine Learning Consulting and AI Model Strategy

We decide which models are worth building, what each one has to be worth, and how good it has to be. Your team or your build partner does the engineering.

Fondo del encabezado del servicio L

Machine learning consulting is the work of deciding which models a business should build, what each one has to be worth before it is funded, and how good it has to be before anyone acts on it. Vivaldi is a strategy firm, so our part is the decision, the model strategy and the evaluation. The engineering is done by your team or your build partner, and we say that plainly because most of the disappointment in enterprise machine learning comes from buying strategy and delivery as if they were one thing.

  • Desde 1999estrategia de demanda, comercial y de clientes
  • Más de 200estudios de caso publicados
  • 10 oficinasen cuatro continentes
  • 5xWirtschaftsWoche: Lo mejor de la consultoría
  • Sin vínculos con proveedoresnuestro modelo y nuestro asesoramiento sobre plataformas no tienen ningún interés comercial

The model performs and nothing changes

Most enterprise machine learning is technically sound and commercially inert. The accuracy is good, the dashboard refreshes, and the decision the model was built to improve is still made the way it was made before: by the same person, on the same instinct, at the same point in the week.

That is a scoping failure rather than a technical one. Work that starts from available data produces models looking for a use. Work that starts from a named recurring decision produces models that get used, because the person who will act on the output was in the room when the threshold was set.

The test is simple. Ask who changed what they did last quarter because of a model. If the answer is a name and a specific decision, the program is working. If the answer is that the team now has better visibility, it is not.

Cuando las empresas nos contratan

  • The data science team delivers and the business does not move. Capable people, respectable accuracy scores, and a P&L that looks much as it did three years ago.
  • Before a model, platform or vendor commitment. A build is about to be scoped before anyone has written down which decision it is meant to change and who owns that decision.
  • When an executive asks what the analytics investment returned. The honest answer is a list of dashboards, and nobody wants to be the one to say it.
  • When the choice between classical machine learning and a language model is being made by fashion. Usually in favor of the language model, usually for a task that does not need one.
  • When models are built and then quietly abandoned. Nobody owns the question of whether a model is still right, so accuracy decays and the answer keeps arriving.

Why a data-first approach does not work

The standard sequence is to audit the data estate, find where the data is good enough to model something, and model that. It is a reasonable way to produce work and a poor way to produce value, because the quality of your data has almost nothing to do with the value of the decisions it could improve.

Starting at the decision inverts the sequence. You name the recurring decision, establish what a better version of it is worth, then ask what prediction would be needed to change it and whether your data can supply that. Three times out of five the honest answer is that the decision can be improved without a model at all, which is a cheaper finding than a model nobody uses.

This also changes who has to be in the room. A data audit is a conversation with the data function. A decision inventory is a conversation with the people who own the P&L lines the models are supposed to move, and they are the ones who decide whether anything changes.

What our machine learning consulting services include

Cuatro etapas de trabajo. La mayoría de los proyectos empiezan por la primera y se detienen cuando se agota el valor.

1. A decision inventory and a value case per model

Before any modeling, we build the inventory of recurring decisions, who owns each one, how it is made today, and what a better version would be worth in money rather than in accuracy points.

  • Las decisiones se clasifican en función del valor en juego, en lugar de según la disponibilidad de datos
  • The decisions that should not be modeled at all, named explicitly, so the program is not judged against them later
  • A value case per candidate model, written before the model exists, so there is something to measure it against

2. Model strategy: which approach fits which workload

A deliberate answer to a question most organizations settle by fashion. Classical machine learning, a language model, a combination, or no model at all.

  • Classical machine learning for structured prediction at volume: forecasting, propensity, classification, optimization. Cheaper, faster, more testable, more explainable
  • Language models for unstructured input, interpretation and generation, including the extraction step that turns documents, calls and tickets into features a predictive model can use
  • Architecture that assumes the models will be replaced, because capability that was frontier two years ago is now mid-tier and available from several suppliers
  • We hold no partnership with any model or platform provider, so the recommendation carries no commercial stake

What the input looks like

Structured and repeating, predict a valueClassical machine learningForecasting, propensity, classification, optimization. Cheaper, faster, more testable and more explainable, and still the correct default for most of what a business actually needs predicted.
Structured and repeating, interpret or explainA thin layer, not a programA language model narrating what the predictive model found is useful and rarely worth a program of its own. Fund it as a feature.
Unstructured and varied, predict a valueWhere the cost runs awayPointing a language model at documents or calls and asking it for a number is the expensive, untestable version. Extract first, then predict with a classical model. Two steps, not one.
Unstructured and varied, interpret or generateLanguage modelsTheir actual job: reading messy context, summarizing it, drafting from it, and handling input that never arrives in the same shape twice.

What the job is

The tool follows the input and the task, not the news cycle. The quadrant that quietly consumes budget is the bottom left, where a language model is asked to do arithmetic it should not be doing.

3. Evaluation design, set before anything is built

The useful question is not what accuracy is achievable. It is what accuracy would change the action, and those are different numbers arrived at by different people.

  • The accuracy, latency and confidence the decision actually requires, taken from the person who makes it
  • The definition of done, written before the build starts, so the project has an end other than exhaustion
  • The monitoring plan: how drift is detected, how often the model is re-evaluated, and who is notified when it stops being right

4. The decision, governed

A prediction only changes something if someone, or something, is allowed to act on it. That is a permissions question wearing a technical costume, and it is where most programs stall.

  • Which calls a system may make alone and which have to escalate, set explicitly on materiality, reversibility and the confidence of the model itself
  • How the output arrives in the workflow where the decision is made, rather than in a report that is read afterward
  • The outcome capture that tells you whether the last recommendation was any good, without which nothing improves after launch day

Not sure which model is worth building?

Start with a working session with one of our partners. We will take a single recurring decision, establish what a better version of it is worth, and test whether a model would change it.

Book a model strategy review

How a machine learning engagement works

  1. Inventory. Interviews with the people who make the recurring decisions, a map of how each is made today, and a ranking by value at stake rather than by data readiness.
  2. Specify. For the decisions worth pursuing, the accuracy, latency and confidence required, the form the answer has to arrive in, and the escalation threshold. The output is a build specification written from the decision backward.
  3. Select and evaluate. The model approach chosen per workload, the evaluation design, and the acceptance criteria your team or your build partner will be held to.
  4. Hand over and measure. The specification goes to whoever is building, and a measurement line runs from the decision to the P&L so the value case can be checked against what happened.

Why companies choose Vivaldi for machine learning strategy

  • Partimos de la decisión, no de los datos. El argumento de valor precede al modelo, que es el paso que determina si se utiliza algo.
  • We are not selling you the build. Our revenue does not depend on the model being large, the platform being chosen or the program running long, which is a different incentive from a firm that staffs the implementation.
  • No vendor ties. We hold no partnership with any model or infrastructure provider, so the architecture advice is about your workload rather than our margin.
  • We will tell you not to build it. A recommendation to improve a decision without a model is a normal outcome of the first phase and the cheapest one available.

How to evaluate a machine learning consultant

Vale la pena preguntárselo a cualquier empresa de tu lista de finalistas, incluida la nuestra.

  • Do they start from the data or from the decision? A firm that opens by auditing your data estate will produce models looking for a use. Ask which decision the first model is meant to change, and who owns it.
  • What accuracy would be enough? The right threshold comes from the decision, not from what is achievable. A firm that cannot set it before building has no definition of done.
  • How does the output reach the person who acts? If the answer is a report or a dashboard, expect the decision to stay where it was.
  • Who is paid for the build? A firm that recommends a model and then staffs the engineering is answering two questions with one invoice. That is not disqualifying, but you should know it before the proposal.
  • What did you advise a client not to model? A firm that has built everything it was asked to build was selling capacity rather than judgment.

Strategy firm, data science consultancy or ML engineering vendor?

Criterio Strategy firm (Vivaldi) Data science consultancy ML engineering or platform vendor
Qué decide Which models are worth building, what each is worth, and how good it has to be How the model is specified, built and validated How it is engineered, deployed and run
Resultados típicos Decision inventory, value case, model strategy, evaluation design, build specification Models, notebooks, validation reports, sometimes a production handover Pipelines, serving infrastructure, monitoring, licenses
¿Quién lo aprueba? The commercial owner and the analytics lead together The analytics or data science lead Engineering and procurement
Interés comercial en la respuesta None in which model, platform or build partner you choose Usually in the size of the modeling team Directo
Ideal para cuando Los modelos funcionan y las decisiones no han cambiado The decision is settled and the model has to be built well The model is proven and has to run reliably at scale

What we do not do

We do not write production code, build data pipelines, run MLOps or operate models. None of our published case studies describe a model we built or deployed, and we are not going to imply otherwise on a page selling machine learning consulting. Our published work is strategy, research, segmentation and brand and commercial measurement, which is also why the proof section of this page is empty rather than borrowed from somewhere else in the estate.

What that buys you is an opinion with nothing attached to it. We will tell you which decision is worth a model, what the model has to achieve, and how to judge the firm that builds it. If you want one party to decide and build, we are the wrong firm, and you should ask any firm that offers both how it handles the conflict.

Trabajos relacionados

  • Predictive and advanced analytics if the prediction is commercial: demand and sales forecasting, lifetime value and churn, price and promotion response, inventory and assortment.
  • Data strategy if the binding constraint is the state of the data rather than the state of the decisions.
  • Generative AI strategy if the input is unstructured and the task is interpretation or generation rather than prediction.
  • AI strategy and new tech for the prior question of which workflows and decisions are worth changing at all.

Este trabajo forma parte de nuestra práctica de Estrategia Tecnológica Empresarial en Vivaldi, la consultora estratégica responsable de «Demand Landscape» y «Interaction Field».

Machine learning consulting in the US and worldwide

Our work runs out of ten offices on four continents, with the US team in New York. Model strategy travels better than models do: a decision inventory built in one market is largely reusable, while a model trained in one market usually misprices the same risk in another, because the alternatives available to the customer are different. We scope the decisions globally and the models locally, and we say which is which before the budget is set.

Empieza por una decisión

In one working session we will take a single recurring decision, establish what a better version of it is worth, and test whether a model would change it. You leave knowing whether the model is worth building, and what it has to achieve if it is.

Book a model strategy review

Machine learning consulting FAQs

What does a machine learning consultant actually do?

Identifies the recurring decisions worth improving, establishes what improving each one is worth, sets the accuracy, latency and confidence the decision requires, chooses the model approach, and designs the evaluation and the workflow the output has to arrive in. At Vivaldi the modeling and engineering are done by your team or your build partner. We do the decision, the model strategy and the evaluation.

Does Vivaldi build and deploy the models?

No. We are a strategy firm. We do not write production code, build data pipelines, run MLOps or operate models, and none of our published case studies describe a model we built or deployed. We specify what should be built and how it will be judged, and we help you hold the firm that builds it to that specification.

When is machine learning the right tool rather than a language model?

For structured prediction at volume, meaning forecasting, propensity, classification and optimization, classical machine learning is cheaper, faster, more testable and more explainable. Language models are for unstructured input, interpretation and generation. Most good systems use both, with the language model turning messy input into features and the predictive model doing the arithmetic.

How much data do we need?

Less than most vendors imply for a first useful model, and more than most organizations expect for a reliable one. The binding constraint is usually label quality and whether outcomes are captured at all, rather than raw volume. If outcomes are never recorded, no amount of input data will make the model improve after launch.

How do you know a model is good enough?

By setting the threshold from the decision rather than from the model. The question is what accuracy would change the action, not what accuracy is achievable, and those are different numbers arrived at by different people. We write the definition of done before the build starts.

How do you keep models accurate over time?

Through monitoring for drift, scheduled re-evaluation and a named owner who is notified when the model stops being right. Models decay as the world moves. The practical question is whether you detect it before it costs you, which is a governance design rather than a technical one.

What is the difference between data science consulting and machine learning consulting?

In practice the labels overlap. Data science covers the wider set of analysis, experimentation and inference, much of which never becomes a deployed model. Machine learning is the subset that ships something that keeps predicting. The useful distinction is not the label but whether the engagement ends with a finding or with a decision that runs differently.

How is this different from your predictive and advanced analytics service?

This page covers general model strategy: which models are worth building, classical machine learning against language models, evaluation design and the governance of the decision. Commercial and customer prediction, meaning demand and sales forecasting, lifetime value and churn, price and promotion response, inventory and assortment, sits on our predictive and advanced analytics page. Most organizations need both.

What if the answer is that we should not build a model?

That is a normal outcome of the first phase and the cheapest one available. A decision can often be improved by changing who makes it, when it is made or what information reaches them, none of which needs a model. We would rather say so than sell you a build, because we do not make money from the build either way.