Phase 2 · Staircase steps three and four

Platform to
Advantage

You have built the lakehouse. The layers that turn it into AI sit on top of it, on the same catalogue, the same identity model and the same storage, and each one makes the next use case cheaper than the last.

01Build on the platform
02Not beside it
03Layer by layer
04Each one compounds

See it layer by layer Map what goes on top

ApplicationThe use case itself: the assistant, the scoring, the thing in the workflow.Then

ModelRegistry, evaluation, guardrails and cost control, in one place.Build next

ServingThe same data, consumable by a model at workflow speed.Build next

SemanticBusiness definitions written once, in a form a model can read.Start here

LakehouseThe one you have built. Medallion, catalogue, governance.Built

Built once. Reused by every use case after

The hard part is done. Everything from here is cheaper because of it

Innovating for the world’s largest brands

Anglo American

Woolworths

PepsiCo

DHL

Transport for London

Discovery Bank

NEOM

RCL Foods

Where you are

Step three makes you faster.
Step four makes you harder to copy.

Nobody skips a step, and you are past the part most of the market is still arguing about. The difference between step three and step four is not more use cases. It is what sits underneath them.

Behind you01No AI engagementIsolated legacy processes. Nothing runs on it yet.

Behind you02Tactical deploymentsUse cases built, in production, serving teams.

This page03AI as enablerEmbedded in workflows. Faster decisions, same people.

This page04Competitive advantageNet-new revenue competitors cannot copy quickly.

Later05AI native businessThe organisation is built around the models.

Decision Inc. view of where the market sits in 2026Nobody skips a step. The businesses at the top got there one build at a time.

Still choosing your first build? Go to Phase 1: Where to Start Building

Layer by layer

What each layer takes,
and what it makes possible

The order is the point. Semantic first, because everything above it inherits the definitions. Application last, because it is the only one anybody ever wants to start with.

Layer

What you build

What it takes

What becomes possible

01Semantic

Metrics and definitions modelled once, on top of the gold layer, and certified.

The shortest of the four, and the most argued about. Most of it is business time, not engineering.

Anyone, or anything, can ask a question in business language and get the same answer twice.

02Serving

Features, retrieval and endpoints, built on the platform rather than as a copy of it.

Short, once the semantic layer exists. Largely configuration on the platform you already run.

A model can read your data at the speed a workflow needs, without anything leaving the estate.

03Model

Registry, evaluation, monitoring, guardrails and cost control, in one governed place.

A capability rather than a project. It does not finish, and it should not.

You can put a model in front of a customer, because you can prove what it does and show when it drifts.

04Application

The use case itself: the assistant, the scoring, the thing embedded in the workflow.

Weeks each, once the three layers below it exist. Most of the effort is the business process, not the build.

A use case becomes an application instead of a project. This is the whole return on the three layers above.

You are not being asked to fund a data project. You are being asked to finish an argument the business already owes itself.

Our work building the frontier

Examples of our
client work

Open an area to see what clients built with us.

01Generative AI in the journeyProducts customers and patients use, not internal demos.

HealthcareDischarge summaries rewritten into plain language. 4,553 patient queries, 1,541 patients, 90% satisfaction.

WoolworthsA decade of Taste content turned into a Gen AI Shopping and Culinary Platform.

EVT30,000 comments a month read, themed and flagged in real time.

02Predictive and governed MLModels a clinician, or a regulator, will accept.

HealthcareReadmission risk scored pre-discharge. Explainable outputs, drift monitored continuously.

PepsiCo30% reduction in staff overtime running cost, optimising utilization, demand and workforce configuration.

ArcelorMittal90% reduction in text ordering and 95% compliance to Contract Management in less than 1 year.

03Agents inside the processWhere the work actually gets done, at volume.

Henry Schein70% of customer chats resolved by an agent, in 11 languages, against live order data.

Kogan40% of enquiries handled end to end: tracking, cancellations and returns.

Austin HealthResearch compliance review from 15 minutes to 1, approvals in a day not weeks.

04The foundation underneathWhere most of the effort goes, and where it fails without it.

OmniaLegacy and siloed to a governed lakehouse for 3,800 people, with certified data products.

MiningA production AI data platform stood up in four weeks, then handed to their own team.

MedTech14.5 TB of reclaimable storage down to 119 GB. An architecture fix, not a clean-up.

Case studies

Built, shipped and
running in production

MiningSouth AfricaMining AI TransformationRead case study

Oil and GasAustraliaBPRead case study

Multi-IndustrySouth AfricaOmniaRead case study

Why the next one costs a fraction

The first build pays for the middle.
Everything after it pays only for itself.

The layers are the reusable part. That is the whole argument for building on the platform rather than beside it.

The use case

Model layer

Serving layer

Semantic layer

Reuses the layers built for the first

The use case

Reuses the layers built for the first

The use case

First buildSecond buildThird build

Relative effort. The thin band on builds two and three is the marginal model-layer work each new use case adds.

What this actually asks of you

Fund the first one differently. It carries the middle of the stack as well as itself, so it will always look poor judged as a single use case.

Judge it on the second. The real test of the first build is what the one after it costs.

Why there are no numbers on this chart

Because we have not seen your platform. The shape holds everywhere; the ratio depends on what is already in place, and that is the first thing the mapping engagement measures.

Where value sits

Which of these is worth
building the middle layers for?

At this step the question is not where to start. Mark the function that already has a use case somebody has asked for, and the one with the data closest to hand.

Finance and accounting

Ledger and close · integration · intra-month forecasting · validation controls

Planning and performance

Consolidation · bottom-up budget and forecast · cash flow · statutory reporting

Operations and supply chain

Planning, routing, yield, quality, maintenance, procurement

Data and analytics

Customer insight · benchmarks · operational reporting · the platform underneath all of it

Risk and compliance

Scenario modelling · single view of risk · framework automation · compliance documentation

People

Employee data · leave and expenses · performance · the internal knowledge hub

Faster is something you can buy.

Harder to beat is something you have to build.

Practical steps

3 Practical Steps
to Accelerate Agentic

Each is fixed in scope, and each one is built on the platform you already own rather than beside it. Choose the statement closest to your organisation.

Which sounds most like you?