AI & Machine Learning
A prediction is not a decision. We build the decision first, then the model.
Forecasting, propensity, detection and optimization, built into the workflow where someone actually has to act.
The people who scope it are the senior people who build it. If a model isn’t the answer, we’ll say so.
From prediction to decisionIllustrative
- PredictionStock runs short in week five
- ThresholdAgreed with planning
- DecisionReorder now
What we build
Five kinds of model. Each one answers a question someone has to act on.
Predict
“Which customers are about to leave?”
Churn · propensity · predictive maintenance · credit and claims risk
Forecast
“How much will we need, and when?”
Demand · inventory · workforce and capacity · revenue
Detect
“Which of these is not normal?”
Fraud and risk · quality defects · equipment and transaction anomalies
Optimize
“What is the best plan within our limits?”
Pricing and promotion · routing · inventory · workforce scheduling · allocation
Recommend
“What should we offer this person next?”
Next best offer · next best action · ranking · personalization
Working with documents, text or conversation? That is usually a generative AI problem. See Generative AI & AI Agents
Why models stall
The algorithm is rarely what fails. Three other things are.
Validated on the wrong decision
An impressive metric, and no economic lift.
Never wired into the workflow
A prediction nobody acts on.
Left alone after launch
The value decays, and nothing raises an alert.
How we are different
Every build starts with a Data IQ™ Decision Brief.
Decision first: the model is the last thing we design. Before any engineering, the brief records who makes the decision, what triggers action and what a wrong call costs. It is the first artifact of our Data IQ™ methodology, and the build is measured against it.
Data IQ™ Decision BriefExample
- The decision
- Which members get a retention offer this week?
- Made by
- Marketing operations, in the campaign tool
- Acts when
- The score clears a threshold agreed with finance
- Measured by
- Incremental repeat visits per offer
- Guardrails
- Margin and frequency caps
- Human override
- Yes, with the reason captured
- Reopened when
- Lift falls below the agreed floor
Frame the decision
Users, workflow, measure of success, threshold, risk tier.
Engineer the data
Readiness, features, labels, and splits that prevent leakage.
Model and validate
Backtests, segment checks and explanations, against the decision.
Deploy and sustain
In the workflow, monitored, with a trigger to retrain.
Proof
Four models that changed what people did.
Recommend · Retail
14%
more loyalty sign-ups
11% lift in repeat visits, at 8% less discount per incremental visit. A national quick-service restaurant chain.
What the model did Chose the offer type, depth and timing for each member, within margin and frequency caps.
Read the case studyOptimize · Manufacturing
12%
fewer delivery miles per order
18% fewer stockouts at key partners and 9% better plan adherence. A large food manufacturer and distributor.
What the model did Sequenced deliveries and consolidated loads, on cleaner demand signals fed into the planning system already in use.
Read the case studyPredict · Financial services
25%
fewer manual comparison steps
15% faster portfolio-fit identification. With one ranked list and one set of visible criteria, alignment between the investment and distribution teams improved 32%. A global asset manager.
What the model did Scored portfolios for suitability and policy fit, showing the criteria behind every recommendation.
Read the case studyForecast · Manufacturing
+20%
forecast accuracy
Used to optimize inventory and eliminate supply shortages across product lines and regions. A global manufacturer.
What the model did Forecast sales and demand from sales, planning and market data, refreshed as new data arrived.
Read the case studyHow each figure was measured is set out in its case study.
In production
Built to keep working after we leave.
We build the whole system in your operating environment, not a model in ours:
- Data pipelines and features built for live scoring
- Model serving, batch or real time, versioned and tested
- Workflow integration where the decision is made
- Monitoring for performance, drift and data quality
- Retraining on a trigger, not a calendar
- Handover to a named owner, with the runbook
Model governance follows NIST AI RMF where it applies, with documentation and audit trails scaled to how much the decision matters.
The same model, two ways of running itIllustrative
Not sure the data can support a model? Two free self-assessments, six to seven minutes each, no email required.
How to engage
Three ways in. One of them is the model in production.
Want proof on one use case first?
AI & Agent Pilot
One use case taken from workflow to a running system on your data, with quality and unit economics measured. Two to four weeks. Fixed scope. Fixed fee agreed in writing first.
- A running system, handed over with its source
- The evaluation harness
- Measured unit economics
- A decision: scale, revise or stop
Have a decision worth improving? · The main engagement
Production ML build
One decision. One production system. Measured in the workflow.
- Starts with
- A signed Data IQ™ Decision Brief
- Proceeds when
- The data can support the decision
- Ends with
- A monitored model in use, handed over
- The decision, its measure and its threshold, agreed first
- Data pipelines and features built for live scoring
- A model validated against the decision, with explanations
- Monitoring, drift alerts and retraining triggers
- A team that can run and extend it without us
Several models already running?
MLOps and model operations
The inventory, monitoring and review routine that keeps a portfolio of models reliable.
- A model inventory with owners
- Shared deployment and monitoring
- Retraining and retirement rules
- A review cadence that holds
Not sure the data is ready? The two-week AI & Data Assessment Sprint replaces assumptions with evidence. Sprint details
How we work. Every engagement follows our Data IQ™ methodology, from diagnosis through implementation and handover. See how we work
Before you commit
Questions we get every week.
Do you build only the model, or the whole system?
The whole system: data pipelines, features, training and serving, monitoring, the integration into the workflow, and the handover that lets your team run it. A model on its own does not change a decision.
How do you avoid models that test well and fail in production?
By validating against the decision, not the dataset. That means time-based splits, backtesting, leakage checks, segment-level performance, and a threshold set with the people who act on it. Monitoring and drift detection are in place from the first release.
Machine learning or generative AI: which do we need?
If the answer is a number, a score, a ranking or a plan, it is usually machine learning. If the work is reading, writing or conversing over documents and knowledge, it is usually generative AI. Many systems use both, and we scope them together when they do.
How do we know whether our data is good enough?
By looking at it, early. Data readiness is the first technical step: access, quality, definitions, and whether the outcome you want to predict has been recorded reliably. If it has not, we say so before modelling starts, and the Assessment Sprint exists for exactly that question.
How is the return measured?
At the workflow, against a baseline set when the decision is framed: cycle time, throughput, loss avoided, false positives, conversion or cost per outcome. The measure is agreed before engineering begins, so the result is not argued about afterwards.
Can you work in regulated settings?
Yes. Access controls, audit logging, model documentation, explanations and review routines are scaled to how much the decision matters. We build the controls and the record; we do not offer legal advice.
What happens after the model is live?
It is monitored for performance, drift and data quality, with a named owner for alerts and an agreed trigger for retraining. Your team runs it, with our support for as long as that is useful.
The people who scope it are the people who build it
Bring us the decision you want to improve. We’ll tell you what it takes.
The approach, the data it needs and a realistic path to production. Thirty minutes with the senior people who would do the work. If a model isn’t the answer, we’ll say so.
Talk to the people who would build itGo deeper.
- Case studyRewards optimization for a national restaurant chain
- Case studyLoyalty personalization for a fast-casual brand
- Case studyA decision layer for a global asset manager
- ArticleModel drift: what to monitor and when to retrain
- ArticleMLOps in practice
- ArticleMachine learning or generative AI: which does the problem need?
- FrameworkThe Data IQ™ framework and maturity model
- AssessmentsFree AI and data assessments
- Case studiesDelivered work, with the figures
A model is only as good as what surrounds it.
Have a decision a model should improve? Thirty minutes with the people who would build it.
