Audiences AICustom Decision Systems
Some decisions should not live in a meeting forever.
We design and build systems for recurring decisions that carry real money and a named owner, yet are too context-specific for a packaged product.
What qualifies
Good candidates are decisions that recur, carry real money, and belong to a named business function. They are being made today through meetings, spreadsheets, heuristics or fragmented systems, and they are plausibly supportable by data that exists or can reasonably be captured.
The test behind every build is consistent. The signal already exists digitally. The same event can legitimately mean different things, and those meanings lead to materially different actions, sometimes including doing nothing. The possible actions are finite enough to encode, the reasoning can be audited, and the logic deploys beside your systems of record.
What we turn down
- One-off analyses and dashboard-only requests
- A generic mandate to implement AI with no decision attached
- Systems with no named owner
- Automation on data that cannot support the decision
- Technology-first mandates where the method was chosen before the problem was understood
Every system makes the same things explicit
The recurring decision. The business owner of that decision. The signals and data needed to support it. The method appropriate to it, which may be rules, optimization, statistics, machine learning, workflow, human review, or a combination. The evidence and assumptions. Uncertainty, abstention or escalation where they are material. The action that follows. The KPI that should move. And the learning loop, with its owner, after launch.
No unauditable decisions. The method can be simple or sophisticated. The degree of explainability, validation, monitoring and human review must match the consequence of the decision. This firm has no quarrel with machine learning. It has a quarrel with unaccountable decision-making.
The build path
diagnose→ establish the data spine→ design the decision logic→ build→ run→ measure→ hand over
The sequence expresses the discipline. Engagements weight the steps differently: a decision on shaky data spends longer in the spine; a well-instrumented one moves to logic quickly.
What the output can be
Recommendation cards · optimizers and schedulers · planning systems · approval gates · workflow and exception logic · operating processes · decision-support interfaces · the data foundation a business needs before any downstream intelligence can be trusted.
What we will not do
Build a system with no named decision owner. Claim a value number the available data cannot support. Automate a decision before the underlying data deserves the trust. Force machine learning or generative AI where rules, optimization or process design fit better. Make a consequential decision impossible to audit at the level its risk demands.
Where the work has run
Recent decision territory spans customer lifecycle and retention, promotion and margin, assortment and merchandising, buying and procurement, inventory and supply chain, production and delivery, scheduling and planning, quality and complaints, store formats and expansion, and AI governance and portfolio prioritization. The industries behind that list range from department-store retail and beauty to apparel, optical, manufacturing, fuel retail, subscription media and education.
Client work is published with permission, or reconstructed to protect confidentiality. See the work →