The reason most AI programmes stall is not the model. It is that the business cannot agree what a customer is, what revenue means this month, or which KPI the CFO and COO are both looking at. Fix that and the AI compounds. Ignore it and no model will save you.
Why do most AI programmes fail at the data layer — and how do you fix it without a two-year replatform?
The reversal
The industry story is that AI needs a modern data platform, and the roadmap tends to be a multi-year replatform costing eight figures. That story confuses the data layer (where information is stored) with the intelligence layer (where the business decides). An organisation can run on a fragmented data estate and still make world-class decisions if the intelligence layer above it is coherent. The reverse is not true: a pristine data platform under a fractured intelligence layer produces beautifully rendered disagreement. IBM's Institute for Business Value 2024 CEO study found that data foundation gaps are consistently cited among the top blockers to AI value — but the highest-performing firms address them with operating-model changes and targeted intelligence layers, not by pausing the business for a replatform.
The insight stack
What actually moves the P&L
Separate the data layer from the intelligence layer
The data layer is warehouses, lakes, pipelines, connectors — the plumbing. The intelligence layer is definitions, KPIs, decision rules, and the live signals executives actually act on. Most organisations conflate them, so every intelligence conversation becomes a platform conversation, and every platform conversation stalls on scope. Decouple them and you can ship an intelligence layer in months against whatever data estate exists today.
Write down the operating definitions before writing the queries
'Customer', 'active user', 'gross margin', 'churn', 'pipeline'. Every organisation has three versions of each and no owner for the canonical one. The single highest-return week in most engagements is the one spent forcing definitional agreement across functions and publishing the results as a governed operating dictionary. Everything downstream depends on it. Without it, AI models trained on one team's definition produce outputs another team refuses to act on.
The intelligence layer is a live signal, not a monthly deck
Performance management collapses when the signal cadence is slower than the decision cadence. A pricing decision that has to be made weekly cannot be governed by a monthly report; a customer-health decision that has to be made daily cannot rely on a quarterly cohort study. The intelligence layer's job is to compress signal-to-decision latency to below decision-cadence — not to look prettier than last quarter's dashboard.
Build outward from the top three decisions
Rather than 'model the whole business', identify the three highest-frequency, highest-leverage operating decisions — pricing exceptions, resource allocation, capacity planning, retention interventions — and build the intelligence layer that serves them first. Each of those decisions typically depends on 8–15 data elements, not 8,000. Shipping intelligence against three decisions in a quarter produces more value than modelling everything in a year.
Governance is metadata, not memoranda
Owner, definition, source of truth, freshness SLA, quality rules, lineage, downstream consumers. If those seven fields are not attached to every material metric, governance is a policy document, not an operating practice. Under both GDPR and the emerging EU AI Act, this metadata is also the substrate of regulatory defensibility — you cannot demonstrate lawful basis or model-input quality if the metric it references has no owner.
AI is a consumer of the intelligence layer, not a substitute for it
The most common failure mode is asking an AI to reconcile the three versions of 'revenue' the business has never bothered to reconcile. The AI does its best, produces a plausible synthesis, and no function trusts it. Fix the intelligence layer first. Then the AI has something coherent to reason over, and the outputs stop feeling like guesses.
Case example
the £120M business that shipped intelligence in one quarter
A £120M industrial-services business had spent 18 months evaluating data-platform vendors and had not shipped a single decision-grade dashboard. The engagement started by ignoring the platform question entirely. In week one, the operating team documented their top five weekly decisions and the 63 metrics required to make them. In week two, canonical definitions were agreed and published for 41 of those 63 metrics; the remaining 22 were classified as unresolved and given owners. In weeks three to twelve, a lightweight intelligence layer was assembled on top of the existing ERP, CRM, and finance systems — no replatform. By the end of the quarter, three of the five decisions were being made against live signal, and the fourth had an owner and a target date. The platform question was still open — but it had shrunk from a two-year strategic bet to a targeted, sequenced upgrade against a working intelligence layer.
Mini-playbook
The intelligence-layer diagnostic
List the top five operating decisions your executives make every week or every day.
For each decision, list the metrics required — usually 8–15, not hundreds.
For every metric, name the owner. If there is no owner, that is the problem.
Write the canonical definition. If two functions disagree, resolve it in writing this week.
Attach the freshness SLA (how stale can it be before it is unusable?) and the source of truth.
Deliver the metric against the existing data estate first — do not wait for the platform.
Publish the operating dictionary somewhere everyone can find it, and make edits go through governance.
Only after the intelligence layer works, ask the platform question — and ask it against a working spec.
How Strategy Labs installs this
Anchored to Performance management system
Strategy Labs installs the intelligence layer through the Consulting Advisory Engine (CAE), which holds the operating dictionary, the metric owners, the freshness SLAs, and the decision cadence as first-class governed objects — not as a Confluence page. Every metric is versioned, lineage-tracked, and tied to the operating decision it serves. Pragmatic DecisionCore (PDC) runs the primary and secondary research that underwrites external benchmarks and market-facing metrics so the intelligence layer is grounded in evidence, not internal narrative.
We do not implement data platforms. We build the intelligence layer that makes the platform question answerable — and then the platform choice becomes a targeted upgrade rather than a two-year strategic gamble.
Frequently asked
Related questions executives ask
- Do we need a modern data platform before we can do AI?
- No. A modern data platform helps, but it is neither necessary nor sufficient. The binding constraint on AI value is the intelligence layer — canonical definitions, owned metrics, decision-grade signal — which can be shipped in a quarter against an existing data estate. Replatforming without fixing the intelligence layer produces beautifully rendered disagreement.
- What is the difference between the data layer and the intelligence layer?
- The data layer is storage and pipelines — warehouses, lakes, connectors. The intelligence layer is definitions, KPIs, decision rules, and live signals executives act on. Confusing the two turns every intelligence conversation into a platform conversation, and platform conversations rarely finish in time to matter.
- How long does it take to build an intelligence layer?
- For a mid-market business, a decision-grade intelligence layer against the top three to five operating decisions can be shipped in a single quarter. Full-estate coverage takes longer, but the value is heavily concentrated in the first decisions.
- How does the intelligence layer intersect with GDPR and the EU AI Act?
- Metric-level metadata — owner, definition, source of truth, lineage — is the substrate of regulatory defensibility. Under GDPR you must demonstrate lawful basis for personal-data processing; under the EU AI Act you must demonstrate input-data quality for high-risk systems. Both obligations collapse without a governed intelligence layer.
Over to you
What is one metric in your business that three different functions define three different ways — and what has that ambiguity cost you? Share the story in the comments.
Continue reading
More AI Operations briefings
Where should AI actually live inside your operating model?
Most AI programmes start by asking what the tool can do. The ones that move the P&L start by asking where the operating model is already leaking value — and work backwards to the capability.
Read briefingWhat separates an AI agent that ships value from one that stalls in pilot?
Scaling a broken workflow with an AI agent does not fix the workflow — it industrialises the breakage. The agents that ship value inherit a redesigned process. The ones that stall inherit the old one at higher speed.
Read briefingWhat does responsible AI governance look like when AI is running live operations?
Responsible AI governance is not a policy document, an ethics committee, or a training module. It is an operating discipline — versioned, instrumented, and audited — that keeps unit economics honest while AI is running live in the business.
Read briefing
Discussion
(…)Comments are moderated before appearing. Your email is only used for moderation and is never shown publicly.
Loading discussion…