analytical leadership
WHAT DOES AN AI-AUGMENTED DATA AND ANALYTICS TEAM LOOK LIKE?
Most data teams still deliver reports and answer requests. The change that matters is that a trusted answer becomes a company service, available inside the decision and carrying its definition, sources and uncertainty with it. Analysts still interpret ambiguity. The governed answer does the retrieval, calculation and tracing around them.
The company has data, but
still waits for answers
At most mid-market companies, Data and Analytics remains a capability spread across IT, finance or marketing. Analysts administer BI tools, maintain dashboards and respond to requests. Where a central function exists, it combines data engineering, analytics engineering, BI analysis, insight partnership and part-time support for definitions and control.
The function sits across the value chain. It informs pricing, segmentation, personalisation, pipeline choices, product decisions, planning, working capital and process performance. Its distinctive contribution is the speed and reliability with which evidence becomes a decision.
Six processes| 01 | BI reporting and dashboard maintenance | 04 | Data-quality monitoring and incident resolution |
| 02 | Ad hoc data requests and analysis | 05 | Source integration, transformation and platform operation |
| 03 | Metric definitions, analytical models and governance | 06 | GDPR data handling, access controls, lineage and data-subject support. |
Where the information lives is the problem. Evidence is dispersed across ERP, CRM, HRIS, commerce or product systems, marketing platforms, workplace tools, BI datasets and spreadsheets. Join logic sits inside report files or analyst notebooks. Definitions sit in finance workbooks, slide decks and people's heads. Two reports can each look coherent and still disagree about revenue or customers.
The reporting queue is not only a capacity problem. It is the visible symptom of definitions, permissions and source logic that the company has never made reusable.
Current scorecards count report delivery, backlog, refresh failures and platform availability. The measures that matter are different: time from question to trusted answer, reuse of certified data, incident recovery and whether an important answer can be traced to its sources.
The unit of output becomes
the answer
The function changes structurally in two ways.
The unit of output moves from reports to governed answers. Routine query generation, charting and first interpretation become inexpensive. Managers answer bounded questions through certified models under their existing permissions. Each answer carries its definition, source lineage and incomplete coverage. Analysts spend less time translating requests into queries and more time testing reliability, choosing the valid definition and interpreting what remains ambiguous.
The data layer becomes an operating product for the whole company. Integration, quality rules, lineage and definitions stop arriving as isolated projects. They run as a continuous service. The central team owns the platform, standards, data contracts and answer quality. Finance, commercial, operations and HR stewards own meaning and acceptable use in their domains. IT retains infrastructure, identity and security.
A manager should be able to ask a cross-functional question and receive a current answer with its meaning, sources, access basis and uncertainty attached.
Where it is weakestthe combination is weakest where evidence is sparse, causality matters or the business definition is contested. Fluent language cannot settle a disagreement about what the company means. It can only spread the wrong definition faster.
A day, a week,
a month
A company that invests now still has a central Data and Analytics team in 2028. The team concentrates on analytical models, connectors, reliability, insight partnership and portfolio choices. Domain stewards remain inside the functions. Capacity released from report production moves into decision support, repeated experiments and data reliability.
The dayReliability agents test freshness, schema changes, volume shifts, broken joins and unusual metric movements. They open incidents with a likely upstream cause and show which reports, answers and workflows are affected. Engineers decide whether the diagnosis is right and approve material fixes.
A manager asks a bounded question through the company's briefing layer. The service runs under that manager's identity, queries certified models and returns the calculation, definition, sources and incomplete coverage. An analyst reviews low-confidence answers, questions that cross domains and requests that imply cause.
An insight partner enters an operating meeting with a pre-read built from current data, recent decisions and material exceptions. Follow-up questions are answered while the discussion is live when the evidence supports them. The meeting moves from reconciling numbers to choosing action.
The weekDomain stewards review changes to definitions, access and data contracts. The central team studies failed questions as a product backlog: misunderstood terms, missing joins, stale sources and requests that should remain human-led. Repeated questions become tested answers instead of repeated tickets.
The monthA metric council approves consequential definition changes. The function head reviews answer reliability, unresolved data debt, sensitive-data use and decisions enabled. Connector and model changes pass through test data, shadow use and approval before production. Access reviews and deletion obligations use the same catalog and lineage records.
Continuous monitoring, routine analysis without a dashboard request, traceable answers inside the meeting and a searchable history of changed definitions. Capabilities once too demanding to maintain become the normal analytical service.
What runs, and what
stays with the person
| PROCESS | WHAT THE AGENT DOES | WHAT STAYS WITH THE PERSON |
|---|---|---|
| BI reporting and dashboards | Drafts queries, visuals, commentary, scheduled briefs and threshold alerts against certified measures | Approval of new measures and consequential narratives; the action leaders choose |
| Ad hoc requests and analysis | Clarifies the question, selects approved sources, writes and tests the query, records assumptions and produces a first interpretation | Causal claims, ambiguous questions and executive recommendations |
| Metric definitions and governance | Finds duplicate measures, drafts definitions from code and documents, identifies conflicts, proposes tests and shows downstream effects | Domain stewards approve meaning; the central team approves implementation |
| Data quality and incidents | Watches tests and anomalies, traces likely causes, identifies affected assets and drafts remediation | Engineers approve schema changes, backfills and source corrections |
| Integration and platform operation | Drafts mappings, transformation code, tests, documentation and impact analysis | Engineers retain architecture, security and release authority |
| GDPR, access and data-subject support | Maps fields to purposes, assembles evidence, checks retention rules and routes exceptions | The DPO or legal owner decides lawful basis and disputed actions |
Four stages on
The Access Ladder
The practical destination begins with a governed warehouse and a small number of valuable source connections. The first domain should answer one repeated cross-system question. A broad lake programme before that question creates infrastructure without proving a decision benefit.
The stages map directly onto the rungs of GRAIL's Access Ladder. The order matters.
Begin with the report inventory, metric glossary, transformation code, dashboard exports, privacy policies and examples of recurring requests. The agent identifies duplicate definitions, documents hidden logic, generates test questions and helps design the operating model. The analyst supplies current context by hand.
Connect the systems behind one valuable question, usually ERP, CRM and the existing BI model. Establish identity, lineage and approved metrics with the connection. Current records can be joined, answers inherit user permissions, quality is monitored continuously and every result can show where it came from.
Add the operational or product source that explains movement in financial and commercial measures. The agent proposes reference-data corrections, tickets and workflow updates. A named person sees the evidence and approves the write through a separate permission path.
Reversible operations can run inside a defined threshold with an owner and rollback path. Financial posting, employee decisions, consequential definition changes and disputed privacy actions retain named human authority. Autonomy arrives only after read-only answers have achieved agreed accuracy.
A warehouse becomes necessary when the answer needs history, cross-system joins, stable identifiers, replayable calculations or company-wide access controls. Point connections cannot maintain one versioned definition across tools.
Headless operation changes the interface, not the accountability. ERP, CRM and warehouse platforms remain systems of record. Every action retains the user, instruction version, sources, generated query, metric version, result, approval and outcome.
Six things we believe,
from building this
Data and Analytics is not becoming a faster reporting department. It is becoming the company service that determines whether an answer is fit to enter a decision. Six beliefs follow from that shift.
A dashboard fixes the questions and leaves interpretation to whoever opens it. A governed answer begins with the live question and returns the calculation, definition, sources and limits together. The function should organise around that moment of use.
Natural-language access over raw tables creates confidence before it creates meaning. Definitions, stable identifiers, certified models and test questions are what turn fluent output into a company answer. The model is the visible part; the analytical contract carries the load.
A central team cannot decide what an active customer, net revenue or productive capacity means for every function. Domain stewards own those meanings. The data team owns their implementation, lineage, testing and consistent use.
Every misunderstood term, missing join and stale source shows where the analytical service is incomplete. Treating those failures as a product backlog improves the whole company. Hiding them behind a polished answer makes the service less trustworthy each time it is used.
The agent can retrieve, calculate, compare, trace and draft an interpretation. The person decides whether the definition fits, whether correlation is being mistaken for cause and what action the evidence supports. Junior analysts still need contact with raw data and reconciliation, because that is where they learn why a plausible answer can be wrong.
The first connection should answer a repeated question that matters across systems. Once the question works, the company can add sources, history and controlled actions in an order tied to use. Starting with a broad platform programme delays the moment anyone learns whether the answer changes a decision.
These six beliefs concern the operating model, not the tools. Together they define who owns meaning, how an answer earns trust and where human judgment enters before evidence becomes action.
Roles, rhythm, and
where it fails
The function head moves from managing a reporting queue to owning the company's analytical contract: which sources count, which measures are certified, who may use them and how answers are tested.
The BI analyst becomes an insight partner and evaluator. The analytics engineer becomes the load-bearing role because reusable transformations, definitions and tests turn source data into business meaning. The data engineer concentrates on connectors, identity, reliability and controlled write paths. Domain stewards remain in their functions. Managers learn to challenge definitions and distinguish correlation from cause.
The rhythmInventory reports, definitions, code and repeated questions for one bounded domain
Establish governed read access to one warehouse domain
Put trusted answers and exception briefs into a changed cross-functional operating rhythm
Add sources and approved actions after read-only use is stable, and redesign roles around the new service
Where it failsA lake programme before a valuable question. Raw tables before definitions. Query volume treated as success. Central analysts claiming meanings the functions must own. Fluent answers without verified questions, citations or review. Write access before read-only use is stable. Review added on top of the old reporting queue. Junior roles stripped of the work that teaches error recognition.
The team does not become smaller by default. Its capacity moves from dashboard maintenance and repeated reconciliation toward data reliability, analytical judgment and decisions that deserve better evidence.