Self-Service BI: How Much Freedom, How Much Governance
The choice is not between control and freedom. It is how much of each applies to which content. A three-tier model gives every piece of BI content a standing, and gives each standing the controls it needs and no more.
Self-service BI is decided badly when it is decided once for the whole estate. Lock everything down and the central team becomes a queue, and analysts export to spreadsheets to get their work done. Open everything up and the organisation acquires several versions of each figure and copies of data nobody is tracking. The question is then put as a trade-off between the two.
It is better put as a question of scope. Different content carries different risk, and governance should be proportionate to it. This article works through that decision: what the published evidence says, a model that sorts content by who relies on it, what the model assumes, and how to tell when the balance has gone wrong in either direction. It concerns BI content only. The governance of data itself is a wider subject.
A word on the evidence. Most detailed guidance on this subject is published by BI vendors about their own platforms. Two of them are used below, and each recommendation that comes from a vendor is attributed to it. One independent source, a review of the research literature, is used to show that the problem is recognised outside vendor material.
What governance is for in a self-service setting
The purpose of governing BI is easy to state the wrong way round. Tableau’s published guidance for its customers puts it the right way: governance “creates accountability and enables, rather than restricts, access to secure and trusted content and for users of all skill levels”1. The same document names the tension directly: “Striking the balance between control and agility is critical”1.
It also describes what happens when the balance tips towards control: “In spite of stringent governance policies, users often go the route of locally saving sensitive data and analytics for quick analysis”1. That sentence is a vendor’s observation about its customers, not a measured finding, and the mechanism is plain enough. A control that stops people doing their work does not stop the work. It moves the work somewhere the control cannot see.
Governance in this setting therefore has two jobs. It protects the content the organisation relies on, and it makes the governed route the easy one.
Independent evidence on the question is thinner than the vendor literature, and it points the same way. A review of published work on implementing self-service BI, presented at the Hawaii International Conference on System Sciences in 2018, begins from a blunt observation: “Although SSBI promises more benefits compared to a traditional BI system, many organizations fail to implement SSBI”2. Two of the challenges it identifies are governance questions. One is “deciding who can use and add data, how long data should be stored, and determining the minimal data quality requirements”2. The other is policy: organisations “need to establish policies for data governance that define the data available for access, how it can be accessed and whether the standard of quality is appropriate for the desired analysis”2. The review synthesises earlier publications, several of them industry reports, so it shows that the problem is recognised in the research literature. It does not measure how often the problem occurs, and no figure is taken from it.
Three tiers of content
The model below sorts content by who relies on it. It is DII’s synthesis of two vendors’ published frameworks, which use different terms for similar ideas. Both describe how to adopt the vendor’s own product. The tiers are used here because the idea does not depend on either product, and where a term or a feature belongs to one vendor it is named as that vendor’s.
Three tiers of BI content and the controls for each
| Governed core | Certified shared content | Personal exploration | |
|---|---|---|---|
| What it is | The official datasets and reports the organisation runs on | Team and departmental content reviewed and marked as reliable | An individual’s own analysis |
| Who relies on it | Executives, regulators, the whole organisation | A department or several teams | The author |
| Who owns it | A central BI or data team | A named business owner, on centrally owned data | The author |
| Definitions | Must use the agreed KPI definitions | Must use the agreed definitions; may add local measures, labelled | Free |
| Review before release | Formal: tested, documented, signed off | Lightweight: reviewed against a checklist by someone authorised | None |
| Change control | Versioned and announced | Owner’s responsibility; recorded | None |
| Access | By role, enforced on the data | By workspace or group | Private by default |
| Marked as | Certified | Promoted or certified | Unmarked |
The governed core
This is the small set of content on which the organisation’s decisions and external reporting depend. Microsoft’s adoption guidance describes the equivalent scope as one where content “is typically managed by a centralized team and is subject to additional governance requirements”3. The controls are heavy on purpose: an error here is expensive and public.
The core should be small. Every report added to it takes on its cost of review and change control, so the test for admission is whether the organisation would be embarrassed by a mistake in it.
Certified shared content
The middle tier is where self-service earns its keep, and where it goes wrong if nothing is done. The workable arrangement is the one Microsoft calls managed self-service, which it summarises as “a blended approach to data and BI”4: business authors build reports, and they build them on data the central team owns.
Two features of this tier matter most. The first is the split itself. Business authors know what their colleagues need to see, and the central team knows how the data should be modelled. Each does the part it is best placed to do. The second is the mark. A reader needs to tell reviewed content from unreviewed. Microsoft’s platform distinguishes a promoted item, where “the item creators think the item is ready for sharing and reuse”, from a certified one, which “can be regarded as reliable and authoritative, and is ready for use across the organization”5. The distinction is useful on any platform: one mark an author can apply, and one that only an authorised reviewer can.
Personal exploration
The outer tier is individual work. Microsoft’s description of the scope is exact: “Personal solutions are, as the name implies, intended for use by the creator”3. Tableau’s guidance recommends giving it an explicit place, a personal sandbox that “can be used for ad hoc or in-progress analysis and hides the content that is not ready for broad release”1.
The single rule for this tier is that it stays personal. Analysis here is ungoverned because nobody else relies on it. The moment it is shared for others to use, it has moved tier and needs the middle tier’s review. A simple promotion route is what keeps this tier from becoming the place where unofficial reports are published.
One caution from the same source applies to the whole model. Decentralised content is not the same as uncontrolled content: “The concept of business-led self-service isn’t the same as shadow IT”4. Business-led work done on a sanctioned platform, visible to administrators, is a governed arrangement. The same work done in an unsanctioned tool is not.
What the model assumes
The tiers rest on preconditions that neither vendor’s guidance can supply for an organisation. These are DII’s reading of what the model requires.
Agreed definitions exist. The middle tier is told to use them. If there are none, each author invents their own and the certified mark certifies nothing. This comes first, and how to define KPIs and metrics that mean the same thing everywhere covers it.
Access is controlled on the data, not the report. If authors can build on any dataset they can reach, sensitive data will appear in reports whose audience was never considered. Access rules belong on the datasets and should follow the classification of the data, as set out in how to build a data classification scheme people will apply. The Access Policy Checklist is a short reference for the rules themselves.
Authors have the skills. Self-service moves work to people who were not hired as analysts. They need to know how to read the certified datasets, what the definitions mean and when a result should be checked with someone. This is the precondition most easily skipped, because tool training is mistaken for it. The Data Literacy Programme Plan sets out the wider capability a self-service model depends on.
Signs the balance is wrong
The balance can fail in either direction, and the signs are different. They are offered as things to look for, not as measured indicators.
Too much control.
- A queue of report requests measured in weeks.
- Regular exports from the BI tool to spreadsheets, followed by analysis there.
- Business units buying or building their own reporting tools.
- A central team that spends its time on formatting changes.
Too little control.
- Several reports showing the same KPI with different values.
- Nobody able to say which version of a report is current.
- Copies of extracts on shared drives with no owner.
- A certified mark that most content carries, or none does.
The last pair is the most telling. A mark applied to everything has stopped distinguishing anything, and suggests the review behind it has lapsed.
The response to each direction is to move a boundary, not to redesign the model. Too much control means the core is too large or the middle tier’s review too heavy. Too little means the middle tier is not being reviewed, or exploration is being shared without promotion. Where years of imbalance have left an estate too large to govern, the tiers have to be applied after a clear-out, as described in rationalising a sprawling BI estate.
Putting the decision on record
The model is set for an organisation by a small number of decisions. Each should be made explicitly and written down.
- What is in the core? Name the reports and datasets. Keep the list short.
- Who may certify? Name the reviewers for the middle tier and the checklist they apply.
- What is the promotion route? State how personal work becomes shared content, and how long that should take.
- Who watches the balance? Assign someone to look at the signs above each quarter and move the boundaries.
Each decision has a sensible starting position. Keep the core to the reports that go to the board, to regulators or to customers. Give the power to certify to a small group drawn from the central team and the departments, so that review does not wait on one person. Make promotion a request that is answered within days. A slow route teaches authors to share their work informally, which is the outcome the tiers exist to prevent. These starting positions are DII’s suggestion.
The last of these is the one that keeps the arrangement alive. Self-service governance that is set once and left will drift towards one failure or the other, for the same reasons any governance arrangement fades, which why data governance programmes stall examines at programme level.
The model is also only as good as the function it sits in. The four conditions that make reporting trustworthy are described in what makes business intelligence trustworthy, and the BI Maturity Assessment gives a way to score where a function stands before the tiers are drawn.
Sources
- Tableau (Salesforce). Governance in Tableau (Tableau Blueprint).
- Lennerholt, C., van Laere, J. and Söderström, E. Implementation Challenges of Self Service Business Intelligence: A Literature Review. Proceedings of the 51st Hawaii International Conference on System Sciences. 2018.
- Microsoft. Microsoft Fabric adoption roadmap: Content delivery scope (Microsoft Learn). Updated 30 December 2024.
- Microsoft. Microsoft Fabric adoption roadmap: Content ownership and management (Microsoft Learn). Updated 30 December 2024.
- Microsoft. Endorsement overview (Microsoft Fabric documentation, Microsoft Learn). Updated 11 July 2024.
This article is general professional information. It is not legal or regulatory advice.