Taupo Consulting All articles
Professional Services

Governing the Tools Nobody Declared: A 60-Day Sprint for Shadow AI

Taupo Consulting
Governing the Tools Nobody Declared: A 60-Day Sprint for Shadow AI

Photo: Daniel Edler (MillwallDan85), CC BY-SA 4.0, via Wikimedia Commons

Shadow AI is a governance issue before it is a technology issue

Generative AI has arrived in UK mid-market firms through the browser, not through the IT roadmap. Employees use public tools to draft client communications, summarise long documents, prepare first-cut analysis and write code. Much of this is sensible and productive. Very little of it appears in board papers.

The exposure is manageable but real. Confidential client information may be entered into consumer tools. Personal data may be processed without a lawful basis or appropriate safeguards. Outputs may be inaccurate and then relied upon. Supplier terms may allow data retention or model training that the firm would not accept in a formal contract. None of this requires a dramatic response, but it does require visibility and boundaries.

A 60-day sprint is enough to move from an unknown position to a governed one, provided the objective is pragmatic. The goal is not to ban experimentation. It is to make approved use easy, high-risk use visible and unapproved high-risk use unacceptable.

Days 1-15: Map actual usage without turning it into a hunt

Start with a short, anonymous survey. Ask which tools are used, for which tasks, with what kinds of data and whether outputs are shared with clients or used in decision-making. Follow up with a small number of interviews across client-facing, finance, operations, legal and technology teams. Where the firm has lawful access to expense, procurement or single sign-on records, use them to identify known tools rather than to monitor individuals.

Classify usage into three groups: known and approved, known but unapproved, and likely but undocumented. The third group is usually larger than expected. Look for workflows where AI is embedded in existing software as well as standalone tools. A meeting transcription feature in a collaboration platform is a governance question just as much as a public chatbot.

Identify high-risk workflows early. These include client confidential material, personal data, legally privileged advice, pricing and bid strategy, HR decisions, security configuration and code that will enter production. The purpose of the mapping phase is not to assign blame. It is to establish an inventory that is good enough to govern.

Days 16-30: Set a pragmatic acceptable-use policy

The policy should be short enough to be read and specific enough to be useful. Set out approved tools for defined data classes, the circumstances in which public tools may be used, the requirement for human review of outputs, and the prohibition on entering confidential client or personal data into unapproved consumer services. Include examples that reflect the firm's actual work rather than generic warnings.

Define roles. A business owner is accountable for the use case. A data protection lead advises on personal data. Technology or security owns tool assessment and access controls. Procurement owns supplier terms. The policy should align with UK GDPR and ICO guidance, but it should not pretend to be legal advice. Where the firm operates in regulated sectors, check sector-specific expectations before approving use cases.

Supplier terms deserve particular attention. Review data retention and training rights, data residency, subprocessors, security commitments, intellectual property positions and the provider's ability to change terms. A tool that is acceptable for public marketing copy may be unsuitable for client matter data. The policy should make that distinction practical rather than theoretical.

Days 31-45: Triage high-risk workflows and establish an approval route

Use a simple risk matrix: data sensitivity, potential consequence and reversibility. A workflow that processes personal data and feeds a client deliverable is higher risk than one that improves internal brainstorming. A workflow that writes code into production is higher risk than one that drafts a first-pass summary for human review.

For high-risk workflows, require controls. These may include redaction, enterprise tenancy, data loss prevention, logging, human review and a named accountable owner. For lower-risk workflows, publish an approved list and let managers proceed within clear boundaries. The approval route should be light-touch: a short form, a risk tier, a service-level target of five working days and a time-bound exception process. The objective is to make the safe route faster than the shadow route.

Days 46-60: Make it stick

A policy alone will not change behaviour. Managers need a short briefing on the boundaries and the approval route. Staff need examples of acceptable and unacceptable use. Procurement and onboarding should include AI tool assessment so that new services do not enter through the side door.

Establish a quarterly review. Track the number of approved tools, open exceptions, reported incidents, training completion and high-risk use cases. Report the position to the board or executive with a clear statement of risk appetite. Update the policy as tools and supplier terms change. Generative AI governance is not a one-off project; it is an operating discipline.

The output of the 60-day sprint is not a lengthy policy document. It is a visible inventory, a usable set of rules, a triage method and an approval route that people will actually follow. That is a more credible position than pretending the tools are not already in use.

All Articles

Related Articles

Interest on the Undecided: Running a Decision Debt Audit Before the 2027 Plan

Interest on the Undecided: Running a Decision Debt Audit Before the 2027 Plan

The Margin Nobody Owns: A Pragmatic Pricing Reset for UK Service Firms

The Margin Nobody Owns: A Pragmatic Pricing Reset for UK Service Firms

The Calendar as a Cost Centre: Auditing the Management Rhythm Before 2027

The Calendar as a Cost Centre: Auditing the Management Rhythm Before 2027