A vendor arrives with a £250k AI proposal. Marketing enthusiasts support it. IT discusses capabilities. Your board approves it.
Yet you wonder: Will this actually generate returns?
Most CFOs struggle to answer this question. The statistics are sobering: 95% of enterprise AI initiatives deliver no measurable return (MIT, 2025). 56% of CEOs report zero ROI in the last year (PwC, 2026). 71% of UK employees are using unapproved AI tools—half of them weekly (Microsoft UK, 2025), with minimal manager awareness.
The core issue isn’t technology itself—it’s governance. AI initiatives receive treatment as IT experiments rather than capital investments. Financial tracking doesn’t exist. Accountability for outcomes is absent.
You wouldn’t approve a £250k ERP system without implementation reports. Why would AI projects bypass this scrutiny?
What AI Actually Delivers
With proper discipline, AI produces three measurable results:
Growth: Customer churn prediction, prospect conversion identification, and pricing optimisation. Mid-market firms report 8–15% improvements in win rates and retention when AI deployment includes structured oversight.
Automation: Invoice processing cycles compress from 5 days to 1. Forecasting accuracy gains 10–20 percentage points. Customer service responses occur in minutes instead of hours.
Scale: A model automating accounts payable transfers to expenses, supplier onboarding, and contract review. Establishing governance for one tool enables deploying five additional solutions in roughly half the time.
The Cost of Inaction
44% of UK businesses have already suffered data exposure from shadow AI (SAP, 2025). Compliance leaders experience stress. Auditors raise difficult questions.
Unmonitored costs accumulate silently. License renewals proceed. Token consumption surges. Integration efforts stall. No oversight exists. No accountability assigned.
The 90-Day Implementation Framework
Step 1: Run a 30-Day AI Inventory (Days 1–30)
Objective: Identify where AI operates within your organisation before creating compliance problems.
Request a single page containing five columns:
- Tool name (ChatGPT, Copilot, AI-enabled SaaS applications)
- User groups (department, approximate user count)
- Primary function (which processes or decisions it supports)
- Data categories processed (customer, personal, financial, confidential, none)
- Risk considerations (example: “customer data transmitted to US servers,” “supports regulatory reporting”)
Execution: Assign one project lead from Finance or Operations. Create a small working team: one IT representative, one HR representative, one Risk/Compliance representative (if available). Set rigid timeline: 30 days maximum.
The goal isn’t eliminating everything. The goal is eliminating blindness.
Step 2: Set a Simple Investment Gate (Days 31–60)
Objective: Starting immediately, all AI proposals exceeding your chosen threshold (suggested: £50k annually) must pass a three-question evaluation before Board consideration.
Request one page addressing:
Problem: What specifically are you solving (quantified)? Examples: “Reduce invoice cycle from 5 days to 1 day.” “Raise forecast accuracy by 15%.” “Decrease customer churn by 8%.”
Baseline: What does the current process cost? Include personnel, duration, errors, write-offs, lost revenue. Absent quantification of current pain, reject the proposed solution.
Monetary Impact: Net annual benefit post-launch (savings plus revenue minus operating costs). Account for cloud expenses, token costs, integration, support, and ongoing license growth. Include implementation investment and payback timeline.
Projects unable to demonstrate this on one page lack readiness.
Step 3: Name Three Owners for Every Project (Days 61–75)
Objective: Clarify accountability roles without creating new governance layers.
For each significant AI tool or initiative, designate three owners:
Business Owner: Typically a functional leader. Bears responsibility for results and for halting initiatives failing to deliver.
Data Owner: Understands what constitutes “quality data” within your organisation. Frequently someone in Finance, Operations, or a data leadership position.
Risk/Compliance Owner: May be your existing Risk officer, Compliance lead, or Head of IT. Monitors ICO, FCA, and sector regulatory alignment.
Document these names. Include them in Board packs for significant AI initiatives.
Step 4: Track Monthly, Not Annually (Days 76–90)
Objective: Monitor progress consistently rather than waiting for annual reviews.
For each major AI initiative, include a brief monthly management report line:
- Forecasted metrics (benefit and cost, financially quantified)
- Current observations (monthly and year-to-date performance)
- Changes affecting usage, expense, data, or risk
- Single status indicator (red/amber/green)
Real-World Outcomes
The 30-day inventory surfaces shadow AI. Typical mid-market companies discover 15–25 tools. Considerable numbers process customer information via unapproved services.
The investment gate eliminates poorly-justified initiatives early. Organisations commonly block 2–4 proposals lacking defensible answers. Each represents £50k–£150k in prevented expenditure.
Monthly tracking enforces discipline. Adding one Board pack page creates financial visibility, which generates operational discipline.
Board Communication Strategy
One clear directional statement:
“We will evaluate AI like any significant investment: defined problem, concrete numbers, named owners, monthly tracking. This approach captures value while maintaining regulatory and stakeholder confidence.”
The Consequence of Delay
AI proliferation continues regardless. Shadow AI expands. Expenses accumulate invisibly. Risks compound.
More critically: you forfeit compounding advantages. Invoice automation governance enables operations teams to apply identical methods to purchase orders. Governance creates the infrastructure for exponential growth.
As Finance Director, you own the result—through design or by default.
Free download: AI Playbook for UK Mid-Sector Companies (.xlsx, 29 KB) — the inventory and investment gate templates referenced in this article.