How do you audit AI tools and workflows in your organisation without turning it into a six-month consultancy project?
A short AI-use review can begin with one team or service line: inventory its tools, trace the data they receive, name owners and record the decisions each tool affects. Use those findings to stop unsafe uses and prioritise deeper checks.
Book an AI Risk & Efficiency Audit
- What AI are we using? (including shadow AI)
- Where does the data go? (and what is sensitive)
- What decisions does it influence? (client impact, automation level)
- Who is accountable? (named owner with authority to pause it)
- What controls exist today? (and where they are missing)
Explore Managed AI Governance & Support
Record decisions and evidence gaps for each item: stop an unsupported use, fix a missing control, formalise a reviewed use or monitor a defined risk. Assign an owner and a next check to each action.
Two review pitfalls to check for
Look for these two possible problems in the review:
- Unmeasured rollout: pilots or Copilot-style deployments without clear ownership, data controls or outcome measures.
- Unimplemented policy: governance documents exist, but approvals, data boundaries and incident responsibilities have not been put into the workflow.
Check both the written policy and what the team actually does. Record discrepancies rather than assuming either a tool rollout or a framework demonstrates control.
Choose the scope of the initial review
For an initial review in professional services, financial services, insurance or B2B SaaS, consider the following areas and define which are in scope:
- Tools: ChatGPT/Claude/Gemini usage, Copilot, niche AI apps, OCR tools, meeting bots, vendor AI features.
- Workflows: prompt libraries, RAG/Q&A bots, document drafting, client comms, reporting, data enrichment, triage, claims/underwriting support.
- Data paths: what leaves the organisation, where it is stored, what is logged, and what is retained.
- Decision impact: advisory vs automated decisions, internal vs client-facing outputs.
Include informal use where it affects the chosen scope. Reviewing only the official project may leave undocumented tools or data flows out of the inventory; do not assume those are always the largest risks.
A 10-step approach for a scoped AI-use review
1) Time-box and define your "audit unit"
Choose the review unit: a business unit, client service line or set of high-impact processes. Estimate discovery, analysis and decision time from the number of workflows, access to evidence and reviewer availability. Agree a stopping point and unresolved-evidence handoff; the ten steps are not a delivery-time commitment.
2) Build the inventory (yes, even the messy bits)
Create an inventory record for every AI tool/workflow you find. Minimum fields:
- Name + vendor
- Business purpose
- Where it runs (browser, desktop, internal app, vendor platform)
- Data inputs (including whether client data is used)
- Outputs (who consumes them, where they go)
- Automation level (drafting vs decisioning)
- Named owner
Record unowned systems as a decision gap. Assign someone with authority to investigate and pause the use; a name on the inventory does not itself establish effective ownership.
3) Classify risk using a simple tiering model
One initial screening scheme uses three tiers; adapt the boundaries to the task, data, automation and applicable requirements:
- Tier 1 (low): internal productivity, low sensitivity, no client impact.
- Tier 2 (medium): supports client work, uses sensitive data, or influences operational outcomes.
- Tier 3 (high): client-facing, automated decisions, regulated outcomes, or material financial impact.
Use a tier to prioritise review, then assess the actual controls needed for the workflow. An internal screening tier is not a regulatory category or a finding that the use is safe.
4) Check the data posture (confidentiality, retention, lawful basis)
For each item, answer:
- Does data leave your environment? Where to?
- Is it stored or used for training by the vendor?
- What is logged (prompts, outputs, documents)?
- What is the retention policy, and is it aligned to your obligations?
Record unanswered data-flow questions as gaps. Pause any use that depends on unresolved data permissions or controls, and assign the missing evidence to a responsible reviewer.
5) Check security and misuse risks (especially prompt injection and data exfiltration)
For GenAI workflows, look for:
- Unrestricted tool access (email, file systems, CRMs) without approval gates
- RAG systems pulling from mixed sources without provenance or tenancy boundaries
- PII leaking into prompts or logs
- No kill switch or rollback process
6) Check quality and drift risks (does it stay correct over time?)
Define how quality changes would be detected: suitable test cases, reviewed samples, reported failures or monitoring. Record a baseline, review owner and trigger for reassessment. Sampling design and frequency depend on the task and do not guarantee that all errors will be found.
7) Check applicable requirements and decision evidence
Select records that support the requirements and decisions in scope, for example:
- Risk assessment
- Approved use-case and constraints
- Testing notes (including edge cases)
- Monitoring plan
- Incident response runbook
Where evidence is missing for a sensitive or high-consequence use, record the gap, assess whether that use must pause and assign a reviewer. The proposed tiers do not determine the legal significance or cost of an incident.
8) Map waste (subscriptions, duplicate tools, manual rework)
For the cost and rework review, track:
- Unused or overlapping subscriptions
- Time lost to prompt fiddling and rework
- Hidden costs (staff workarounds, poor outputs, escalations)
Compare subscription cost, review effort and rework with a recorded baseline. Report savings or risk changes only where the review evidence supports them; the exercise may instead identify additional control costs.
9) Make decisions: stop, fix, formalise, or scale
For each item, choose one:
- Stop: unacceptable risk, no owner, or no value.
- Fix: valuable, but missing controls (data/security/monitoring).
- Formalise: write the policy, assign ownership, document how it is used, and set a review cadence.
- Scale: consider a defined expansion only after relevant testing, control evidence, measured benefit and required approvals are in place.
10) Set review ownership and change triggers
Set a review interval proportionate to the workflow and requirements; monthly or quarterly meetings are examples. Reassess affected uses after material vendor, data, configuration or incident changes. Name the person responsible for retiring a tool and checking that its access and data arrangements are closed out.
A simple checklist you can use tomorrow
- Do we have a complete inventory of AI tools and workflows (including shadow AI)?
- Does each item have a named owner and a risk tier?
- Do we know where data goes and what is retained?
- For Tier 2/3, do we have testing notes, monitoring, and a kill switch?
- Have the responsible reviewers checked the applicable requirements and unresolved evidence gaps?
- Have we compared duplicate tools, unused seats and rework against a cost baseline, including any new control costs?
Conclusion
Finish the scoped review with a decision record: which uses are paused, which controls need work, who owns each action and what evidence is needed next. Keep unreviewed teams and workflows explicit so that a bounded review is not presented as an organisation-wide finding.
For a scoped review, confirm the current service, scope and terms before following the AI Clarity Consultation. Confirm any proposed deliverables and timing with the provider. The ten-step example does not establish the scope, price or availability of that offer.
Need a controlled AI route for a regulated team?
If this issue affects a live financial, insurance or advisory workflow, we can help you separate useful AI opportunities from conduct, data, supplier and governance risks.