AI opportunity audit
Find the right starting point for AI.
Before deploying agents, we identify where they should actually operate. Get the findings, opportunities, controls, and roadmap for one business function.
- The operation. A map of how the function works today.
- The findings. What the map shows about where effort and risk sit.
- The opportunities. Each workflow, classified and ranked.
- The plan. A sequence you can start on.
- The business case. What the change is worth, measured against a baseline from your own operation.
Audit my operation.
Start with the workflow you want to improve. If an audit is the right next step, we agree the scope together before implementation.
What the audit examines across one business function.
- WorkflowsSystemsHuman decisionsExceptionsManual interventionsData dependenciesRisk and controlsOperational cost
When to start
Start where work gets stuck.
The audit begins with the business, not with ideas about AI. These are the signs that a function is ready for it.
People spend their time gathering context from several systems before they can decide anything.
The same decisions come round again and again, and each one is made by hand.
The same exceptions come back every week, and the knowledge to handle them sits with a few people.
Approvals stall because the person approving cannot see the evidence quickly.
Reconciliation and reporting are assembled manually from exports and spreadsheets.
Work moves between teams by email, and nobody can say where a case is.
Every process depends on switching between systems that do not talk to each other.
Cycle time is long, and the cause is waiting, not work.
What we examine
Eight things we look at across a business function.
We follow the work as it really happens, through your systems and your teams, and write down what we find.
Workflows
The steps a piece of work goes through from trigger to done, and how long each one takes.
Systems
Every system a workflow touches, what it holds, and how work moves between them.
Human decisions
Where a person decides, what they look at to decide, and how often the answer is the same.
Exceptions
The cases that fall outside the normal path, how they are found, and how they are resolved.
Manual interventions
The copying, checking, chasing and re-keying that keeps the process running.
Data dependencies
Which data each step needs, where it lives, and how reliable it is when it arrives.
Risk and controls
The controls that exist today, the ones that only exist on paper, and what an agent would be allowed to touch.
Operational cost
Where the effort goes, so that the value of changing a workflow can be measured against a baseline.
Classification
Every step is sorted into one of three kinds.
The audit does not assume that everything should become an agent. It decides which steps should not.
Rules and conventional automation. If a fixed rule can decide the step every time, a fixed rule runs it.
Run by: conventional automation
Contextual judgment, adaptive execution, research and orchestration. The step needs information gathered from several places, or a decision that varies with the case, inside limits we can define.
Run by: an agent, inside controls
High-accountability, strategic, sensitive or not yet bounded judgment. People decide. Agents may gather the evidence.
Run by: a person
Deterministic work belongs to conventional automation, which is cheaper and easier to trust than an agent. Agents are for the steps in between, and people keep the steps where accountability has to sit with a person.
What you receive
Not another AI strategy presentation. An executable plan.
The audit connects workflow findings, opportunities and controls, and an implementation roadmap.
The operation
A map of how the function works today.
- Systems graph and department map
- Workflow inventory
- People and handoff map
The findings
What the map shows about where effort and risk sit.
- Where effort goes: coordination, gathering, review, exceptions and strategic work
- Exception taxonomy
- Broken integrations and bottlenecks
- Risk register and system fragmentation
The opportunities
Each workflow, classified and ranked.
- Deterministic, agentic or human
- Complexity, risk, value and dependencies
- Recommended level of autonomy for each
The plan
A sequence you can start on.
- Priority workflows and deployment sequence
- Required integrations, risk and control requirements
- 30, 60 and 90 day and six month roadmap
The business case
What the change is worth, measured against a baseline from your own operation.
- Time recovered and cycle time
- Error reduction and revenue impact
- Working capital impact where it applies
- Deployment cost and projected return
A sample
What an audit looks like.
A short extract from an audit of accounts payable exceptions. It shows the shape of the work, and it is not a real customer.
Illustrative sample. The company, the function and every finding are made up to show the format. Real audits use your systems and your baseline.
Workflow map: handling an invoice exception
Each step, who does it today, the systems it touches and how it is classified.
| Step | Who today | Systems | Kind | Finding |
|---|---|---|---|---|
| Invoice arrives | AP clerk | Email, Shared drive | Invoices arrive by email and by portal. Reading and filing them follows fixed rules. | |
| Match to the purchase order | AP clerk | ERP | Matching an invoice to an order and a receipt is rule based. It can run without an agent. | |
| Investigate a mismatch | AP specialist | ERP, Email, Vendor portal | Each exception needs evidence from several places: the order, the goods receipt and the vendor thread. This is where the time goes. | |
| Decide the adjustment | Controller | ERP | A quantity or price adjustment changes what the company owes. The decision stays with the person accountable for it. | |
| Post the adjustment | AP specialist | ERP | Once approved, writing the correction and recording the reason is routine, and can be done inside limits. | |
| Reply to the vendor | AP specialist | The reply follows the outcome and quotes the evidence. Each message is drafted from an approved template. | ||
| Release the payment | Treasury | Banking system | Moving money stays with people, whatever else changes. |
Opportunity matrix
Each workflow ranked on complexity, risk and value, with what it depends on and where to start.
| Workflow | Kind | Complexity | Risk | Value | Depends on | Start at |
|---|---|---|---|---|---|---|
| Invoice intake and filing | Low | Low | Medium | Mailbox and shared drive access | ||
| Purchase order matching | Low | Low | Medium | ERP read access | ||
| Exception investigation | Medium | Low | High | ERP, receipts and vendor thread, read only | ||
| Quantity and price adjustments | Medium | High | High | Adjustment limits and a named approver | ||
| Vendor replies | Low | Medium | Medium | Approved reply templates | ||
| Payment release | Low | High | Low | Banking controls |
Autonomy recommendations
The level each agentic workflow should start at, and the ceiling we would set for it.
| Workflow | Start | Ceiling | Because |
|---|---|---|---|
| Exception investigation | The agent only reads. Its findings can be checked against what people decide, so authority can grow as the record builds. | ||
| Quantity and price adjustments | The action changes a balance. It starts with a person approving each one, and only routine, low-value cases can ever run alone. | ||
| Vendor replies | Replies go outside the company, so a person approves them until the approved templates cover the ordinary cases. |
Roadmap
The sequence, in the order the work would happen.
30 days
Agree the scope and the baseline
- Confirm the workflows in scope and who owns them
- Measure today: effort, cycle time, exceptions
- Agree read access to the ERP and the mailbox
- Build the sandbox from historical cases
60 days
Run in shadow
- Process live cases beside the team, with nothing executing
- Review every difference between the agent and the team
- Name approvers and set thresholds with the controller
90 days
Supervised production
- The agent prepares adjustments and replies
- The controller approves each one, with the evidence in front of them
- Record every approval and every override
6 months
Review and extend
- Review the record against the requirements for the next level
- Decide, with the workflow owner, what authority the agent has earned
- Add the workflows next to it, on the same controls
Who it is for
For the leader who owns an operation.
The audit works best when there is one function to look at and someone who can open up the systems.
It is written for the COO, CFO, CRO, head of operations or head of a function who carries the cost of work that gets stuck. You do not need an AI strategy, or a view on which model to use. You need a clear picture of your own operation.
We start with a conversation about where work stalls. If the audit is the right next step, we agree its scope with you before it begins.
Functions we usually start with
After the audit
A plan for a focused implementation.
The audit ends with a sequenced plan, so you know what to expect before you commit to anything. If you decide to deploy, the plan becomes the first stages of the deployment path: sandbox, shadow and supervised.
If the audit finds that a workflow should stay with fixed rules or with people, it says so.
How we deployFAQ
A few questions, answered.
Recurring work with documents, records, approvals or exceptions. The audit separates useful AI opportunities from work better handled by rules or people.
We plan around the tools you already use. The audit identifies the connections, data and access an implementation would need.
A workflow owner, the people doing the work and someone responsible for system access. We agree permissions and human approval points together.
The first conversation identifies the operation and the questions to explore. We agree the audit scope before work starts; its findings guide an implementation proposal.