Every agent must earn its place in the accounts
We start from the financial statements, not the technology, and hold each AI agent to the line it is meant to move, at its full cost.
How every agent is judged
The same four financial pillars appear in every engagement, on every industry page and in every board report.
Revenue growth
Conversion, pricing, retention and new services that AI makes possible, counted as gross profit, not headline revenue.
- Revenue growth
- Conversion improvement
- Customer retention
- Gross profit contribution
Operating cost efficiency
Hours, handoffs and rework removed from core processes, net of the full cost of running and supervising AI.
- Cost per transaction
- Process operating costs
- Productivity gains
- Verified expense reductions
Capital productivity
Cash released from inventory, receivables and work in progress, and better returns on the assets already owned.
- Working capital
- Inventory efficiency
- Receivables
- Asset utilisation
Risk reduction
Fewer losses, errors, incidents and compliance failures, counted only when there is evidence to show it.
- Expected operating losses
- Error and remediation costs
- Compliance-related exposure
- Control effectiveness
Revenue | Operating costs | Capital | Risk
For each category: the baselines that matter, where AI can help, what the organisation needs, where value comes from, what it costs, how it is measured and attributed, and the oversight it needs.
Revenue
- Financial baselines
- Revenue by segment and channel, conversion rates, average price and discount, retention and churn, gross margin.
- Potential AI use cases
- Faster quotes and credit decisions, lead qualification, next-best-action prompts, pricing support, proactive service and retention.
- Required capabilities
- A clear growth strategy, joined-up customer data, consistent customer processes, and sales and service teams ready to work with AI.
- Sources of value
- Gross profit from extra conversion, better price realisation and retained customers. Counted as gross profit, not revenue.
- Implementation and recurring costs
- Integration with customer systems, data preparation, model usage, sales enablement and oversight of customer-facing outputs.
- Key performance measures
- Conversion, time to quote or decision, retention, revenue per customer or adviser, price realisation.
- Measurement and attribution
- Controlled roll-out or matched comparison groups wherever possible; otherwise before-and-after against a stated baseline, with other changes noted.
- Risks and human oversight
- Mis-selling, unfair or biased treatment, brand damage and inaccurate commitments. Customer-facing actions need clear limits and human review above thresholds.
Operating costs
- Financial baselines
- Cost by process, hours per transaction, error and rework rates, cost-to-income or expense ratio, outsourcing spend.
- Potential AI use cases
- Document preparation, data entry and checks, reconciliation, claims and case handling, service requests, reporting.
- Required capabilities
- Documented processes, measured cycle times and costs, accessible data, and managers willing to redesign roles.
- Sources of value
- Hours removed from processes, lower error and rework, avoided hiring and lower outsourcing spend.
- Implementation and recurring costs
- Process redesign, integration, model usage, human oversight, change management, validation and assurance.
- Key performance measures
- Hours per transaction, unit cost, error rate, cycle time, cost-to-income or expense ratio.
- Measurement and attribution
- Freed capacity counts as value only when it is redeployed to revenue work or replaces hiring that would otherwise happen.
- Risks and human oversight
- Errors at scale, loss of know-how, over-reliance on one vendor. Sampling, exception queues and kill switches.
Capital
- Financial baselines
- Inventory days, receivables days, work in progress and lock-up, asset utilisation, capital spend, cost of capital.
- Potential AI use cases
- Demand forecasting and replenishment, collections and billing accuracy, maintenance timing, capital project prioritisation.
- Required capabilities
- Reliable operational and financial data, integration with planning and billing systems, finance involvement.
- Sources of value
- Cash released once from working capital, deferred or avoided capital spend, more output from the same assets.
- Implementation and recurring costs
- Data integration, planning system changes, model usage and oversight by planners and finance.
- Key performance measures
- Inventory days, days sales outstanding, lock-up days, asset availability, capital spend against plan.
- Measurement and attribution
- One-off cash release is shown separately from recurring benefits. The financing value is calculated at the cost of capital, not added twice.
- Risks and human oversight
- Stock-outs from over-aggressive targets, customer friction in collections. Service levels set by people, agents work within them.
Risk
- Financial baselines
- Losses, provisions and write-offs, fraud, incidents, regulatory findings and penalties, compliance cost.
- Potential AI use cases
- Fraud and anomaly detection, monitoring and early warning, compliance checks, alert triage, quality inspection.
- Required capabilities
- An AI inventory, defined decision rights, testing and monitoring, security and vendor oversight.
- Sources of value
- Lower losses, fewer incidents and findings, lower compliance cost per case.
- Implementation and recurring costs
- Validation, monitoring, assurance, security controls and specialist oversight.
- Key performance measures
- Loss rates, incidents, false-positive rates, findings, cost per compliance case.
- Measurement and attribution
- Risk benefits are counted only after enough history to show them, typically a full cycle of 12 months or more.
- Risks and human oversight
- Missed risks from false negatives, model drift, adversarial behaviour. Clear escalation and human review of material cases.
An adjustable ledger scenario
Hypothetical illustration. Change the assumptions to explore the method. The results are not a forecast, a quote or a client result.
Five steps from the accounts to the agent
- 01
Read the accounts
We start with the annual report, management accounts and segment results, and compare them with listed peers. Where does the business trail its sector, and which lines carry the most weight?
- 02
Map the value levers
For each line that matters, we trace the operating drivers behind it and the decisions inside those drivers. Only then do we ask where an agent could move the number.
- 03
Build a case for every agent
Each proposed agent gets a one-page case: the line it moves, today's baseline, the expected effect, the full cost of running it, a named owner and the conditions under which it is switched off.
- 04
Decide and sequence
Cases are ranked as a portfolio, by value, time to evidence and risk, and each agent is placed on the delegation map: act, recommend, or inform only.
- 05
Measure and assure
After deployment, results are measured against the baseline with evidence an auditor would accept, and reported to the board as a quarterly ledger. Agents that do not earn their place are stopped.
Two numbers, never confused
Before deployment, the ledger holds an estimate built from stated assumptions. After deployment, it holds realised results measured against the baseline. Both are reported, side by side, and only realised results are claimed.
Business case
Built from the baseline and stated assumptions before an agent is approved. Used to decide and prioritise.
Measured result
Measured against the same baseline after deployment, with evidence an auditor would accept. Reported quarterly to the board.
Rules we hold to
Capacity is not a saving
Freed hours count only when they are redeployed to revenue work or replace hiring that would otherwise happen.
Count the full cost
Model usage, integration, data work, human oversight, validation and assurance all go on the ledger.
No baseline, no agent
If a result cannot be measured before deployment, it cannot be claimed after.
Risk benefits wait for evidence
Lower losses or fewer incidents are counted only once there is enough history to show them.
Roles kept separate
Our advisory work is independent. Where we also lead delivery, or use technology from a company in which we hold an interest, we say so in writing and never assure our own work.
Worked ledgers for each sector
Each industry page includes a hypothetical worked ledger with every assumption shown.
Start with the accounts
Send us your latest annual report. We will show you where agents could move the numbers, and where they should not.