The AI Agent Business Case Is a Queue, Not a Headcount Cut

A finance-grade calculator for deciding whether one workflow should receive agentic investment

Altivate | Published 21 June 2026 | Updated 26 July 2026

1. Start with the queue the business already owns

Most AI agent cases begin in the wrong unit. They start with headcount, model capability or the number of tasks an agent might perform. Finance cannot govern any of those as a benefit.

Start with one queue:

• supplier invoices waiting for resolution;
• service cases waiting for evidence and routing;
• orders waiting for an exception decision;
• maintenance work orders waiting for triage;
• contracts waiting for first review;
• account records waiting for validation.

A queue has observable economics. Work arrives at a rate. Some cases follow a standard path, some need judgment and some must never leave accountable human control. Cases wait, age, get reopened, escalate and occasionally cause loss. People spend time on them, but time is only one part of the cost.

This is the first discipline of the investment case:

The unit should be narrow enough that a process owner can answer four questions from evidence:

1. How many cases arrive?
2. What happens to each case today?
3. What outcome tells us the case is complete and correct?
4. What is the consequence when the outcome is late or wrong?

Do not write “finance agent” or “customer-service agent.” Write “classify an inbound invoice exception, retrieve the relevant purchase order and policy, draft the proposed resolution, and route it to the authorised approver.” That description exposes the reads, actions, owner and residual human work that the calculator needs.

Use RAG, tools, workflow or agent to select the simplest architecture, then apply the Enterprise AI Agent Readiness Test to each action the system may take.

2. The calculator has five evidence blocks

The calculator below is a worksheet, not a forecast. Every operating input is blank by design, and its completeness gate refuses to show NPV or financial approval until each required input is explicit. Use an observed baseline, a contracted price or a bounded assumption with an owner and review date.

Workload and baseline

Do not multiply all handling time by labour cost and call it savings. First establish what happens to released capacity. If work disappears but headcount, overtime, contractor spend, backlog and service output do not change, no financial benefit has been realised.

Addressable scope

Eligible cases

annual cases x in-scope share x data-ready share x adoption

The product of these inputs is often much smaller than the original volume. That is useful. It prevents a business case from taking credit for work the first release cannot reach.

Agent performance and residual work

Measure the outcome, not the agent’s claim. Anthropic’s evaluation guidance makes this distinction directly: a transcript may say that a booking was made, but the outcome is whether the reservation exists in the system. That is vendor guidance, and it is also sound measurement practice.

The correct-accepted, exception and accepted-error rates must be mutually exclusive and total 100% of eligible cases. This prevents an exception rate from being deducted once in volume and again inside a per-case time formula.

Benefit economics

Keep benefit classes separate:

Baseline annual workload

eligible cases x baseline handling minutes

Future annual workload

correct accepted cases x review minutes + exception cases x exception handling minutes + accepted-error cases x (review minutes + rework minutes)

Released hours

(baseline annual workload - future annual workload) / 60

If future workload is equal to or greater than baseline workload, released hours and capacity value are zero or negative. Do not hide that result.

Capacity value before realisation

released hours x loaded hourly cost

Recognised capacity benefit

capacity value before realisation x finance-approved realisation factor

The realisation factor is not a technology input. It expresses what the organisation will do with the capacity. It can be zero. A zero is more credible than pretending all time saved becomes cash.

Full cost and failure exposure

Expected annual failure loss

annual agent actions x material failure probability x average loss per material failure

This is not permission to hide severe tail risk inside an average. A low-probability action with a catastrophic consequence needs a hard authority boundary, not a favourable expected-value cell.

3. Turn the worksheet into three scenarios

Do not debate one forecast. Compare a downside, base and upside case using the same formulas.

Then calculate:

Annual net benefit

recognised benefits - annual run cost - annual control cost - expected failure loss

Simple payback in months

initial investment / positive monthly net benefit

Payback is not defined when monthly net benefit is zero or negative. Report the result as “no payback in the modelled horizon,” not as an error or an infinite-looking number.

Net present value

sum of discounted annual net benefits - initial investment

Use the organisation’s approved discount rate and planning horizon. Do not extend the horizon simply to make the case positive. Model refresh, platform change, re-evaluation and process drift as real operating costs.

The downside case should be uncomfortable but plausible. It should assume slower adoption, more review, lower completion, higher exceptions and a longer stabilisation period. If the investment only works when every uncertain input moves in the favourable direction, it is not ready for approval.

Worked scenario

Assume 20,000 eligible cases at 18 baseline minutes and a loaded rate of AED 180 per hour.

These are arithmetic examples, not Altivate benchmarks. Apply the finance-approved realisation factor, annual run and control cost, and expected failure loss before calculating net benefit.

Download the polished AI agent business-case workbook with an executive scorecard and scenario charts, or use the open CSV model.

4. The business case is an operating contract

A good investment case assigns each benefit and cost to someone who can cause or verify it.

This table prevents the technology team from owning a benefit it cannot realise. Product teams can deliver a capable system. Only the business can change work allocation, retire cost, increase throughput or accept a different risk position.

OpenAI’s current use-case guidance recommends collecting specific work opportunities and comparing impact with effort. Anthropic recommends the simplest sufficient architecture because agentic complexity adds latency and cost. Both are vendor guidance. The finance implication is broader: every layer of complexity needs an incremental value claim and an owner.

If a deterministic workflow creates 80 percent of the expected value at half the run and control cost, the agent does not win because it is more advanced. It loses the business case.

5. Benefits that do not survive scrutiny

“Hours saved”

Ask what changes after the hours are released. Accepted answers include lower overtime, avoided hiring, increased completed volume, reduced backlog or reassigned capacity with a measurable output. “People will do higher-value work” is an intention until that work, owner and measure are named.

“Fewer errors”

Define the error, current incidence, detection point and unit consequence. Also measure new failure classes created by the agent, including plausible but unsupported decisions, unauthorized action and correlated failure at scale.

“Faster cycle time”

Identify the step the agent changes and the outcome affected by time. A two-hour reduction matters only if it changes cash, service, capacity, loss or a binding deadline.

“Better employee experience”

Measure task abandonment, rework, active use, override behaviour and time spent correcting output. A satisfaction survey can supplement operating evidence, not replace it.

“Strategic option value”

Option value can justify a bounded learning investment. It cannot be presented as realised ROI. Record the decision the pilot will enable, the evidence required and the expiry date of the option.

6. Put a stop threshold in the approval

The pilot should be authorised with a pre-agreed stop or redesign threshold.

Example structure:

• stop if material authority violations exceed zero;
• stop if the correct-outcome rate stays below [enter] after [enter] evaluated cases;
• redesign if exception handling consumes more than [enter] minutes per eligible case;
• redesign if adoption stays below [enter] after the agreed change actions;
• do not scale if the downside-case payback exceeds [enter] months;
• do not expand tools or actions until the current action passes its evaluation and control gates;
• re-open the investment case if model, platform, price, process or policy changes materially.

This protects the organisation from “pilot momentum,” where continued spending becomes the reason to continue spending.

It also improves learning. A stopped pilot can still be a successful investment if it resolves an important uncertainty cheaply and prevents a larger loss.

7. The executive sign-off checklist

Before approval, the sponsor should be able to answer yes to each statement:

• ☐ One workflow and one accountable process owner are named.
• ☐ The current queue, handling, cycle, rework and failure baseline are evidenced.
• ☐ The eligible scope excludes cases the first release cannot safely reach.
• ☐ A simpler workflow or tool-based option was costed.
• ☐ Correct completion is measured from final state, not agent self-report.
• ☐ Residual review and exception work are included.
• ☐ Capacity has an accepted conversion mechanism.
• ☐ Each quality, cycle, revenue or risk benefit has a causal formula.
• ☐ Initial, run, control, change and product-ownership costs are included.
• ☐ Material failure exposure and non-negotiable authority boundaries are recorded.
• ☐ Downside, base and upside scenarios use the same formulas.
• ☐ The pilot has a stop threshold and a dated investment review.

If several boxes remain open, approve discovery, not deployment.

8. A better board question

The weak question is:

The better question is:

That question can produce “not yet,” “use a workflow,” or “pilot one action.” Those are investment decisions, not failures of ambition.

An agent earns scale when its evidence improves the business case while the operating controls remain intact. Until then, fund the next uncertainty, not the largest story.

Continue with How to Start with AI Agents Without Starting with an Agent, or return to White Papers.

Sources and verification status

Checked 26 July 2026. Vendor guidance is identified as such and is not used as an independent ROI benchmark.

• OpenAI, Identifying and scaling AI use cases – vendor guidance on workflow discovery and impact/effort prioritisation.
• OpenAI, A practical guide to building agents – vendor implementation guidance on agent components, guardrails and intervention.
• Anthropic, Building effective agents – vendor guidance on workflows, agents and the cost/latency trade-off of complexity.
• Anthropic, Demystifying evals for AI agents – vendor guidance on tasks, trials, traces, graders and outcomes.
• NIST, AI Agent Standards Initiative – current standards initiative covering security, identity, interoperability and evaluation.