AI agent ROI is calculated as:
ROI = (realised financial benefit − full agent cost) ÷ full agent cost × 100
The difficult part sits inside those two inputs. A generated answer is not a benefit, and token spend is not the full cost. A defensible model counts accepted workflow outcomes, verifies the resulting business state, includes human review and recovery, then converts only realised value into money.
That gives leadership a decision it can act on: expand the workflow, repair the operating design or stop funding it.
Why headline AI agent ROI figures cannot become your business case
Salesforce’s August 2026 survey of 2,025 agentic AI decision-makers reported that the 30% already in production reached meaningful ROI in about eight months. Respondents also reported 53% employee adoption and a 29% average lift in customer satisfaction.
A January 2026 Forrester Total Economic Impact study commissioned by Microsoft modelled 120% ROI and a 15-month payback period for a composite 10,000-person company. Its three-year model included $44.5 million in risk-adjusted benefits and $20.2 million in costs.
These figures show that buyers are measuring returns across adoption, revenue, labour, external spend and operating cost. They do not supply a transferable forecast. Salesforce’s outcomes were self-reported. The Forrester study was vendor-commissioned and built around a large composite organisation using Microsoft products.
Use external benchmarks to challenge an assumption, set a range or explain why the return takes time. Your investment case still needs the volume, cost and verified outcomes of the workflow being funded.
Start with one accepted workflow outcome
Choose a unit that survives contact with the operating system.
“Customer support” is too broad. One billing query resolved without repeat contact is countable. A sales agent produces value when an opportunity meets the agreed qualification rule, reaches the CRM and is accepted by the account owner. An accounts-receivable workflow earns credit when cash or a verified payment commitment changes the ledger state.
Anthropic’s agent evaluation guidance makes the same distinction technically: a transcript can claim that a flight was booked, while the outcome is whether the reservation exists in the environment database. ROI needs that end-state discipline.
Define five fields before modelling the return:
| Field | Required decision |
|---|---|
| Unit of work | The bounded case the agent handles |
| Accepted outcome | The destination state that proves completion |
| Baseline | Current volume, handling time, cost and quality |
| Value basis | The financial mechanism attached to each accepted outcome |
| Measurement window | The period long enough to capture adoption, exceptions and downstream effects |
This boundary prevents activity from masquerading as value. Messages, tool calls and documents describe production. The CRM, ERP, support platform or financial record confirms whether production changed the business.
Build the full cost denominator
Separate one-time investment from recurring operation. That keeps payback visible and stops implementation work from disappearing inside an annual percentage.
One-time investment
Include workflow discovery, source preparation, integration, permissions, evaluation design, deployment, training and recovery testing. Internal subject-matter expert time belongs here alongside supplier fees.
The Forrester composite illustrates the scale of costs hidden behind platform pricing. Its three-year model assigned $15.1 million to agent development, $2.6 million to planning, deployment and ongoing management, and $2.5 million to subscriptions and consumption. The exact numbers fit a 10,000-person composite, but the composition is the useful lesson: technology consumption was one part of the cost base.
Recurring operating cost
Record these lines monthly:
- model, retrieval and paid tool usage;
- platform licences, infrastructure, storage and observability;
- active human review and approval time;
- exception handling, retries and failed runs;
- correction, reversal and incident work;
- source maintenance, regression testing and release support;
- workflow ownership and reporting.
OpenAI’s Agents SDK records generations, tool calls, hand-offs and guardrails as traces and spans. That execution evidence can attribute usage and failure to a run. The ROI record must join the trace to reviewer effort and the final state in the business system.
A cheap run can still be an expensive outcome. Weak retrieval may force a specialist to reconstruct the case. A failed write can trigger manual reconciliation. An agent that finishes quickly but creates repeat contact transfers cost into the next queue.
Convert benefits only when the value mechanism is real
Time saved is operational evidence. Cash benefit appears when the company changes what happens with that time.
Use four benefit classes and state the conversion rule for each:
- Avoided cost: a contract, software line or planned hire is removed from the budget.
- Released capacity: paid hours become available for other work, without claiming a cash saving.
- Incremental contribution: additional revenue is counted at contribution margin after fulfilment cost.
- Loss avoided: a measured reduction in rework, errors, concessions, bad debt or another recurring loss.
Revenue uplift deserves particular care. Use gross profit or contribution margin rather than top-line revenue, subtract any delivery cost created by the increase, and require a credible comparison period or control. A faster response has commercial value only when the team can connect it to conversion, retention or another accepted metric.
The Forrester model applies this discipline in one useful place: it converts modelled revenue gains into net income using a margin assumption. Its labour benefit also comes from positions opened by leavers that were not refilled, rather than treating every saved hour as payroll reduction.
Keep capacity and cash in separate columns. Leadership can still fund an agent to release scarce specialist time, but the approval should rest on that operating choice rather than an inflated savings claim.
Use three equations, not one percentage
A single ROI percentage hides timing and workflow quality. Calculate the following together.
1. Cost per accepted outcome
Cost per accepted outcome = total operating cost ÷ accepted outcomes
The denominator excludes drafts, abandoned runs and outputs rejected after review. Compare it with the current cost for the same completed unit.
2. Payback period
Payback months = one-time investment ÷ monthly net realised benefit
Monthly net realised benefit equals monthly realised benefit minus recurring operating cost. When that number is zero or negative, payback has not started.
3. ROI for the measurement window
ROI = (realised benefit − one-time and recurring cost) ÷ one-time and recurring cost × 100
State the window beside the result. A three-month pilot ROI and a three-year programme ROI answer different capital questions.
A worked example exposes the assumptions
Consider an internal commercial-operations workflow processing 1,000 cases each month.
The current process takes 15 minutes per case at a fully loaded labour rate of £40 per hour, creating a £10 baseline handling cost. The agent reaches an accepted outcome on 650 cases. Human review averages three minutes for those accepted cases, while exceptions and corrections add £900 in monthly labour. Model, tool and platform spend is £1,200 per month. Ongoing maintenance and monitoring cost £1,000. Implementation cost £24,000.
The monthly calculation is:
| Line | Calculation | Value |
|---|---|---|
| Baseline labour for 650 accepted cases | 650 × £10 | £6,500 |
| Review labour | 650 × 3/60 × £40 | £1,300 |
| Exceptions and corrections | measured monthly | £900 |
| Model, tools and platform | provider usage | £1,200 |
| Maintenance and monitoring | allocated monthly | £1,000 |
| Total recurring agent cost | sum of recurring lines | £4,400 |
| Gross released capacity | baseline minus recurring cost | £2,100 |
The workflow releases £2,100 of monthly capacity, but that is not automatically a £2,100 cash benefit. If the company uses the capacity to absorb growth and avoids £1,500 of planned monthly contractor spend, the realised benefit is £1,500. Recurring cost still exceeds that cash benefit, so the implementation does not yet pay back on cost avoidance alone.
If accepted cases also produce £4,000 of measured monthly contribution margin, total realised benefit reaches £5,500. Monthly net benefit becomes £1,100, creating a payback period of 21.8 months before financing or discounting.
That answer is less flattering than a calculation that values every saved minute as cash. It is also usable. The team can now see whether acceptance, review burden, case value or implementation scope needs to change.
Check whether reviewers can deliver the forecast volume
A team with 20 review hours cannot accept a forecast that requires 40. Start with the time the named reviewers can actually reserve, then work back to the volume whose benefit the business can realise.
Microsoft’s queue-based load levelling guidance explains the underlying constraint: when the average producer rate exceeds the consumer rate, the queue grows and latency increases. A review queue has the same capacity problem. Buffering drafts absorbs a burst; it does not create more expert attention.
For a segment where every case requires human review, calculate:
Review workload per case = initial review minutes + (share requiring another review × additional review minutes)
Reviewed-case ceiling = available reviewer minutes ÷ review workload per case
Accepted-outcome ceiling = reviewed-case ceiling × acceptance rate
The following inputs are illustrative assumptions, not client results or industry benchmarks. Assume a team reserves 20 hours each week for review. Each case takes four minutes initially; 10% require one additional eight-minute review. After those reviews, 90% of completed cases meet the acceptance criteria.
| Capacity input or result | Calculation | Weekly result |
|---|---|---|
| Available reviewer time | 20 × 60 | 1,200 minutes |
| Review workload per case | 4 + (0.10 × 8) | 4.8 minutes |
| Maximum completed reviews | 1,200 ÷ 4.8 | 250 cases |
| Maximum accepted outcomes | 250 × 0.90 | 225 outcomes |
| Review time for 500 cases | 500 × 4.8 ÷ 60 | 40 hours |
Producing 500 drafts would demand twice the available review time. Under these assumptions, reviewers can accept at most 225 outcomes per week. Absence and uneven arrivals reduce that ceiling; the forecast still requires a test with real cases. Recognise benefit from observed completions and account for any justified automatic route separately.
Record active review time separately from elapsed waiting time. The former enters labour cost; the latter affects whether an otherwise acceptable result arrives before its deadline. Add queue age and overdue cases to the pilot success criteria, so a fast agent cannot pass while its work expires awaiting approval.
If capacity is insufficient, lower intake or reserve more qualified review time. Improve the evidence packet when reviewers spend time rebuilding the case. Change the human approval boundary only after testing the relevant action and consequence; a backlog is no justification for bypassing an approval.
Before buying more delivery capacity, name the internal person who will accept the extra work. The consultancy versus in-house ownership test helps separate a build bottleneck from an approval bottleneck. Supplier capacity cannot compensate for an unavailable decision owner.
Run sensitivity before approving the investment
Model a downside, base and upside case. Change the variables that production can actually move:
- eligible monthly volume;
- accepted-outcome rate;
- active review minutes;
- exception and reversal rate;
- value per accepted outcome;
- model and tool cost per run;
- adoption among intended users;
- implementation and maintenance cost.
Do not hide risk inside one arbitrary discount. Show which assumption breaks payback.
The Salesforce study gives two useful operating signals for sensitivity design. Clean, accessible data and tightly bounded use cases were the strongest reported success factors for autonomous deployments. Companies that unified relevant data before deployment reported meaningful ROI in 7.3 months, compared with 8.8 months for those addressing data gaps afterward. Those are survey associations, not causal guarantees, but they justify testing data-readiness and scope assumptions directly.
Review ROI by segment before expanding authority
Pilot averages become less reliable as scope widens. The cleanest cases enter first; expansion introduces weaker evidence, rare exceptions and actions with larger consequences.
Segment the economics by case type, customer group, source quality, action, agent version and review route. Then inspect the marginal cases entering the workflow. Another 10% of volume can require stronger reasoning and specialist review while producing less value per case.
Use this expansion gate:
- Did realised value exceed full cost for the segment?
- Which exceptions consumed skilled attention?
- How many accepted actions needed correction or reversal?
- Did the workflow remove work or move it into another queue?
- Which source, tool or release changed the result?
- Can finance reproduce the calculation from stored evidence?
A dashboard number without traceable inputs turns into a negotiation between functions. A case-level economic receipt preserves the basis for the decision.
Keep one economic receipt per case
Each completed case should retain:
- workflow, case and operation identifiers;
- agent, model, prompt and tool versions;
- source evidence and permissions used;
- execution cost, retries and elapsed time;
- evaluation result and human review time;
- proposed, accepted, rejected and reversed actions;
- recovery work and owner;
- verified outcome, timestamp and destination system;
- financial value rule applied to that outcome.
Review the receipts weekly during rollout. Fix the largest source of economic drag before widening scope. The constraint may be oversized context, a missing integration, an approval packet that forces reconstruction or an action that remains expensive to reverse.
Model Operator is a Dubai-based AI product studio, personally led by Alexander Chrisostomou. Its AI Initiative Consulting engagement maps where AI should be used and how the work will enter the team’s operating routine. For an ROI decision, that means tracing a proposed workflow through to the person who accepts its output and the financial outcome that justifies the investment.
Bring one recurring workflow, its current baseline and the outcome finance already recognises. Start a Model Operator build conversation or email alexander@modeloperator.io to turn it into an implementation and measurement plan.