The DN Alpha Thesis: the winner is the workflow with a short path to realised value

The usual automation pitch begins with a task volume and a claim about minutes saved. A credible business case begins with the constraint those minutes relieve. If customers wait because a support queue is understaffed, freed capacity may improve service. If a finance team pays contractors to process routine documents, accepted automation may reduce a specific bill. If employees simply finish an unchanged workload sooner, the benefit may be real but remain a capacity gain rather than a cash saving.

That distinction changes which deployments deserve funding. A technically impressive agent can struggle commercially because its output requires expensive review, its errors create costly exceptions or the organisation cannot use the capacity it frees. A modest workflow can pay back sooner when the acceptance rule is clear and the avoided expenditure is visible.

DN proposes indexing workflow economics within industries before comparing industries. “Accounting” is too broad a unit: extracting invoice fields, approving supplier payments and completing a monthly close have different authority, review and failure requirements. Preserve those differences in the evidence record.

What the research supports, and what it cannot tell you

These sources justify testing the local workflow instead of importing a headline percentage. They do not establish which industry has the highest agent ROI. DN’s inference is that task mix, review effort and benefit realisation belong in every deployment record. The proposed framework below is original editorial methodology. [1] [2] [3]

The industry workflow matrix

This is a measurement map, not a ranked performance table. DN has no verified cross-industry ROI sample in this edition. Each row describes what a comparable future record would need.

A useful industry comparison would match task complexity, output acceptance, labor rates, review policy and observation period. Without those controls, a short payback in one deployment says little about a different buyer’s likely result.

The DN Realised Value Bridge

Start with baseline human minutes per accepted task. Subtract the human minutes still needed after automation, including review and correction, across the same workload. This produces net freed capacity. Apply the fully loaded hourly labor cost, then apply a realisation percentage representing the share that actually changes financial outcomes.

Add separately evidenced monthly contribution margin from incremental business, not gross revenue. Subtract additional error and incident loss relative to the baseline. Subtract monthly usage, platform, integration maintenance and other incremental operating costs. Keep one-time setup separate.

Capacity hours: volume × (baseline minutes − human minutes with automation) ÷ 60.

If automation increases human time, the calculator charges the added labor at the full hourly rate rather than discounting that cost by the realisation share.

Monthly benefit before operating cost: capacity hours × labor cost × realisation share + attributable incremental contribution margin − additional loss.

Monthly net benefit: monthly benefit − incremental recurring operating cost.

First-year ROI: [12 × monthly net benefit − setup cost] ÷ [12 × recurring operating cost + setup cost] × 100.

Simple payback: setup cost ÷ positive monthly net benefit. This is a steady-state model, without ramp-up, discounting, taxes or working-capital effects. Enter incremental costs, not expenditure that remains identical in the baseline.

A fictional scenario that exposes the capacity-to-cash gap

Consider 1,000 monthly tasks. Baseline human time is 12 minutes each; automation leaves six minutes of human review and correction. That frees 100 hours. At 30 currency units per hour, the capacity value is 3,000. If only 50% becomes avoided expenditure, the realised labor benefit is 1,500.

Add 600 in separately evidenced incremental contribution margin, subtract 100 in additional loss and 500 in monthly operating costs. Monthly net benefit becomes 1,500. With 6,000 of setup cost, steady-state payback is four months. First-year net value after setup is 12,000; first-year incremental cost is also 12,000, producing 100% first-year ROI.

Set realisation to zero while keeping the other inputs unchanged and monthly net benefit falls to zero. The agent still frees 100 hours, but the model no longer recovers setup cost. This is why a dashboard showing hours saved cannot settle the investment case. Every number in this example is invented for arithmetic demonstration.

Which operating model fits the evidence?

No vendor is recommended or assigned LIVE status here. Check current service availability, access restrictions, data-processing terms and the actual product before procurement. This edition’s action is a measured pilot, not a platform signup.

Build an evidence record that survives the finance review

Run baseline and automated work against comparable task populations. Record accepted output counts, human time, recurrence, review effort and errors. Include unsuccessful attempts. Faster handling of easy tasks is not evidence of improved performance on the full workload.

Separate pilot economics from steady-state economics. A pilot may include heavy onboarding; a mature deployment may need sustained monitoring. Report both instead of selecting the cheaper period. Record time to reach production, because a four-month steady-state payback can follow a lengthy implementation delay.

Use invoices or finance-approved records for avoided spending. For incremental margin, retain the attribution method and subtract associated variable costs. If the business cannot distinguish agent impact from seasonality, promotion or staffing changes, label the benefit uncertain and run a conservative scenario with it removed.

Methodology and the future index

Index status: DN-ROI v1.0 is a framework edition. The six-row matrix is a proposed classification dataset, not a database of deployment returns. The external studies are contextual evidence about AI assistance. They are not inputs used to assign industry ROI values.

Future publication rule: only publish sector distributions when comparable deployment records exist. Show sample counts, observation windows, inclusion criteria and uncertainty. Do not rank industries using calculator defaults, unverified vendor claims or model-generated averages. A future median must specify whether it includes failed and abandoned deployments.

Limits: the calculator assumes fixed monthly volume and constant benefits. It excludes ramp-up, financing, taxes, depreciation and terminal value. It is decision support, not a financial forecast or investment recommendation. A reported positive return cannot excuse unacceptable security or quality failures.

Maintenance: review research quarterly and revise each measured record after material workflow, model, pricing or control changes. Change log: 1 October 2026, initial framework, industry matrix and calculator. Corrections: send the claim and supporting evidence through DN’s contact page.

Sources

1. Brynjolfsson, Li and Raymond: Generative AI at Work, NBER Working Paper 31161. This article identifies the working-paper estimate rather than mixing publication versions.
2. METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025.
3. METR: We are Changing our Developer Productivity Experiment Design, February 2026.

Frequently asked questions