Approved partner · Anthropic & OpenAI

Measured uplift,not promised uplift.

Baseline before build, re-measure after
Deliverable: the evaluation harness

S.01 · What We Measure

The operational metric, not the press release.

Every buyer asks the same question: how do you prove the AI delivered value? Nugawi measures AI uplift with a human-verified baseline taken before the build, the same metric instrumented in production after, and the difference reviewed with the client. The evaluation harness ships as a deliverable, so your team keeps grading the system after we leave, and the Agent Audit re-grades systems we did not build. We are a member of the Claude Partner Network and an approved partner of both Anthropic and OpenAI, and the counting rules never depend on the vendor: we deploy the right model for each workflow, not one model for everything.

What is uplift measurement?

The measured change in one operational metric between a human-verified baseline and the same metric after go-live, counted on identical rules. If a number cannot be counted the same way twice, it is not uplift.

What gets counted, and what never does.

S.02 · How It Works

Baseline, instrument, re-measure, review.

For a leading freight forwarder the metric was quote turnaround, counted on the same clock before and after. The rules are fixed before the build and hold whichever model runs.

01

Human-verified baseline

Drawn from real historical workflow data before anything changes: real cases, real timestamps, real costs. Your team verifies it line by line and signs it. If your own people will not sign the baseline, it does not stand.

02

Instrumentation in production

The new path is instrumented so the same metric is counted as live cases run: same fields, same clock, same exclusions, recorded automatically rather than sampled by hand.

03

Re-measurement on the same rules

After go-live the baseline counting rules are applied again to production data. Same metric, same definitions, same exclusions. The delta is arithmetic, not argument.

04

Joint review of the delta

The before and after are reviewed with the client, evidence attached: which cases moved, which did not, and why. You sign the delta with us, not after us.

S.03 · What Ships

The measurement outlives the engagement.

Three artefacts land in your repository, not ours. All three are client property.

01

The evaluation harness

Golden cases drawn from your real decisions, a regression set that starts before cutover, and quality gates wired into the release path. Versioned in your repository and running without us.

02

The measurement report

The baseline, the re-measured numbers, the delta, and how every figure was counted: sources, rules, exclusions and evidence. Written to survive an audit, not a pitch.

03

The capability to re-run it

Your team can re-grade the workflow whenever models, prompts or volumes change. Measurement is not one number in a deck; it is an instrument your operators keep using.

S.04 · How It Connects to Engagements

One counting discipline, every engagement.

The same baseline discipline runs through every Nugawi engagement, so the numbers stack across the whole path.

01

Discovery Sprint

Establishes the baseline: one consequential workflow measured as it runs today, verified by your team before anything changes.

02

Lighthouse Deployment

The 90-day build releases its final tranche against a re-measured baseline, so the last payment tracks measured uplift, not a demo.

03

Agent Audit

For systems already live, the audit re-grades them against the outcomes that matter, with evaluation runs as evidence.

S.05 · Questions Buyers Ask

The questions before the measurement.

How do you measure AI uplift?

Against a human-verified baseline taken before the build. One operational metric of the workflow, usually cycle time, touch count, exception rate or cost per document, is counted on fixed rules from real historical data. After go-live the same rules are applied again to production data, and the difference is the uplift. The rules are written down before the build starts, so the number is arithmetic, not argument.

What if the baseline is disputed?

Then it is disputed before the build, not after it. The baseline is built from your own historical data and verified by your team before anything changes: your operations people check the cases and sign the counting rules. If a number cannot be agreed, the rule is rewritten until your team will sign it, and the build does not start on a baseline your own people reject.

Who owns the measurement?

You do. The counting rules, the evaluation harness and the measurement report are client property: they live in your repository, documented so your team can re-run the measurement without us. We never hold the only copy of the evidence, and no part of the measurement depends on our continued presence.

Does the client keep the harness?

Yes. The evaluation harness ships as a deliverable of every production build: golden cases drawn from your real decisions, a regression set that grows with every defect caught, and quality gates wired into the release path. It is versioned in your repository and runs without us, so the grading continues after the engagement ends.

How does this affect how we pay?

Structure only: engagements can tie the final payment milestone to the re-measured baseline, so the last tranche is released against measured uplift rather than a demo. All figures are quoted on application. Engagement structure: https://nugawi.ai/pricing.

S.06 · Start

Uplift you can check, not trust.

A vendor who cannot say how uplift will be measured before the build starts owns the number at the end. We put the counting rules in writing first: the baseline your team verifies, the same metric instrumented in production, the delta reviewed together, and the harness left behind so the grading continues. Figures are quoted on application, and the payment structure is written into the fixed scope before you commit.