How we grade AI deployments
Every Tuesday, Deploy Grade grades one real AI deployment. Each one is scored 0–4 on five criteria, the scores are weighted, and the result maps to a letter. Every grade is Deploy Grade’s opinion, based only on the public evidence available on the date we publish.
The five criteria
| Criterion | Weight | A 4 looks like | A 0 looks like |
|---|---|---|---|
| Evidence | 30% | Audited or regulator-filed numbers (e.g., SEC filings), consistent across sources | Anonymous or unsourced claims |
| Measured outcome | 25% | Defined metrics with a baseline, period and trend | No measurable outcome |
| Cost and payback transparency | 20% | Cost of the system and dollars saved or earned, enough to estimate payback | No cost information |
| Durability and risk controls | 15% | Human fallback, monitoring, published failure modes, results sustained 12+ months | No controls disclosed, or known failures unaddressed |
| Transferability | 10% | A typical 10–500-person company could copy it with off-the-shelf tools | Requires unique scale, data or budget |
From score to letter
Weighted score = 0.30 × Evidence + 0.25 × Outcome + 0.20 × Cost and payback + 0.15 × Durability + 0.10 × Transferability.
| Letter | Weighted score |
|---|---|
| A | 3.4 or higher |
| B | 2.6 to 3.39 |
| C | 1.8 to 2.59 |
| D | 1.0 to 1.79 |
| F | below 1.0 |
Worked example (Deploy Grade #1, Klarna): Evidence 4, Outcome 3, Cost and payback 2, Durability 2, Transferability 2. That’s 1.20 + 0.75 + 0.40 + 0.30 + 0.20 = 2.85, a B.
What a grade means, and what it doesn’t
A low grade means the public evidence is weak. It doesn’t prove the system performs badly. Companies can raise their grade by publishing better data, and we’ll re-grade and note the change.
We don’t contact graded companies before publication, and no company pays for or reviews a grade. Sponsors never touch grades. Grades are not investment advice.
Sourcing standard
- Every number links to a primary source where one exists: filings, official changelogs and pricing pages, the company’s own announcements.
- Vendor press releases are labelled as vendor claims.
- Anything we couldn’t verify is marked unverified. Anything not public is marked not disclosed.
- Our own calculations are labelled “Deploy Grade calculation” and show their inputs.
- We never invent numbers, quotes or testimonials.
Corrections
Spot an error? Write deploygrade@pm.me. We fix errors in the web version and note the fix at the top of the issue.