● Deploy anywhereGet API key

Home / Add-ons / Samples

Sample deliverables: what a service engagement actually produces

Short answer. Each Plugsky service ends with a document, not a conversation. This page shows the structure of all three: a cost baseline from onboarding, a quality diff from migration, and a gap register from the compliance audit. The examples are illustrative; the fields and ordering are the ones we use.

Every example below is invented. No customer data, names or figures are used.

1. Onboarding — cost baseline

The baseline is five lines of arithmetic on the customer's own volume, plus a plan fit and a watch list. The point is that a reviewer can redo the sum without asking for a spreadsheet.

Illustrative baseline for a support assistant. Volumes are invented.
InputValueSourceType
Requests per day1,400Gateway logs, 7 daysmeasured
Average tokens per request2,800Sample of 20 callsmeasured
Peak concurrent calls5Gateway logsmeasured
Binding constraintCostCustomerstated
Plan fit and first limit reached.
PlanPriceRequests/dayTokens/dayFits?First cap hit
Hobby$81,0002,000,000NoTokens at ~714 requests/day
Starter$202,0005,000,000YesNeither at current volume
Builder$60no daily capno daily capOver-provisionedConcurrency only

Watch list: (1) tokens/day at 3.92M of 5M — 78% of the cap and rising 4% week over week; (2) requests at 1,400 of 2,000 — 70%, so tokens bind first; (3) peak concurrency 5 of 6, so one more parallel worker would queue.

2. Migration — quality diff report

The diff compares the old and new provider on the same requests, grouped by input class. A single aggregate pass rate hides the class that fails, so the report never uses one.

Illustrative pass rate after replaying 600 redacted requests.
Input classCasesPassCosmeticStructuralSemanticPass rate
Short free-form18017631097.8%
Long context12010468286.7%
Structured extraction15014901099.3%
Tool calling9071019078.9%
Multilingual605721095.0%
An extract of the 43 disagreements in this example, one per category. The full list is in the report.
#Input classKindWhat changedSeverity
1Tool callingStructuralArguments arrive as a JSON stringHigh — parser patch
2Tool callingStructuralMissing finish_reason on one callMedium
3Long contextSemanticAnswer omits the second conditionHigh — stop cutover
4Long contextSemanticDifferent number in the summaryHigh — stop cutover
5Short free-formCosmeticDifferent opening phraseNone

Recommendation in this example: migrate after patching the tool-call parser, and do not migrate the long-context class until the two semantic differences are understood. The tool-call failures are structural — arguments arrive as a JSON string rather than an object — which is a parser change, not a model regression.

Latency and tokens: median tokens fell 6% on the new provider, while p95 latency rose from 2.1 s to 2.8 s on long-context requests. Neither number is a benchmark; both were measured during this replay.

3. Compliance audit — gap register

The register is the work plan. Each row names the system, the obligation, whether a control exists, whether it can be evidenced, and the target date that follows from the severity.

Illustrative extract from a gap register. Framework: EU AI Act.
SystemObligationControlEvidenceSeverityEffort (days)Target
Support assistantArticle 50 — disclose AI interactionNone in the widgetNone (checked 22 Nov 2026)Critical0.530 days
Support assistantArticle 50 — machine-readable marking of generated outputPartialDesign note only (checked 22 Nov 2026)High360 days
Document extractorAI literacy for staff operating the systemOnboarding deckNo attendance record (checked 22 Nov 2026)High160 days
Document extractorVendor due diligenceContract signedClause identified (checked 22 Nov 2026)Medium0.590 days

What makes it useful: the critical row can be closed by shipping a one-line disclosure, so it is scheduled first. The medium row is already largely evidenced, so it waits. The register is dated, and a register older than a quarter is treated as history rather than a current position.

How to read the baseline

Start at the bottom, not the top. The plan fit table decides the answer; the inputs only exist to fill it. In the example, Hobby hits its 2,000,000-token cap at about 714 requests a day — before its 1,000-request cap. The token cap binds first, and a reviewer can check that division in under a minute. That is the point of keeping it to five lines.

The "type" column matters as much as the number. A measured value comes from logs or a sample; an estimate is labelled and can be challenged. Mixing the two silently is how a planning document becomes a false promise. If a customer cannot supply logs, the baseline still runs, but every row says estimate and the summary says so at the top.

The watch list is the part that survives the meeting. A plan that fits today can fail after a launch, a new customer or a prompt change. Naming the three limits most likely to be hit — and what to do when they are — turns a static document into something the team can act on next month.

What the diff deliberately omits

It does not score "quality" on a five-point scale. A number like 4.2 out of 5 invites comparison between providers that were never measured on the same cases. The report uses pass, cosmetic, structural and semantic because those four categories lead to different actions: ignore, keep, patch the parser, or stop the cutover.

It also does not hide the sample size. Six hundred requests across five input classes is enough to find structural differences and not enough to prove a long-tail guarantee. The report says what was replayed so a reader can judge the coverage rather than trust an adjective.

Why the register is dated

An AI system register is a photograph, not a live feed. Systems are added, models are swapped, suppliers change and staff move. A register without a date cannot be defended because nobody knows which version was true. The report carries the date on the cover and in the evidence appendix, and a register older than a quarter is treated as history.

The severity column is what makes it actionable. "Compliant" and "non-compliant" are not useful labels when a critical row can be closed with a one-line disclosure and a medium row needs a supplier negotiation. Ordering by exposure, with a target date, lets a small team start on Monday instead of commissioning another assessment.

What all three have in common

Questions

Are these real customer documents?

No. They are illustrative samples built with the same templates we use on engagements. Names, volumes and findings are invented so the structure can be inspected without exposing a customer.

Will my report look exactly like this?

The structure and fields are the same. The number of rows, the severity mix and the recommendation depend on your systems and evidence.

Can I reuse the templates?

The samples show the format. The editable templates are delivered during the engagement; they are not a substitute for the assessment itself.

Sources

Want the real thing?

Each service starts with an intake form. Scope and price are confirmed in writing before anything is booked.

See the services Ask a question