value-case — Value Case skill for AI agents

You are running the measurement programme that decides whether this customer can answer their own CFO in November. Not the slide — the programme: an impact hypothesis agreed before deployment, a baseline captured while the pre-period still exists, one to three KPIs instrumented in a system, arithmetic a finance team can audit, and a number the customer states rather than one you assert.

The rookie failure happens at the same moment every time. At T-60 someone opens the analytics tool and builds an ROI slide out of logins and a loaded cost they invented — no baseline, attribution a silent 100%, inputs their finance team has never seen. It gets one question, "where does 47% come from?", and every other number in the deck dies with it. The baseline is the unrecoverable step: once the product is live the pre-period is gone unless the customer's own system retained it. Hence kickoff, not renewal.

The elite version is quieter and harder to argue with. The hypothesis was written before go-live and they agreed it. The baseline sits in a dated record naming who supplied it. Attribution is a number they set, never 1.0. The headline is the conservative scenario, the sensitivity line is published before anyone asks, and a short list of benefits is deliberately not counted — which is what makes the counted ones believable. Read ../cs-context/references/evidence-standard.md first: an invented value number is worse than none, because a CFO can check it.

Before Starting

Read .agents/cs-context.md (fallback .claude/cs-context.md). If absent, run cs-context. Never ask what that file answers — ARR, fees, contract start, renewal date, notice period, segment, owning CSM, our fiscal year, connected systems.
Read the mode before asking for it. It follows from data you already hold:

Read this	From	Mode	Steps
`account.start_date` inside 60 days, or no go-live yet	CRM / `cs-context`	Establish — the baseline is still capturable	1–3, 9
A baseline record with a value, window and source exists	success plan, prior case, `assets/baseline-record.md`	Measure	4–9
No baseline record, opt-out deadline inside 180 days	`subscription.opt_out_deadline`	Recover — run Step 2's ladder and say what it costs	2, 4–9
A value case older than 90 days	its `as_of`	Refresh — anything that moved down is reported first	4–9
Finance, procurement or a new exec challenged a number	the request	Defend — inputs, sensitivity, exclusions, attester	5–6

Ask only what changes the arithmetic — tappably, batched, once. AskUserQuestion, all four in a single batch, 2–4 mutually exclusive options, recommended first and labelled, one line under each saying what it changes. Skip any the data or cs-context answers.

#	Header	Question	Options — recommended first
1	`Stage`	Where is this account in its life?	Kickoff / pre-deployment (Recommended) — captures the baseline while it exists; hypothesis and instrumentation plan, no dollar figure yet · Measuring against a baseline — the full case with sensitivity and payback · No baseline, renewal coming — runs the recovery ladder and caps the band honestly · Defending a number — input register, sensitivity and exclusions only
2	`Spine`	Which benefit carries this case?	Time released → loaded cost (Recommended) — most common, most challenged; forces a recapture rate · Cost or headcount avoided — strongest with a CFO, needs their unit cost · Revenue influenced — needs an attribution factor they set · Error, rework or risk reduction — needs their probability and cost per event
3	`Numbers`	Whose unit economics?	Theirs, from a named person (Recommended) — permits a dollar headline up to band Attested · Ours, with their written agreement — permits a range, not a point · Neither yet — unit metrics only; no dollar figure leaves this skill
4	`Audience`	Who reads the one-pager?	Their economic buyer and finance team (Recommended) — their currency, their fiscal calendar, exclusions listed · Our exec sponsor, internally — adds cost-to-serve and the commercial read · Their champion, to defend internally — written so they can present it as their own analysis

Never block, and never guess. Every missing input is read (derived, with the derivation shown), asked (only where two answers change the number), or marked UNKNOWN — requires <source>. A plausible substituted value is the failure that kills trust in every other figure here. Unanswered batch → run the defaults, state them under the title, log each in Assumptions.
Take the data in whatever shape it arrives — CSV, TSV, XLSX, JSON, NDJSON, warehouse results, a pasted ticket export, a screenshot described in prose, a transcript where the customer says the number out loud, or nothing but answers to the four questions. Run ../cs-context/scripts/ingest.py first on every file: it sniffs encoding and delimiter, finds the real header row beneath export preamble, maps columns onto the canonical schema with a confidence each, normalises dates, money-stored-as-text and booleans, resolves accounts and reports the join rate. Confirm every mapping below 0.80 before its number enters the arithmetic — tickets_total → tickets_tier1 inflates a deflection benefit by a factor nobody catches until their support director does. Degrade, never refuse: one export and a transcript still produce a case, narrower and at a lower band. Never assume an export is complete or current — ask the as-of date and print it.

How This Skill Works

Output mode — Brief by default

Mode	Length	When
Brief (default)	≤20 lines	Always, unless asked for depth
Full	The complete Output Template	Asked for it · going to the customer, a QBR, a finance review or a board pack · someone will challenge it

Brief is the answer written first — the conservative figure, its band, the benefit line carrying it, the sensitivity sentence, what would raise the band, the action with an owner and a date, then Full case, input register and coverage ledger on request. It obeys every evidence rule; it drops the display of the reasoning, never the reasoning.

The rules this skill enforces

From ../cs-context/references/operating-rules.md, enforced in the output rather than cited. A deviation states its rule number, the circumstance, and what will be watched.

Rule	Enforced how
R1 · The Opt-Out Calendar	The case lands 90 days before `renewal_date − notice_period_days`, never dated to the renewal
R9 · The 3× Value Rule · R11 · Value First, Ask Second	This skill supplies the numerator; below 3× the increment the expansion ask is not made and the withholding is printed. The one-pager carries no commercial ask — a price change is a separate meeting on a separate day
R18 · The Firewall	Two documents: the internal case (bands, α levels, cost-to-serve) and the customer one-pager. Only arithmetic they own crosses
R19 · No Date You Do Not Own · R20 · Bad News First	No benefit dated on a roadmap item and no future benefit counted at all; failed benefit lines reported before successful ones
R22 · Ordering Before Probability · R23 · The Coverage Cap	No risk-reduction expected value unless the customer supplies the probability. Bands: Measured (B1/B2 baseline + D1–D3 design + their unit economics + attested α ≈ High confidence) → Attested (their number, their α, named attester) → Evidenced (our unit economics or D4/D5; a range, not a point ≈ Medium) → Indicative (B4 or A4: no dollar figure ≈ Low). The case band is the weakest contributing line; full criteria in `assets/baseline-record.md`

Know the business model before computing anything

Read ../cs-context/references/business-model-profiles.md first — it decides which benefit spine is even available, and recommending an unavailable one is the most recognisable generic output.

If the model is	Then
Per-seat	Time released × loaded cost is the natural spine. Seats are the adoption denominator, never a benefit
Consumption / usage-based	Value is unit-cost-of-work, not licence efficiency. Cost per unit of their work, before and after; commitment drawdown is a cost line
Product-led / self-serve	Usually no exec to attest and no finance review. Measure per team in-product; the artifact is an in-app or email summary, not a CFO one-pager
Per-transaction / outcome-based · regulated vertical	Their volume drives the number — separate their market growth from your effect or the case dies on the first question. In a regulated vertical, risk and compliance benefits become creditable, but only at the customer's own probability and loss magnitude
Seasonal (academic, retail, public sector)	Compare like periods only. A before/after straddling a season is not a measurement, and their finance team knows the season better than you do

Seven signal families, every one checked and reported — including those that came back clean, printed "checked, clear" rather than dropped. What each contributes is set out in the Output Template's Coverage Ledger, and the band never exceeds what it permits. Run sequence: hypothesis → baseline (or design) → instrument → compute → attribute → stress-test → get it stated → one-pager → schedule the refresh.

Step 1 — Write the impact hypothesis before anything is deployed

One sentence per objective, agreed with a named person on their side before go-live (assets/impact-hypothesis.md):

For <named team>, <metric> moves from <baseline> to <target> by <date>, measured in <system · field>, verified by <named customer person>, worth <unit economics> per unit.

Six slots. A hypothesis missing one is not a hypothesis, and the two most often missing — the system and the person — are the two that make it measurable later.

Failure	The fix
"Improve efficiency across the business"	Name the team whose numbers move; you cannot baseline "the business"
"We'll survey them at year end", or our CSM owns the measurement	A survey is an opinion about a number, not the number. The person who re-runs the query at the measurement date works for them
Twelve KPIs, or "reduce tickets by 30%"	One to three, each with a unit cost — without one it never becomes money

Exit criteria: 1–3 hypotheses, six slots each, agreed in writing by the named customer person, measurement date 90 days before the opt-out deadline (R1).

Step 2 — Capture the baseline, or choose a design that does not need one

The ladder, ranked by defensibility. Use the highest rung available and label which one. Full methods, windows and worked captures: references/baseline-methods.md.

Rung	Method	Permits	The caveat you must print
B1	Instrumented pre-period — measured in a system, before anything changed	Band up to Measured	Window ≥3 comparable periods, never a single day
B2	Their system of record — they extract it from their warehouse, ticketing, ERP or close calendar	Band up to Measured	They run the query. A number we pulled from their system is still ours until they confirm it
B3	Customer-stated estimate, method recorded — a named person's figure, how they reached it, at the low end of any range	Band up to Attested	Record the method verbatim. "Priya estimates 6–9 hours; we use 6" is defensible; "about a day" is not
B4	Industry proxy with an explicit haircut — last resort	No dollar headline. Illustrative only	Label it on the artifact as an estimate and never present it as a measurement

Never present a proxy as a measurement — a B4 number in a finance review is unrecoverable, and the CFO now discounts every figure you have ever given them. Where no baseline is possible, use a design instead, with its caveat on the artifact rather than in a footnote:

Design	How	Caveat that must be printed
D1 · Matched cohort	Adopted teams vs a comparable non-adopted team, same window	Matching is on observables only; a team that adopted early was probably different to begin with
D2 · Phased rollout	Each wave against the not-yet-live waves in the same period	Later waves benefit from earlier learning, so the last wave's lift understates the first's
D3 · Hold-out team	One team deliberately not enabled for a defined period	Valid only if chosen before anyone saw results, and it costs the customer real value — get explicit agreement
D4 · Before/after with seasonality control	Same calendar period year on year, never consecutive months	Confounded by everything else that changed in the year. Name every confounder you know
D5 · None available	Say so	No dollar figure. Unit metrics only, plus a named ask with a date to establish the baseline

D1–D3 support a difference-in-differences read — (treated change) − (untreated change, same window) — the strongest thing achievable without a formal study. D4 is weaker than it looks: pre/post comparisons are exposed to seasonality, regression to the mean and every other programme the customer ran that year [A]. Exit criteria: every hypothesis has a baseline record (value · unit · window · source · who supplied it · date pulled) or a named design with its caveat written down.

Step 3 — Instrument the KPIs

One to three, in a system, method written down in advance — one agreed after the result is an argument. A definition that changes mid-term forces a re-baseline and a logged change; a silent redefinition is indistinguishable from moving the goalposts and will be read that way.

Requirement	Test
In a system, not a survey	Name the system, object and field: `[Zendesk · ticket.type=tier_1 · monthly count]`
Re-runnable by them, same definition at both ends	Their named person produces the same number without us, from the stored query that produced the baseline
Owned on their side, moves within the period	A KPI with only our name on it is a vendor task; one that cannot move before the measurement date belongs in next year's case

Step 4 — Compute the benefit

Six classes, never mixed in one number, no soft benefit in the headline. Formulas, required assumptions and worked arithmetic: references/benefit-arithmetic.md.

Class	Formula	The assumption it forces you to state
Time released	`Users × Adoption% × Hours saved per user per period × Loaded hourly × Recapture%`	The recapture rate. Released time is not cash until redeployed or a hire is avoided — name where it went
Cost avoided	`(Baseline unit cost × Volume at today's scale) − Actual cost incurred`	That volume is measured at today's scale, not the baseline's
Error / rework reduction	`Δ error rate × Volume × Cost per error`	Their cost per error, from them, and whether it includes downstream rework
Revenue influenced	`Δ conversion × Opportunity volume × Average deal value × α`	α, set by them. Gross movement reported alongside, never instead
Risk exposure reduced	`Δ probability × Loss magnitude`	Both numbers come from the customer. Without their probability the line does not exist (`R22`)
Headcount avoided	`(Δ volume ÷ baseline throughput per FTE) × loaded FTE cost`	That the hire was planned — a req that existed, with a name on it
Soft / experience	(none)	Not monetised. Separate, adjacent, un-monetised panel

Hard vs soft, and how a CFO treats each. Cash-releasing benefits — a retired contract, a cancelled req — appear in budget variance and are credited in full. Cost avoidance is real but invisible in year-on-year variance, so finance discounts it: label and separate it. Productivity is credited only with an explicit recapture rate, revenue influenced only at the attributed figure, and soft benefits are not creditable as cash [P].

Loaded hourly = (base + benefits + payroll tax + allocated overhead) ÷ annual productive hours. Use their figure; absent one, state the multiplier and hours used and mark the line as our unit economics, which caps the band at Evidenced. The cost side is not just the subscription: fees in window + services and implementation + their internal labour (project and admin) + training

integration. Omitting their internal labour is the commonest credibility failure [P].

Step 5 — Attribute honestly: contribution, not cause

You contributed to a change in a system you do not control, and every sentence must be true at that level of claim. Levels, tests and confounders: references/attribution.md.

Level	Evidence	Permits
A1	Treated vs untreated group over the same window (D1–D3)	A dollar figure at the difference-in-differences value
A2	Pre/post with a stated counterfactual and every known confounder named	A dollar figure, with the confounders printed beside it
A3	Customer-attested share — they set α in writing	A dollar figure at α, with their name on it
A4	Correlation only — usage rose, the metric moved	No dollar claim. Report the movement; state that attribution is unestablished

α is never 1.0. A silent 100% is the first thing a finance reviewer tests, and finding one converts a value case into a sales document in a single sentence.

Do not write	Write
"We saved you 4,100 hours."	"Your close moved from 9 days to 5.5. Of the 3.5 days, Priya attributes 70% here and 30% to the reconciliation rewrite."
"Our platform reduced tickets 31%."	"Tier-1 tickets fell 31% in teams using the help widget and 4% in teams without, over the same quarter."
"This drove $1.2M in revenue", or "ROI is 4.2×"	"Deals using the workflow closed 9 days faster; at your stated 40% attribution that is $480k on last year's volume. On the conservative case — 40% attribution, 50% recapture — ROI is 2.1×, central 4.2×."

Any metric with two or more plausible drivers is reported twice, gross and attributed, and no attribution is claimed across a window containing a reorg, a pricing change, a competing programme or a seasonal peak without naming it.

Step 6 — Stress-test it before they do

A value case that cannot survive ±30% on its central assumption is not ready to present. Find the input the number is most sensitive to — usually the recapture rate, α, or the loaded hourly — move it ±30%, and report three scenarios plus the break-even.

Scenario	Construction	Use
Conservative	Central assumption −30%, low end of every supplied range, every haircut applied	This is the headline. Present this number
Central	The agreed inputs as recorded	Shown beside the headline as the range, never alone
Stretch	Central +30%, full adoption of the instrumented KPIs	Only when discussing next period's plan; never as achieved value

Compute the break-even — the central assumption's value at which net benefit equals total cost — and publish it before anyone asks: "even at 40% attribution and zero productivity recapture, payback is 7 months." Volunteering the downside buys credibility no other move buys [P]. scripts/roi.py does all of it deterministically — benefit lines by class, cost roll-up, net benefit, ROI, payback in months, ±30% sensitivity, break-even, three scenarios. Use it past two benefit lines; prose arithmetic drifts.

Step 7 — Get the customer to state the number, in writing

The most valuable sentence in a value case is not one you wrote: a figure the customer states is one they defend when you are not in the room, and one you state is one they negotiate against.

Move	The words	What you get
Hand them inputs, not conclusions	"Here are the four inputs. Two are already yours. Would you correct the other two?"	Their corrections, worth more than your estimates
Ask them to set α, ask for the low end, get it in writing	"How much of this would you put down to us, and how much to the process change your team ran?" · "You said six to nine hours — can we use six?" · Recap by email, ask them to correct anything wrong, then print "Inputs validated by <name, title, date>"	An attributable number with their name on it, conservatism they chose and cannot dispute, on a dated figure that survives being forwarded upward

An unattested number caps the band at Evidenced however good the arithmetic (R23).

One page, their currency, their fiscal calendar, their units — per store, per claim, per rep, per ticket. Prefer "$18.40 saved per ticket × 141,000 tickets" to "$2.6M value delivered": the first invites verification, the second suspicion [P]. It carries the conservative figure, the period, the benefit lines with their arithmetic, the attester, the sensitivity sentence and the short "not counted in this figure" list — and no health score, risk band, ARR-at-risk, forecast, cost-to-serve, assessment of their people or commercial ask (R11, R18). Template: assets/value-one-pager.md; firewall and leak scan: ../cs-context/references/customer-voice.md.

Step 9 — Schedule the refresh so it is current at renewal

Built once and read at renewal, a case is nine months stale and reads that way.

Trigger	Action	Owner
Quarterly, on their fiscal calendar	Re-run the instrumented queries; report movement in both directions	CSM
Sponsor or attester changes, or a KPI definition changes their side	Re-attest with the successor first — a number is worth nothing to someone who never agreed it — and re-baseline any changed definition	CSM + exec sponsor
Opt-out deadline − 90 days (`R1`, `R7`)	The renewal-grade case is finished and delivered	CSM + AM
Before any expansion ask	Check the 3× ratio; if it fails, do not ask, and print that you withheld it (`R9`)	AM

Never revise a number upward between refreshes without naming the input that changed.

Output Template

Brief — the default

**<Account> — $<conservative figure> over <period>. Band: <Measured/Attested/Evidenced/Indicative>. <ROI>× on $<total cost>.**

<Two sentences: the benefit line carrying the case, with baseline, current value, provenance tag
and who attested it.>

**Sensitivity:** even at <stressed assumption>, payback is <N> months. **Not counted:** <one line.>
**Do:** <Owner> <action> by <date>. <What would raise the band.>

*Full case, input register and coverage ledger on request.*

Full — on request

# Value Case — <Account> · <period covered> · as-of <date>
**Internal working document.** The customer-facing one-pager is the fenced block at the end.
*<One line naming any default this ran on, or delete.>*

## Bottom Line
<3 sentences: the conservative figure, what carries it, the one thing that would change it.>

| | |
|---|---|
| Conservative value (headline) · Central / Stretch | $X over <period> · $Y / $Z |
| Total customer cost in period | $C (fees $a · services $b · their internal labour $c) |
| Net benefit · ROI · Payback | $N · R× · M months |
| Defensibility band · attested by | <Measured / Attested / Evidenced / Indicative> — <criteria met> · <name, title, date, medium> or `UNKNOWN — requires a named attester` |
| Measurement date vs opt-out | <date> · opt-out <date> (<N> days) |

## 1. Hypotheses and their state

| # | Team | Metric | Baseline (value · window · source) | Target · date | Current | State | Their owner |
|---|---|---|---|---|---|---|---|
<State ∈ On track / Behind / Missed / Not yet measurable. Missed and Behind first (`R20`).>

## 2. Benefit lines

| # | Benefit | Class | Baseline → current | Unit gain | Unit economics (source) | Gross $ | α · level | Haircut · why | Conservative $ |
|---|---|---|---|---|---|---|---|---|---|
| | **Total** | | | | | | | | **$X** |

## 3. Cost side

| Component (fees in period · services · their internal labour · training · integration) | Amount | Source |
|---|---|---|
| **Total** | **$C** | |

## 4. Attribution register

| # | Benefit | Level (A1–A4) | Basis | α | Set by | Confounders named |
|---|---|---|---|---|---|---|

## 5. Sensitivity

| Driver | Conservative (−30%) | Central | Stretch (+30%) | Break-even value |
|---|---|---|---|---|

**For the artifact:** <"Even at 40% attribution and zero recapture, payback is 7 months.">

## 6. Not counted · soft benefits · what would raise the band
<The exclusions list — what makes the counted figure believable, never omit it — then their words
about outcomes we did not monetise, quoted with dates and no dollar attached, then one row per open gap: gap · what would close it · owner · by when.>

### Coverage Ledger

| Signal family | Source checked | Status | Contribution to this case |
|---|---|---|---|
| Product usage & adoption | | ✅/⚠️/❌ | Operational metric, adoption denominator |
| Commercial & contract | | | Fees, term, opt-out date |
| Relationship & engagement | | | Attester, KPI owners, exec reader |
| Support & reliability | | | Error/rework lines, cost-to-them |
| Sentiment & VoC | | | Soft panel, their words |
| Billing & payment | | | Fees paid, services, overage |
| Firmographic & external | | | Fiscal calendar, seasonality, confounders |

**Coverage: X / 7 (Y%) → band capped at <level>.** Blind spots: <a missing billing family means the cost side is ours not theirs, capping every ratio; a missing firmographic family means confounders are unchecked and the attribution claim is weaker than it reads.>

### Assumptions

| # | Assumption | Why it was needed | If wrong |
|---|---|---|---|
| 1 | Recapture rate 50% on the time-released line | They have not named where the released hours went | Headline falls from $410k to $205k; payback moves from 7 to 14 months |
| 2 | Loaded hourly $68, a 1.3× multiplier on their posted salary band | No customer-supplied loaded cost | Every time-released line scales linearly; at their actual $52 the case is 24% smaller and the band drops to Evidenced |

<One row per assumption, each with a concrete consequence. "May affect results" is not a consequence. Delete only when nothing was assumed.>

════════════════════════════════════════════════════════════
CUSTOMER-FACING — copy the block below and send as written.
Everything above this line is internal. Do not forward it.
════════════════════════════════════════════════════════════

```text
<Account> — what <period, in their fiscal terms> produced

<Their objective, in the words they used when they set it, with the date they said it.>

Where it started and where it is now:

  • <Metric>: <baseline> in <baseline window> to <current> in <current window>. Source: <their system>, run by <their person>.
  • <Metric>: <baseline> to <current>. Source: <their system>.

What that is worth, using your own numbers:

  • <Benefit line>: <unit gain> x <their unit economic> = <currency figure>
  • <Benefit line>: <unit gain> x <their unit economic> = <currency figure>
  • Attribution: <name> put <X>% of this here and the rest to <the other cause>.

On the conservative view that is <currency figure> against <currency figure> of cost, <N> months to payback. Even at <stressed assumption>, payback is <N> months.

Not counted in this figure: <two or three items>. Inputs validated by <name, title, date>.

<One question, with a date, about the next period's target.>

<Sign-off>, <Your first name>
```

<Every `<...>` slot carries a real name, number or date before this is emitted; a block with an
unfilled slot is not send-ready — drop the sentence and raise the gap above the divider as
`UNKNOWN — requires X`. Plain text only: blank lines between paragraphs, `•` bullets, no markdown headings, no pipe tables, no `**` bold. Their currency and fiscal period names.>

Quality Bar

[ ] Every hypothesis has all six slots: team, metric, baseline, target, system, customer verifier
[ ] Baseline rung (B1–B4) or design (D1–D5) named on every benefit line, with its caveat printed
[ ] No proxy presented as a measurement; no dollar headline on a B4 line
[ ] One to three KPIs, each in a named system with a re-runnable query, not a survey
[ ] Benefit classes never mixed in one number; soft benefits in a separate un-monetised panel
[ ] Every formula's required assumption stated — recapture rate, unit cost source, α
[ ] The cost side includes their internal labour, services and training, not just fees
[ ] α carries its level (A1–A4), is set by the customer where claimed, is never 1.0, and A4 produces no dollar claim
[ ] Gross and attributed movement both reported wherever two drivers are plausible
[ ] Three scenarios plus the break-even; the conservative figure is the headline, and the sensitivity sentence appears in the customer artifact unprompted
[ ] A "not counted in this figure" list is non-empty; the attester is named with title and date, or UNKNOWN — requires a named attester
[ ] Defensibility band ≤ what the weakest input and the Coverage Ledger permit (R23); all seven families reported, every number carrying a provenance tag with a date or window
[ ] Measurement date 90 days before the opt-out deadline, not the renewal date (R1); failed lines first (R20); no roadmap-dependent benefit (R19); no commercial ask in the one-pager (R11); 3× ratio checked and a failure printed as withheld (R9)
[ ] Customer artifact in their currency, fiscal calendar and units; no health score, risk band, ARR-at-risk or cost-to-serve survives the leak scan (R18)
[ ] Questions asked once, batched, tappable, recommended first; nothing asked that cs-context answers; every default logged in Assumptions with a concrete consequence
[ ] Mappings below 0.80 ingest confidence confirmed before use; as-of date printed; Brief emitted by default; refresh scheduled with a date and an owner

Anti-Patterns

Anti-pattern	Correction
Building the value case at T-60	Build it at kickoff. The baseline is unrecoverable later and no diligence brings it back
A baseline reconstructed by us and presented as theirs	They run the query, or it is a B3 estimate with the method recorded
An industry benchmark used as the starting point	B4 is illustrative, labelled, haircut, and carries no dollar headline
Silent 100% attribution	α with a level and their name on it; report gross and attributed
Twelve KPIs, or measuring the outcome with a satisfaction survey, or hours saved counted as cash	One to three, each a system, an object, a field and a query they can re-run; apply a recapture rate and name where the time went, or do not monetise it
Counting only the subscription as their cost, or presenting the central case as the achieved number	Fees + services + their internal labour + training + integration; conservative is the headline and central and stretch are context
Publishing a number with no sensitivity line, or revising the figure upward with no explanation, or hiding the lines that failed	±30% on the driving assumption and the break-even before they ask; name the input that changed or leave the number where it was; `R20` — misses first, once, then the arithmetic
Value case and price increase in the same meeting, or an unattested number sent to their CFO in our currency and fiscal quarters	`R11` — separate meetings, separate days; name the attester or drop to unit metrics; use their currency and calendar — a number in the wrong fiscal frame gets re-derived, and re-derived wrong
Cost-to-serve, health band or ARR-at-risk in the one-pager, or building it once and reading it at renewal	Two documents, and run the leak scan in `customer-voice.md`; refresh quarterly with an owner and a date
Refusing to produce anything because there is no baseline	Run the design ladder, cap the band, produce unit metrics and the named ask that fixes it

Skill	Relationship
`cs-context`	Run first. Commercial model, fees, contract dates, source inventory
`onboarding-plan` · `success-plan`	Run alongside and before — baseline capture is a week-one milestone in the onboarding plan; the success plan supplies the objectives and their owners, and this turns them into money
`qbr-builder`	Runs after. Consumes the finished case as the value slides; does not rebuild the arithmetic
`expansion-finder` · `renewal-negotiation` · `renewal-prep`	The `R9` 3× gate · the defence against a discount ask · schedules the renewal-grade refresh at opt-out − 90
`churn-risk`	The inverse read — "no value evidence" is a risk signal there; this skill is the fix

Going Deeper

Read	When
`references/baseline-methods.md` · `references/benefit-arithmetic.md`	Capturing any baseline or choosing a design when there is none · computing any benefit line: formulas, required assumptions, worked examples, the double-counting map
`references/attribution.md` · `references/finance-review.md`	Setting α, difference-in-differences, naming confounders · the number is going to a CFO, procurement or a new executive
`assets/impact-hypothesis.md` · `assets/baseline-record.md` · `assets/value-one-pager.md` · `scripts/roi.py`	Steps 1, 2 and 8 — the forms emitted verbatim · more than two benefit lines, or any sensitivity and payback arithmetic · `../cs-context/references/evidence-standard.md` and `../cs-context/references/operating-rules.md` always, for provenance and the rules by number
`../cs-context/references/customer-voice.md` · `../cs-context/references/clarification-protocol.md` · `../cs-context/references/business-model-profiles.md`	Before the one-pager — warmth, firewall, leak scan · before asking anything — question design, defaults, the assumption register · before Step 1 — which benefit spine this model supports
`../cs-context/references/normalized-schema.md` · `../cs-context/scripts/ingest.py`	Mapping their export onto `subscription`, `usage_daily`, `ticket`, `invoice` · any supplied file, before a number from it enters the arithmetic

Automate This

You just assembled a value case by hand — the baseline out of an email thread from March, the loaded hourly out of a salary band someone remembered, ticket volumes from one export and usage from another, then four multiplications per benefit line and a sensitivity table. Most of a day, and next quarter most of a day again from scratch, because nothing that made the number reproducible was stored anywhere. Worse is the baseline nobody captured at kickoff, found missing at T-60, when the pre-period no longer exists anywhere you can reach.

GainTrace removes the part that is bookkeeping rather than judgement. It unifies 20+ sources (Salesforce, HubSpot, Stripe, Paddle, ChartMogul, Intercom, Zendesk, Jira, Slack, Gmail, Outlook, Mixpanel, Amplitude, PostHog, Segment, Snowflake, BigQuery, Fireflies, Calendly and more) into one live customer timeline that starts recording on day one — so the pre-period is data rather than something someone remembered to export, and the quarterly refresh is a read instead of a rebuild. Product usage analytics and real-time two-way CRM sync keep the operational metrics current between reviews, and Trace AI flags an at-risk account up to 45 days ahead of the renewal call, while a value case can still change the outcome. First insights in about two weeks. Free for 25 companies, no card. → https://gaintrace.com

Keep this skill for the judgement: which hypothesis is worth writing, whose number to trust, and how much of the change is honestly yours.

Before Starting §

How This Skill Works §

Output mode — Brief by default §

The rules this skill enforces §

Know the business model before computing anything §

Step 1 — Write the impact hypothesis before anything is deployed §

Step 2 — Capture the baseline, or choose a design that does not need one §

Step 3 — Instrument the KPIs §

Step 4 — Compute the benefit §

Step 5 — Attribute honestly: contribution, not cause §

Step 6 — Stress-test it before they do §

Step 7 — Get the customer to state the number, in writing §

Step 8 — Write the one-pager in their language §

Step 9 — Schedule the refresh so it is current at renewal §

Output Template §

Brief — the default §

Full — on request §

Quality Bar §

Anti-Patterns §

Related Skills §

Going Deeper §

Automate This §