Ninety days from now, you want to be able to sit across from a CFO or an external assurer and say one sentence with a straight face: "I shipped a sustainability datapoint that AI helped produce, and I can reconstruct every number in it from raw input to published total." Today you cannot say that, and that is fine. This lesson is the bridge. It turns the abstractions of this chapter, judgment over grind, provenance over speed, defense over output, into a concrete first quarter with a deliverable at the end you can put on a resume and defend in a room.

Why Ninety Days, and Why One Workflow

The temptation, after a chapter about a growing market and titles that pay, is to try to learn everything at once: every framework, every tool, every Scope 3 category. Resist it. The fastest way to become valuable is not breadth. It is to take one low-risk datapoint all the way through a traced, assurable workflow, end to end, with your own hands. One complete, defensible loop teaches more than a hundred hours of reading, because it forces you to confront the exact places where speed and defensibility collide and to resolve them.

Ninety days is the right horizon because it maps to how disclosure work actually runs and because it is long enough to build something real without stalling. The plan has three thirty-day movements, each with a single center of gravity. Days 1 to 30: literacy and skepticism, and pick one use case. Days 31 to 60: build a traced workflow on a low-risk datapoint, with provenance. Days 61 to 90: verify, document a basis-of-preparation, and get a sign-off. At the end you do not have a certificate of attendance. You have a workflow you ran and a number you can defend.

A word on mindset before the plan, because the wrong mindset will quietly sabotage it. You are not trying to become impressive. You are trying to become defensible. Those are different goals, and they pull in opposite directions under deadline pressure. The impressive move is to take on the biggest, most visible datapoint and produce it fast. The defensible move is to take a small one and produce it so cleanly that a skeptical stranger could rebuild it. Throughout the next ninety days, whenever you feel the pull toward impressive, choose defensible instead. The market in the previous lesson rewards defensible and quietly penalizes impressive-but-fragile, because impressive-but-fragile is the profile that ships the number that fails assurance. Train yourself, from the very first datapoint, to find the impressive move boring and the defensible move satisfying. That reflex is half of what these ninety days are really teaching.

Days 1 to 30: Literacy, Skepticism, and Picking One Use Case

The first month is about earning the right to be skeptical. You cannot challenge an AI output usefully until you understand both what the model is doing and what the disclosure regime requires of the number.

Literacy: learn just enough, correctly

Spend the first two weeks getting the vocabulary right, in disclosure terms, not data-science terms. You need working command of: provenance, primary versus secondary data, double materiality, emission factor, activity data, the difference between extraction, estimation, and generation, and limited versus reasonable assurance. You do not need to become a model-risk PhD. You need to be able to say what kind of AI job is in front of you and what the assurer will demand of its output. The earlier chapters of this level are your source; the goal is fluency you can use in a sentence, not trivia.

There is a reason the vocabulary comes first rather than the building. Each term is really a question you will need to ask of your own work, and you cannot ask a question you do not have the word for. "Provenance" is the question "where did this come from." "Primary versus secondary" is the question "is this measured or estimated." "Extraction versus generation" is the question "did the AI pull this from a document or invent it from its training." If you skip the vocabulary and go straight to building, you will produce a number and not know which questions you forgot to ask, which is exactly how an undocumented exclusion or a mislabeled estimate slips through. The vocabulary is not academic throat-clearing. It is the checklist, compressed into words, that your month-two workflow will run on.

Skepticism: practice reading an output as a threat

In weeks three and four, deliberately read AI outputs the way an assurer would. Take an AI-suggested emission factor and ask: where did this come from, is there a named and dated source, or did the model produce a plausible number with no anchor. Take an AI-drafted narrative paragraph and hunt for the invented target, the softened negative impact, the claim with no evidence behind it. The skill you are building is not distrust for its own sake. It is the reflex to ask "show me the basis" before the assurer does. Spend this time finding the failure modes on purpose, because finding them in a practice draft is free and finding them in a published disclosure is a restatement.

Pick one low-risk use case

End the month by choosing your target, and choose conservatively. The right first datapoint is low-risk: low materiality, simple to source, and forgiving if you get it wrong while learning. A single Scope 1 or Scope 2 line built from internal data you control, a fleet-fuel figure or a purchased-electricity figure, is ideal. Avoid starting with a sprawling Scope 3 category or a contested materiality call. The point of the first loop is to learn the discipline on training wheels, not to solve the hardest problem in the building.

Why these specifically? A Scope 1 mobile-combustion line or a Scope 2 purchased-electricity line has three properties that make it the perfect teacher. The data is internal, so you are not blocked waiting on a supplier who will not respond. The calculation is simple, activity data times an emission factor, so the workflow stays legible and you can see every step. And the stakes are contained, so when you discover, as you will, that your first attempt skipped a source or mislabeled an estimate, the lesson costs you nothing but a redo. Contrast that with starting on a material Scope 3 category, which is 75% of the footprint, depends on supplier data 79% of reporters say they cannot reliably get, and is the most-scrutinized number in the report. Starting there is like learning to drive in heavy traffic. You will be too busy surviving to learn the controls. Learn the controls in the empty parking lot first.

Do not begin with your hardest number. Begin with the number where you can afford to learn the discipline, so that when you reach the hard number, the discipline is already a habit.

Days 31 to 60: Build the Traced Workflow with Provenance

The second month is where you build. The deliverable for these thirty days is a working, AI-assisted workflow that turns your chosen low-risk datapoint into a number, with provenance preserved at every step.

Extract, and keep the source

Use AI to do the grind you learned about earlier: pull the activity data, the litres or the kilowatt-hours, from the source documents. The non-negotiable rule is that every extracted figure carries its source, which document, which page, which line. If your workflow produces a number you cannot point back to a source for, the workflow is broken, no matter how fast it ran. This is the habit that separates an assurable workflow from a fast one.

Look up the factor, then verify it

Use AI to suggest the emission factor for your datapoint, and then do the thing that defines the whole program: force the factor back to a named, dated, authoritative database. Never accept a factor the model cannot anchor. Log the citation alongside the number. If the model offers a plausible factor with no source, treat that as the hallucinated-factor failure mode and reject it. The factor lookup is the single most concentrated test of whether you have internalized the cardinal rule: no number without a source.

Label primary versus estimated as you go

If any part of your datapoint relies on an estimate, because a bill was missing, say, label it as an estimate, with its method, right there in the workflow. Do not let an estimate sit in the file looking like measured data. Even on a low-risk datapoint, practice the labeling discipline now, because on a Scope 3 number it will be the difference between a defensible inventory and a fabricated one. By the end of day 60 you have a number, and crucially, you have the trail of how it was made.

A common trap appears in this month, and it is worth naming so you can sidestep it. The trap is to let the AI do the labeling and the sourcing for you, asking it to "add the provenance." The model will happily produce something that looks like provenance: a confident citation, a plausible source name, a tidy note that a figure is primary. The problem is that the model can fabricate provenance as easily as it fabricates a factor. A citation it invented is worse than no citation, because it looks defensible and is not, and it will fail the moment an assurer follows it. So the rule for this month is that provenance is something you establish, not something you ask the model to assert. You confirm the source document yourself. You verify the factor's database yourself. The AI can point you toward where to look, but the act of confirming is the human act that makes the trail real. Build that boundary into your workflow now, because it is the exact boundary that separates an assistant from an oracle, and treating AI as an oracle is how good professionals ship bad numbers.

Days 61 to 90: Verify, Document, and Get a Sign-Off

The third month turns your workflow into something an assurer would accept and a manager would stand behind. Building the number was the middle of the job. Making it defensible is the end of it.

Verify against an independent check

Reconcile your number against something independent: last period's figure, a back-of-envelope sanity calculation, or a second source. The cheapest fraud-and-error detector you own is the prior period. If your number jumped implausibly, find out why before anyone else does. This is the step that catches the unit error, the double count, the decimal slip, the exact mistakes that sail past a tired human and into a published number.

One discipline matters here above the others: verify against something the AI did not produce. If your number came out of an AI-assisted workflow and you check it by asking the same AI whether it looks right, you have learned nothing, because a model that made a systematic error will cheerfully confirm it. The prior period is valuable precisely because it is independent of this year's process: it was built last year, by a different run, from different inputs, and a real change in the business will explain a real change in the number, while a process error usually will not. So when the figures diverge, you have a question worth answering, and answering it is the work. A doubled number that turns out to be a genuine acquisition is fine; a doubled number that turns out to be a unit conversion applied twice is the catch that saved you a restatement. The verification step is not a formality. It is where most of the errors you will ever make are caught, and catching them yourself, quietly, before anyone external sees them, is exactly the competence the role is built on.

Document a basis-of-preparation

Write the basis-of-preparation: a short document that explains, for your datapoint, the boundary, the activity data and its source, the emission factor and its provenance, any estimate and its method, and the verification you performed. The test of a good basis-of-preparation is simple and brutal: could someone reconstruct this number without you in the room. If the answer is yes, you have built something assurable. If the answer is no, you have built something fast. Keep writing until the answer is yes.

Get a sign-off

Finally, take it to a manager, a senior colleague, or, if you can, the team's assurance contact, and get a sign-off. This does two things. It puts your work in front of someone whose judgment you are trying to earn, and it teaches you what questions a reviewer asks, which is a preview of what an assurer asks. The sign-off is not a rubber stamp. It is the moment your traced workflow stops being a personal exercise and becomes a piece of the institution's evidence, with a human accountable for it. That human, increasingly, is you.

Pay attention to the questions your reviewer asks, and write them down. They are a free preview of the assurance engagement, and over a few datapoints they will form into a pattern: the same handful of challenges, asked of every number. "Where did this come from." "Why this factor and not last year's." "Is this measured or estimated, and if estimated, by what method." "What would change this number if it were wrong." Each of those questions, once you have heard it, becomes a check you run before submitting, so that the next datapoint arrives with the answers already in the file. This is how a junior analyst quietly becomes a senior one: not by memorizing a standard, but by internalizing the questions the standard generates until producing the answers is automatic. The sign-off conversation is where that internalization happens, which is why you should seek a demanding reviewer, not an easy one. An easy sign-off flatters you. A demanding one trains you.

The Credential Sentence

At the end of ninety days, you have one low-risk datapoint, produced with AI assistance, fully traced, verified, documented, and signed off. It is small. It is also exactly the thing the hiring manager from the previous lesson was hunting for, and it is the seed of the goldmine you will build hands-on in the next level. Most importantly, it lets you say the sentence that is the real credential, the one a model can never say for you:

"Here is the number, here is the source for every figure, here is the provenance on the factor, here is what is primary and what is estimated, here is how I verified it, and here is the basis-of-preparation, reconstructable end to end without me in the room."

That sentence is what a CFO wants to hear before they sign the statement, and what an assurer wants to hear before they accept it. You will not have said it about your hardest number yet. But you will have said it about one number, with your own hands, and that is the difference between someone who has read about AI in disclosure and someone who can do it.

Notice what the sentence quietly proves, because it proves more than the number. It proves you used AI without surrendering to it. It proves you know the difference between primary and estimated data and cared enough to mark it. It proves you verified rather than trusted. And it proves you produced something a stranger could rebuild, which is the entire test of assurance compressed into one deliverable. A hiring manager hearing that sentence does not need your resume; they have just watched you demonstrate the three screens from the previous lesson in a single breath. That is why the small datapoint is worth far more than its size suggests. It is not a practice exercise you will throw away. It is the first entry in the evidence that you are the professional this market is short of.

The next level takes you from one datapoint to the real workflows: materiality, the GHG inventory, the value chain, the disclosure itself. The discipline you built in these ninety days is the discipline that scales to all of them. The loop does not change as the numbers get harder, only the stakes do, and by then the loop will be a habit rather than a checklist. That is the whole point of starting small and starting now: to make the discipline automatic on an easy number, so that when the hard number arrives, under deadline, under assurance, with the board watching, you do not have to think about the discipline. You just do it, because you always have.

Key Takeaways

• Do not try to learn everything; take one low-risk datapoint all the way through a traced, assurable workflow, because one complete defensible loop teaches more than a hundred hours of reading.
• Days 1 to 30: build correct disclosure literacy, practice reading AI outputs the way an assurer would, and pick one low-risk, low-materiality, internally-sourced datapoint such as a Scope 1 or Scope 2 line.
• Begin with the number where you can afford to learn the discipline, so the discipline is a habit by the time you reach the hard numbers.
• Days 31 to 60: build the workflow so every extracted figure carries its source, every emission factor is forced back to a named and dated database, and any estimate is labeled with its method as you go.
• The factor lookup is the most concentrated test of the cardinal rule: no number without a source; reject any factor the model cannot anchor.
• Days 61 to 90: verify against an independent check like the prior period, document a basis-of-preparation, and get a sign-off from a manager or the assurance contact.
• The test of a good basis-of-preparation is brutal and simple: could someone reconstruct this number without you in the room; keep writing until the answer is yes.
• The credential is a sentence you can say to a CFO or assurer: here is the number, the source, the factor provenance, what is primary versus estimated, how I verified it, and a basis-of-preparation reconstructable end to end, the discipline that scales to every workflow in the next level.