By Errin O'Connor — Founder & Chief AI Architect, EPC Group · Microsoft Solutions Partner across all six designations · G2 Leader in BI consulting
On July 7, Ethan Mollick — the Wharton professor whose essays get forwarded around more executive teams than any analyst report I have seen in a decade — said the quiet part out loud: prompting tricks are over. His team's Prompting Science research at Wharton had been dismantling the folklore for months — the magic phrases, the tips and threats, the fictional expert personas — and the July verdict landed clean. The skill that matters now is not the clever prompt. It is the specification: define the goal, define the output, define the quality bar, define the test. Manage the AI like a capable, literal executor. Stop treating it like a spellbook.
He is right. And I want to take his conclusion one step further than he did, because there is a version of this argument that lives inside every Microsoft enterprise tenant in North America, and almost nobody has connected the two.
In a Copilot deployment, you do not write the spec in a prompt box. You write the spec in the semantic model. And most enterprises have never written it.
The spec already has a file format. You've been ignoring it.
When Mollick says the future belongs to people who can define goals, outputs, quality bars, and tests, every Power BI architect should feel a shock of recognition — because that is a literal description of a governed semantic model. The Tabular Model Definition Language is the specification language. The explicit measure is the output definition. The description field on every table, column, and relationship is the goal statement. The certified-dataset endorsement is the quality bar. The deployment pipeline with tests is, well, the test.
Copilot does not read your mind. Copilot reads your metadata. When a CFO asks Copilot "what was Q2 gross margin," the model goes looking for a measure whose name, description, and lineage tell it what gross margin means in your enterprise. If that measure is explicit, documented, and certified, Copilot returns your number. If the measure is implicit — an ad-hoc aggregation someone dragged onto a canvas in 2021 — Copilot guesses. It guesses fluently. It guesses confidently. And a fluent, confident guess in a financial workflow is not a productivity gain. It is a deposition exhibit with a delay on it.
This is why prompt engineering was always a temporary discipline in the enterprise. You cannot prompt your way around a semantic layer that does not say what your business means. The prompt is a map. The model and the data are the territory. Tricks decorate the map. Specifications align it with reality.
Twenty-nine years of watching the same movie
I have been in enterprise Microsoft consulting since the Project Tahoe beta became SharePoint in 2001, and I was on the Power BI beta when it was still code-named Project Crescent. Every platform generation has a phase where the market sells incantations — the perfect web part, the perfect workflow, the perfect prompt — and a phase where the adults show up and write the specification. The incantation phase is exciting. The specification phase is where the money gets made and the lawsuits get avoided. Mollick's July research is the formal announcement that generative AI just crossed from one phase to the other.
Here is what the specification phase looks like inside the average Fortune 500 tenant we audit — and I mean the average one, across healthcare, financial services, government, manufacturing, energy, education, retail, and technology. Between forty and two hundred duplicate or orphaned workspaces. Implicit measures scattered across hundreds of reports. Description fields blank on the majority of tables and columns. Certified endorsements used as decoration rather than gate. Published industry research says data teams juggle roughly four hundred sources on average and spend sixty percent of their time cleaning data. That environment is the "spec" your Copilot is reading right now. Would you accept a statement of work written like that? Then why are you accepting a semantic layer written like that — and paying per user, per month, for an AI to read it aloud?
The framework: the Spec Stack
Here is how I tell clients to translate Mollick's management thesis into their Microsoft estate. I call it the Spec Stack, and it has four layers, bottom to top.
Layer one — the vocabulary. Metadata enrichment across every table, column, measure, and relationship, authored in TMDL and version-controlled in PBIP. If a new analyst couldn't understand your model from its descriptions alone, neither can Copilot.
Layer two — the sentences. Explicit measure discipline. Every number an AI is allowed to reason over gets authored deliberately, documented, tested, and certified. Implicit measures are the single most common cause of plausible-but-wrong Copilot answers we find in the field.
Layer three — the grammar. Workspace ownership, lineage, and deduplication — one certified source of truth per subject area, an accountable named owner, and a retirement plan for the duplicates.
Layer four — the permission slip. An approved-for-AI matrix: which semantic models may be grounded by which models, for which audiences, under which retention policies — written down, integrated into Microsoft Purview, and signed by the General Counsel.
Stand those four layers up and something remarkable happens: the prompt stops mattering. Any reasonably phrased question returns the certified number, because the specification — not the phrasing — is doing the work. That is exactly the outcome Wharton's research predicts, delivered in the file formats your tenant already supports.
What I tell clients to do
One. Run the metadata audit this month. Count blank descriptions, implicit measures, and duplicate workspaces. That number is your real Copilot readiness score — not your license count.
Two. Pick the one semantic model your CFO actually uses and take it through all four layers of the Spec Stack. Prove the pattern on the model that matters before scaling the pattern to the models that don't.
Three. Put the approved-for-AI matrix in front of your General Counsel before your next license true-up. The conversation is thirty minutes now or thirty days in discovery later.
Four. Retrain your "prompt engineering" budget as specification budget. The people you need are not prompt whisperers. They are the modelers, stewards, and architects who can write down what your business means — which is the work EPC Group's Power BI practice has been doing since the product was a beta code-named Crescent.
The Spec Stack in the repo
Layer 1: descriptions authored in TMDL, PBIP in Git, PR review required for model changes — the semantic model gets the same SDLC as code.
Layer 2: implicit-measure lint — CI check that fails a build introducing implicit aggregations on certified models.
Layer 3: workspace registry — owner, lineage, certification status, dedupe backlog with retirement dates.
Layer 4: approved-for-AI matrix stored as data (not a slide), enforced via Purview labels + tenant settings; reviewed quarterly with the GC; every Copilot-groundable model carries the endorsement + label pair.
Where I land
Mollick closed the book on prompting tricks. I am closing it on the enterprise excuse that followed them — the idea that AI underperformance is a phrasing problem. It was never phrasing. It was always the spec. And in the Microsoft world, the spec has a name, a file format, a version-control story, and a certification path. Your semantic model is the new spec. Write it like the money depends on it — because it does.
The data behind this (sources and verification)
- Wharton Prompting Science Reports 1-4 (Meincke, Mollick, Shapiro) — Prompt tricks contingent and declining in value; Mollick July 7, 2026 synthesis: goals, outputs, quality bars, tests.
- EPC Group field baseline (Fortune 500 tenant audits) — Directional finding: 40-200 duplicate or orphaned workspaces in the average enterprise Microsoft tenant.
- Industry research (1,000+ employee enterprises) — Data teams juggle ~400 sources on average; ~60% of data-scientist time is cleaning data, ~19% finding it.
- Microsoft WorkLab — 11-by-11 finding — Approximately 11 minutes/day for ~11 weeks as the habit-forming tipping point for Copilot adoption.
Third-party figures above are attributed to their named sources as of the Last verified date. EPC Group audit figures are directional findings from client engagements. Items marked [VERIFY] must be confirmed before external quotation.
Frequently asked questions
Why does Copilot give wrong numbers in Power BI?
Usually implicit measures and undocumented metadata: the model guesses which aggregation "gross margin" means because your semantic model does not say. Explicit, certified, documented measures fix the class of error. If your certified dataset has a documented Gross Margin measure with a clear description and lineage, Copilot returns your number. If it does not, Copilot infers one from the data model — fluently, confidently, wrongly.
What is TMDL and why does it matter for AI?
Tabular Model Definition Language — the human-readable, version-controllable format for Power BI and Analysis Services semantic models. TMDL is where the "specification" Wharton says now matters actually lives in a Microsoft estate. Every measure, description, relationship, and calculation group is authored in TMDL, version-controlled alongside your app code in PBIP + Git, and reviewed in pull requests. It is the file format your semantic layer needed for a decade and finally has.
Is prompt engineering completely useless now?
Format and context still matter; magic phrasing does not. Investment moves from prompt tricks (persona-priming, threats, tips, made-up expert names) to specification quality: metadata, measures, matrices, and tests. If a well-phrased and a poorly-phrased question both return the certified number because the model is well-specified, that is the state you want.
What is an approved-for-AI matrix?
A signed register defining which semantic models may be grounded by which AI models, for which audiences, under which retention policies. Approved and reviewed quarterly with the General Counsel; enforced technically via Microsoft Purview sensitivity labels and tenant-level Copilot settings. The point is not to be restrictive; it is to make grounding decisions auditable and provable to a regulator, litigator, or insurer.
How long does Spec Stack readiness take?
One priority semantic model through all four layers in weeks, not months — start with the model your CFO actually quotes from. Once the pattern is proven on the model that matters, it scales to the models that do not. The alternative — parallel readiness projects across every workspace — is how programs die of committee-review latency.
Ready to act on this?
Start with the practice most relevant to your estate, or reach out directly for a senior-architect conversation.
📧 contact@epcgroup.net · 📞 (888) 381-9725
Multiple models. One truth.