Consistency is the real AI skill gap
As AI tools become part of Excel-based modelling, freeform prompting produces inconsistent output across a team, even on an identical task.
This webinar asks whether prompting is the right skill to teach at all. Instead of another session on writing better prompts, it compares freeform prompting with a structured, skills-based approach to AI assistance on the same real modelling task. It shows what changes when house convention is built into the assistance itself, instead of being left to whoever is typing.
What you will learn
- Why freeform prompting produces inconsistent modelling output across a team, even on an identical task
- How freeform prompting and skills-based assistance compare when run side by side on the same task
- How house conventions can be built into AI assistance instead of relearned prompt by prompt
- A practical framework for deciding when to prompt and when to build a reusable skill
The full session, in writing
“… it has to be bankable, it has to be auditable, and that is really driving the discussion we are having here.”
The webinar opens a planned series on AI in financial modelling. It separates three ways of directing an AI tool. A prompt is a one-off instruction typed into a chat. A prompt library is a set of tested, detailed prompts stored centrally and copied in by hand each time. A skill is a standard that lives inside the tool itself, in Excel, Copilot or Claude, and is used automatically whenever that tool is used, without being retyped or remembered.
The central argument is that a simple prompt leaves many small modelling decisions to the AI: the base date for an escalation, whether it compounds mid-period or at the end of the period, whether a year is split into quarters or days. Modellers make these decisions almost without thinking, from training and convention. An AI model given only a short instruction has to guess, and each small guess compounds over a long project life into a material difference.
“A prompt is just your chat. A skill is a standard that sits within whatever tool you are using, and it is referred to whenever you use that model.”
This was shown live. Both presenters gave the same simple cost escalation prompt to the same AI model on two different machines. Both results were mathematically correct but structured differently: one had a header and the other did not, one used quarters in a year while the other put the day-count assumption in a separate unlabelled cell, and formula references were placed inconsistently across columns. The two outputs looked close but were not identical, a small difference that came from how each split a year into periods. In a capital-intensive project finance model, that kind of variance compounds into a real difference in NPV, not just a formatting inconsistency.
The webinar then shows where these conventions live in different tools. Copilot’s skills sit in a set OneDrive folder structure, one markdown file per skill. A newer feature called Rules needs a specifically named, visible worksheet with the rules in column A. Claude’s skills can sit in the sidebar, be shared across an organisation centrally or sit inside a larger project alongside example files. When the presenters repeated the escalation test with a shared, purpose-built skill instead of a bare prompt, the two outputs came back with matching headers, formulas and cell formatting. Applying the same skill to a different calculation, revenue escalation instead of cost, gave the same consistent structure again.
“Even though you would have a consistent skill, the same prompt will definitely give you different results. That is the risk associated with prompts.”
The session closes with a note of caution. Reviewing calculations built by AI needs the same discipline as reviewing any model: checking individual cells and period-to-period movements, not just annual totals, because an AI model can hide a material error inside a total that looks right. Skills change how a model is built or edited, not how it is audited: a model audit still needs a person to take accountability for signing it off. Data security also shapes what goes into an AI tool in practice, with models usually kept on a local device or SharePoint instead of being uploaded into the AI environment. On VBA, current frontier models can write macro code but cannot yet run it inside Excel, and cannot yet judge whether the code is well optimised. That still rests on the modeller’s experience.
“The short answer is that model audit has not changed. The skills are much more useful for the building or the editing of a model, but the audit is not changing, because someone still needs to be accountable in signing off.”
Presenters
Joshua Grimm
Director, Forvis Mazars APAC Energy, Singapore
Hitansh Doda
Director, Forvis Mazars APAC Energy, Sydney