Financial Modelling Services & Training

AI Enablement

Practical AI adoption for the teams who build and rely on financial models. We measure how your deal and modelling workflow runs today, design the AI methods and governance around that reality and coach your people until they run it without us.

Send enquiry
  • 1 focusAI in financial modelling,
    not general AI training
  • MonthlyCapability built month by month,
    not handed over in a workshop
  • Your modelsDesigned and piloted on your
    own models, not samples
  • Any stackClaude, ChatGPT, Copilot
    and Gemini all in scope

The problem

The tools are licensed. The modelling has not changed.

Most modelling teams we speak to have already taken the harder step. Claude, ChatGPT or Microsoft Copilot is licensed, security has signed off and people are using it. What has not happened is any measurable change in how deals get assessed, or how financial models get built, reviewed and relied upon.

The gap is rarely the technology. It is that nobody has decided where AI may be relied upon in the deal and modelling workflow and where it may not, what a human must still check and what evidence exists that the check happened. Absent those decisions, sensible people use the tools for notes and emails and quietly avoid them everywhere that matters.

The aim is a team that uses AI where it genuinely helps, avoids it where it fails, and can show the basis of every number that matters.

What usually prompts the call

  • Copilot or ChatGPT has been rolled out and modelling output looks exactly as it did before
  • More opportunities arrive than the team can properly assess, and screening is where they stack up
  • Analysts are pasting model extracts into public AI tools and nobody has set the rules
  • A board, credit committee or client has asked who checked the AI-assisted analysis
  • Model documentation and review are the bottleneck, and AI is the obvious untried answer
  • Everyone agrees AI matters to financial modelling; nobody owns what that means in practice

The evidence

Where AI helps in financial modelling, and where it fails

We do not advise from vendor claims. We seeded 44 errors into a hypothetical project finance model, benchmarked leading AI tools against them and published what we found. Those results shape every engagement.

Where AI earns its place

  • First-pass model review. A structured AI pass clears a real share of the obvious issues in minutes, so expert review starts from a cleaner base.
  • Sense-checking outputs. Consistent definitions, plausible ratios, results that tell a coherent story. The strongest category in our testing.
  • Explaining an inherited model. Walking the logic of an unfamiliar workbook and tracing where a number comes from.
  • Drafting. Model documentation, assumption registers and clear, cell-referenced review comments, written by AI and checked by people.

Where it cannot be trusted

  • Completeness. The strongest configuration we tested found 27% of the seeded errors. On a single pass, AI misses most of what is wrong with a model.
  • Repeatability. The same tool on the same model returns different findings on different runs. Useful in a tool, disqualifying in a reviewer.
  • Structural and timing errors. Detection was weakest on exactly the subtle issues that decide whether a model can be relied on.
  • Carrying the responsibility. AI output still has to be checked by someone who can stand behind the number.

None of this argues against using AI. It argues for deciding, task by task, where output can be relied upon and where a human must still sign. That is what the recommend step produces.

The biggest performance driver in our benchmark was not which tool was used. It was how the tool was asked. A structured review prompt roughly doubled detection; handing the tools precedent models changed nothing. Method separates the teams that get value from AI, and method can be taught.

Read the benchmark study

How we work

We start inside your team, not with a slide deck

Engagements run over months, and most settle into a monthly coaching rhythm once the first methods are live. The shape flexes with adoption; the sequence does not. Only the first two steps need agreeing at the outset. Everything after that follows what the diagnostic finds.

  1. 01

    Diagnose

    How your team builds, reviews and relies on financial models today, and where AI is already in use, sanctioned or not. We sit with the people who do the work, and we leave with a measured baseline you can hold later work against.

  2. 02

    Recommend

    Where AI may be used, at what level and the gate between build and reliance. Claude, ChatGPT, Copilot and Gemini are weighed against your security requirements, not ours.

  3. 03

    Deploy

    Working methods and a versioned skills library encoding your own conventions, proven on your own live work. We keep what measurably helps against the baseline and drop what does not.

  4. 04

    Coach

    Hands-on sessions with your modellers, side by side on real tasks, at a monthly rhythm. New methods stick when they are practised on live work, not in a workshop.

  5. 05

    Certify

    Structured modules with assessment, taught against your own conventions, using your own models. A defensible answer when a lender, board or credit committee asks who did the work.

  6. 06

    Sustain

    The tools change monthly, so the playbook cannot be static. We update methods and prompts as the models change, re-audit a sample of AI-assisted work and stay on as coaches for as long as that is useful.

Scope

What an engagement can cover

Every engagement is shaped by the diagnostic, so no two look alike. These are the areas teams draw on most, whether the constraint is deal volume or the reliability of the models decisions rest on.

SCREEN

Screening and document extraction

Documents in, structured assumption sheet out, every input traced to source. A standard screening model with its reliability tier stated on its face.

BUILD

AI-assisted model build

Drafting structure, writing and refactoring formulas, plus speeding up the mechanical parts of a build inside your own conventions.

REVIEW

Model review and QA

Structured first-pass review prompts, error screening and sense-checking, tiered by module against your own model architecture so the weakest detection categories stay under human review. Layered under your existing review process, which stays in place.

DOCUMENT

Documentation and reporting

Documentation, assumption registers, handover notes and model outputs turned into clear written analysis for boards, credit papers and investment committees. Drafted with AI and checked by people.

GOVERN

Policy and governance

Reliability by task, tiered human checkpoints decision by decision and a written position on what may go into which tool, at what data classification and what a human must still sign.

TOOLING

Tool selection

An independent view across Anthropic’s Claude, OpenAI’s ChatGPT, Microsoft Copilot and Google’s Gemini, matched to your security posture, not to a reseller agreement.

What it takes from you

The real question is access, not budget

Every failed AI programme we have seen failed because it was run on sample data by people who do not do the work. These are the inputs that make the difference.

DEAL

One recent deal file

A single opportunity followed end to end, from the day it arrived to the day it was decided. It shows how the work actually flows, which no model on its own can.

ACCESS

Two or three representative models

The workbooks your team actually works in, with their history and their habits. Not a sanitised extract.

PEOPLE

Time with the people who do the work

One-to-one time with the analysts and modellers who run the processes, rather than with those who own them on a chart. Then regular sessions they actually attend, spent on live work.

ALIGNMENT

Half a day in one room

Modelling, risk and IT together. Rules made separately by each function do not survive contact with live work.

PILOT

Live work to test on

A model or task at a live but not commercially sensitive stage, so the results hold up outside a demo.

OWNERSHIP

One named internal owner

The single most reliable predictor of whether any of this is still live a year after we step back.

Why this team

Audit evidence, teaching experience and published research

We are not a general AI consultancy. Financial models are the only thing we audit, build and teach, and the only place we advise on AI. Three things underpin the advice.

01

We see what goes wrong in models

The number one ranked global model audit practice by IJGlobal for FY2025, with more than 2,000 models reviewed. We advise on AI from direct evidence of how financial models fail.

02

We teach modellers for a living

More than 20,000 professionals trained in our modelling methodology. Embedding new method in a modelling team is what our training practice does every week.

03

We test the tools ourselves

We benchmark AI tools against seeded-error models and publish the results, including where Claude, ChatGPT and Copilot fall short. The 44-error study is one example, and the testing continues as the tools change.

Independent model audit remains a separate engagement, carried out by a different team. We say so up front because the question always comes.

Ask about AI enablement

Tell us how your team assesses deals and builds models today. We will show you where AI genuinely helps, and coach your people through the change.

Send enquirySee our AI modelling course