Learning & Professional Development

What Are the AFM and CFM Worth in the Age of AI?

On craft, judgement, accreditation – and why I am still working towards the CFM.

There is a slightly uncomfortable question I have been avoiding.

I hold the Advanced Financial Modeler accreditation, the AFM. I am also still working towards the Chartered Financial Modeler, the CFM, and I still very much want to get there.

Neither is particularly easy. The AFM asks you to sit down for four hours, under supervision, with Excel and a business case, and build an integrated three-statement financial model from scratch. There is no template waiting patiently for you, no helpful colleague appearing just as you reach the unpleasant bit, and no AI quietly offering to sort out the formula you have been pretending not to stare at for the last twenty minutes.

I rather like that. The trouble is that outside the examination room, the world is changing rather quickly.

AI can now help build financial models considerably more effectively than it could even a year ago. ICAEW noted in June that there had been a genuine step change in the quality of AI-generated models, while also warning that they still require close human review because errors, hallucinations and wonderfully confident nonsense have not gone away. Ian Schnoor, CFA, CFM of Financial Modeling Institute (FMI) makes essentially the same point: AI modelling is improving rapidly, but using it properly still requires somebody who understands the ingredients well enough to challenge the result.

So there is a fair question sitting in the middle of all this.

What, exactly, are the AFM and CFM worth in the age of AI?

It is a particularly inconvenient question for someone who already holds one of them and is still stubbornly pursuing the other. Which, I suppose, is one reason to ask it properly rather than merely reassure myself.

Let us make the case against them first

There is a very easy version of this article. AI is changing everything. Human judgement is therefore more important than ever. Professional accreditation is therefore more important than ever. Everyone agrees, nobody has been upset and we can all go for lunch.

I do not think the issue is quite that tidy.

The AFM remains, at heart, a practical build examination. Candidates receive a business scenario and have four hours to turn it into an integrated financial model. That is a serious test of applied skill rather than a multiple-choice test of modelling vocabulary.

It is also a test centred on an activity that AI is becoming increasingly capable of helping with.

That tension becomes considerably harder to ignore when you read The Human Financial Modeler, the first white paper of the Financial Modeling Global Leaders Council. The Council consists of 63 modelling professionals across 26 countries and five continents. The report itself is careful about what that means: this is an experienced expert panel, not a statistically representative survey of every financial modeller in the world.

There is another disclosure worth making. I am one of those 63 Council members.

The report does not recommend changes to the AFM or CFM, and it does not set out an FMI accreditation strategy. It deliberately presents the Council’s findings without prescribing what should follow from them. The conclusions I draw here are mine.

One of those findings is nevertheless rather inconvenient for any build-heavy assessment.

The Council was asked to rank six phases of the modelling lifecycle according to where human expertise would remain most critical as AI improves. Scope came first. Build came last. Sixty-three per cent of Council members ranked Scope – the work of understanding why a model exists, who will use it and what decision it is supposed to support – as the single most critical phase. Only one of the 63 put Build in first place.

That does not mean building has ceased to matter. But it should at least make us curious about what exactly a future modelling qualification ought to certify.

If the part of the job most exposed to automation is also the part an examination spends four hours testing, there is a genuine risk that assessment becomes exceptionally good at certifying yesterday’s centre of gravity.

I do not think that makes the AFM obsolete. Quite the opposite, as I will come to. But dismissing the question because the current exam is difficult would be too easy.

The machine is no longer just helping with the homework

The rest of the GLC research makes matters more awkward before it makes them easier.

Eighty-six per cent of Council members expect AI to reduce significantly the time required to build financial models within the next five years. Seventy-eight per cent expect the profession to shift from building models towards supervising model-building systems. Perhaps most strikingly, 65% believe that AI can already perform at a level that would pass a typical entry-level financial modelling interview task.

Apparently the machine is no longer merely helping with the homework. It may already be capable of passing the interview.

There is an important qualification here. An accreditation examination does not have to reproduce an ordinary working day. In fact, there may be very good reasons why it should not. Exams routinely isolate foundational abilities that everyday tools normally support, precisely because the purpose of the exercise is to find out what the person understands rather than what the surrounding technology can do for them.

So the argument cannot simply be: the workplace is changing, therefore the exam must imitate the workplace.

Still, AI changes what we ought to ask of assessment. If a machine can increasingly produce the output, the finished output tells us less about the unaided capability of the person sitting beside it.

We could respond by saying that AI is not good enough yet. That is true, but it has the unfortunate weakness of depending on AI remaining not good enough.

Technology has rarely been kind enough to organise itself around our preferred professional boundaries.

ICAEW’s current guidance is more useful because it takes the capability seriously. It suggests treating an AI-generated financial model as a draft that needs to be interrogated: check its structure, assumptions, hidden content, debt schedules, capacity constraints, formulas and internal checks, and watch for dead numbers or artificial balancing plugs.

That is not a description of useless technology. It is a description of useful technology that still needs a competent professional on the other side of it.

And that, I think, is where the more interesting case for accreditation begins.

We may be confusing production with competence

There are two questions here that are easy to mix up.

Will humans continue to build every financial model themselves?

And do humans still need to know how financial models are built?

The answer to the first may increasingly be no. The second is a different matter.

Despite ranking Build last in future human criticality, 94% of the GLC agreed that knowing how to build models by hand remains essential for developing strong financial judgement, even as AI improves. The same 94% said human oversight and validation will remain essential for AI-generated models used in decision-making.

At first glance that sits oddly beside the low ranking of Build. I think the apparent contradiction disappears once we distinguish between production and development.

Build may become less valuable as a production activity while remaining essential as a developmental activity.

We may spend less of our careers constructing every component manually while still needing to understand construction well enough to supervise, review and challenge whatever does more of that work for us.

There are plenty of precedents. Accountants no longer maintain ledgers by hand, yet we have not decided that double-entry bookkeeping is an eccentric interest best left to museums. Engineers use software for calculations nobody expects them to reproduce manually throughout the working day, but an engineer who did not understand the calculation would still make people nervous. Pilots fly aircraft containing a quite remarkable amount of automation; aviation did not respond by concluding that understanding the aircraft was now somebody else’s problem.

Financial modelling may simply be arriving at its own version of that distinction.

There is some independent evidence pointing the same way. CFA Institute reported this year that financial employers increasingly want people who combine AI and coding literacy with traditional analysis, financial modelling, strategic judgement, critical thinking and communication. It specifically notes that Excel and modelling knowledge are useful precisely because they allow analysts to interrogate and audit AI-built models rather than simply admire them.

This is slightly irritating for those of us hoping that new skills might politely replace old ones and leave the total amount of required competence broadly unchanged. It increasingly looks as though we may have to know both.

Perhaps the AFM is becoming less representative – and more important

That creates an odd possibility.

The AFM may gradually become less representative of how every professional model is produced in day-to-day work without becoming less relevant. It may even make one part of its purpose clearer.

If the exam is trying to answer, Can this person build faster than AI?, its future looks fairly bleak. Machines are annoyingly well suited to competitions involving speed, repetition and not needing coffee.

But if the question is, Does this person understand the craft deeply enough to operate independently of the machine?, the examination looks rather different.

Can the candidate translate a business case into financial logic? Do they understand how the statements fit together? Can they construct schedules that actually reconcile rather than merely look convincing? Can they recognise when an output is mathematically possible but commercially absurd? And could they still function when AI is unavailable, inappropriate, prohibited or simply wrong?

Seen that way, the AFM becomes something like a verified human competence floor.

There is also no need to pretend FMI has failed to notice any of this. Its current AFM positioning explicitly acknowledges that AI can generate financial model outputs faster than before and places increasing emphasis on challenging assumptions, validating logic, explaining conclusions and bringing judgement to the output. Ian Schnoor has publicly made the same broader argument: strong modelling skills remain important because somebody has to understand what the AI has built and ask sensible questions of it.

In fact, there is a pleasant irony in some of the strongest challenges to traditional modelling accreditation emerging from research convened by FMI itself. That does not suggest an institution refusing to engage with the change. It suggests one that has already started asking the uncomfortable questions.

For me, the remaining question is what those findings should eventually mean for the assessment itself.

The AFM may be testing exactly the capability we most need to preserve, while no longer testing enough of the capabilities the profession will increasingly need around it.

That is a more interesting problem than deciding whether the qualification is either timeless or doomed.

The real danger may be us

The most obvious risk with an AI-built model is that the AI makes a mistake. We know how to deal with mistakes. We review them, test them, reconcile them and, with any luck, fix them before the investment committee gets involved.

The more uncomfortable risk is what happens if fewer people remain capable of doing that review properly.

When GLC members were asked about the risks of AI-generated models, their leading concerns were strikingly human. Reduced modeller understanding was selected by 59%, hidden or opaque logic by 51%, automation bias by 48% and skill atrophy by 46%.

The problem is therefore not simply that AI may continue to make mistakes. It is that humans may gradually become worse at recognising them.

And, paradoxically, better AI could make that problem more serious.

Obviously poor output is easy to reject. The dangerous model is the one that looks excellent and is wrong in a way that requires experience to notice.

The GLC report contains a particularly unpleasant example from a Council member who had seen AI effectively use a hard-coded balancing adjustment to make a balance sheet look correct. On first inspection, the model appeared plausible.

That is a much more interesting failure than a #REF! error because it attacks the reviewer’s confidence rather than their eyesight.

AI will get better at avoiding these problems. It will also get better at finding them. I have little doubt about either. But somebody still has to decide whether the model can be relied upon and whether the review has been sufficient for the decision being made.

At that point, technical foundations stop looking like nostalgia and start looking rather more like risk management.

There is an assumption here worth challenging

I should also challenge my own argument.

I am implicitly assuming that learning to build models remains one of the best ways to develop modelling judgement. The GLC strongly leans that way, but a survey of experienced practitioners does not prove that there is no equally effective route.

It is entirely possible that we will develop better training methods. AI-assisted exercises could expose juniors to more varied situations than a traditional apprenticeship ever could. Simulation might give someone deliberate practice in problems they would otherwise encounter only after several years. Reviewing dozens of deliberately flawed AI-built models might even develop certain instincts faster than spending months copying formulas across worksheets.

The GLC itself is not unified on the traditional route. Forty-eight per cent favour juniors learning to build by hand before using AI extensively, while 19% think training should shift more towards reviewing and validating AI outputs. Another 17% expect entirely new junior roles to emerge.

I would not bet against new forms of training. In fact, I hope we find them.

What I am not yet comfortable doing is assuming that we can remove a large part of the old apprenticeship before we know what is replacing the learning that came with it.

That would be a rather large experiment to run on an entire profession.

Where does the next generation acquire the scars?

Part of the reason this bothers me has very little to do with exams.

How do people become experienced financial modellers?

Historically, one part of the answer has been that they spend quite a lot of time being inexperienced financial modellers.

They build schedules, trace formulas, break things and repair them. They inherit workbooks containing formulas apparently written during a period of considerable personal distress. They discover that a model which balanced yesterday has chosen not to today, presumably for reasons of its own. They lose an unreasonable amount of time to a misplaced sign and, at some point, begin to develop a sense for when something is wrong before they can fully explain why.

A number feels wrong.

The formula works.

The check says OK.

The formatting is immaculate.

Still, something is off.

That instinct is not magic. It is accumulated experience wearing a nicer jacket.

The GLC report asks, in effect, what happens to that developmental path if much of the routine work that trained junior modellers is automated. Nearly half of the Council still wants juniors to learn the craft by hand first. Others see training moving towards review, validation or entirely new roles, and a smaller group thinks the traditional junior role could largely disappear.

Nobody knows yet which of those routes will dominate.

What matters is that automating the apprenticeship task does not automatically preserve the apprenticeship lesson.

If AI removes some of the work through which people used to learn, we need to be rather deliberate about what replaces that learning. Otherwise, a few years from now, we may find that we have become exceptionally efficient at producing models and rather less efficient at producing modellers.

This is where accreditation might have a role beyond signalling competence to employers. It can create deliberate practice in skills that a modern workplace may no longer force people to acquire naturally.

Seen from that angle, a four-hour human-only build begins to look rather less quaint.

The AFM and CFM do not face the same AI question

It is worth separating the two qualifications because AI does not challenge them in quite the same way.

For the AFM, the question is fairly obvious: if manual construction becomes a smaller part of everyday production, why continue to certify it?

My answer, increasingly, is that this may be exactly the foundation worth protecting.

The CFM faces a subtler question.

FMI positions it as an advanced accreditation for AFM holders who want to demonstrate higher-level modelling ability, creativity and complex problem solving. Candidates have four hours to resolve a selection of discrete modelling problems in Excel. They must identify the source of the issue, develop an appropriate solution and execute it. The learning materials also include areas such as model checking and error correction, so it would be unfair to describe the CFM as merely a harder build exam.

That already feels closer to the future role of the profession.

It is less Here are the instructions; please build this and more Something complicated has happened; work out what is going on.

I still think the CFM raises another question, though. If advanced modelling increasingly involves judgement, review and accountability, should advanced assessment eventually make more of those abilities explicit?

The GLC data makes that worth considering. Asked what would distinguish the best modellers in five years, 92% selected Business Judgement, 75% Communication and 67% Domain Expertise. AI prompting and orchestration mattered too, at 51%, but it did not displace those more human capabilities.

The picture that emerges is not of a future senior modeller who has merely become extraordinarily quick at Excel. It is someone who knows why the model exists, what matters commercially, which assumptions deserve suspicion and how to explain the answer to somebody who does not care in the slightest whether the formula used LET.

That suggests a distinction I find increasingly useful.

The AFM should continue to prove the craft. The CFM could increasingly prove the judgement built on top of it.

Assessment itself is having the same argument

FMI is hardly alone in having to think about this.

AI creates a basic problem for almost any form of assessment: if a machine can produce the artefact we traditionally asked a person to produce, the artefact becomes a weaker proxy for what the person themselves understands.

UNESCO has argued that assessment in the AI age needs to place more weight on higher-order thinking, critical evaluation, creativity and ethical reasoning, and explicitly points towards exercises in which students critique or improve AI output rather than merely reproduce something a machine can already generate.

That does not mean examinations become less useful. There is a good argument that the opposite happens. Once assisted output becomes easy to produce, independently verified competence becomes harder to observe.

The important point is that trusted assessment remains valuable only if it continues to measure the right things.

What seems worth preserving

I would be very reluctant to lose the AI-free AFM build.

Give the candidate Excel, the case and four hours, and see what they can actually do.

No Copilot. No ChatGPT. No agent quietly doing the awkward bit while the candidate adopts a suitably strategic expression.

That is not because I think every professional modeller should spend the next twenty years working that way. I simply want an AFM to continue telling me that the person holding it understands the machinery underneath an increasingly automated process.

Oddly enough, I think that may be more valuable now than it was before AI.

Where I would be tempted to go further

At the advanced level, I would be tempted to introduce more ambiguity.

This sounds slightly cruel, although professional life has never shown much reluctance to do the same.

Real business problems rarely arrive with a beautifully reconciled requirements document explaining that there is one correct modelling treatment and wishing you a pleasant afternoon.

Give candidates conflicting assumptions, incomplete information or an existing model containing something plausible but wrong. Give them a choice between two technically defensible approaches with different commercial consequences. Ask them to distinguish a material issue from something merely untidy.

Make them choose, explain and defend the decision.

Not instead of modelling competence. On top of it.

The worst response to AI would be to abandon the craft. The second-worst might be to preserve the craft and stop there.

And one day, perhaps let AI into a different examination room

Not the AFM core. I think there remains substantial value in proving what someone can do without assistance.

But unaided competence and augmented competence are not the same thing, and eventually both may be worth testing.

Imagine an assessment in which a candidate receives a substantially AI-generated model. It looks good, which is of course what makes the exercise interesting. Somewhere inside it there is a straightforward formula error, an assumption with no commercial justification, an inconsistency between operating and financing logic, and perhaps one treatment that is technically fine but has an implication the model user needs to understand. For good measure, something else that looks suspicious turns out to be perfectly correct.

The candidate may use AI. The point is no longer to beat the machine but to remain professionally in control while working with it.

They review the model, decide what is material, use AI where it helps and ignore it where it does not, make the necessary corrections and explain what still requires independent validation. At the end, they have to decide whether they would be prepared to rely on the result.

One assessment is effectively asking: can you do the work?

Another could ask: can you remain responsible when AI does some of it?

That distinction interests me far more than any qualification in prompt engineering. Today’s clever prompting technique has an unfortunate tendency to become tomorrow’s button. Tools and interfaces will change much faster than professional responsibility should.

The goal should not be to certify today’s AI product. It should be to develop professionals who remain competent when tomorrow’s one turns up.

And perhaps make them explain themselves

I also like the idea, at least as a thought experiment, of a short professional discussion after an advanced assessment.

Why did you choose this structure? Which assumption concerns you most? What would you independently verify before anyone used the model? Would you put your name underneath the result?

Oral assessment brings obvious problems of its own. Examiner consistency matters, as do bias, cost and scalability, so I am certainly not proposing that FMI add a twenty-minute viva next Thursday and see what happens.

But the underlying idea feels relevant. In an AI-heavy environment, the ability to defend a professional judgement may eventually tell us something that merely possessing a polished workbook does not.

There is also the rather stubborn matter of responsibility

The GLC report becomes particularly interesting here.

Ninety per cent of Council members said sign-off on model outputs should never be fully delegated to AI, while 86% said the same about ethical judgement calls within the modelling process.

That sounds reassuring until the next question is asked.

Who should bear primary accountability if an AI-built or materially AI-modified model causes a significant error or financial loss?

On that, there was no clear consensus at all. No option received more than a quarter of the responses.

I find that more interesting than another debate about whether AI can write a formula.

Being a professional surely involves more than possessing a collection of technical skills. At some point somebody has to stand behind the work. Someone tells an investment committee that the downside case has been tested. Someone tells a lender that the debt service calculations can be relied upon. Someone explains why a valuation is reasonable, where it is uncertain and what would cause the conclusion to change.

AI can increasingly help produce that analysis.

Responsibility remains rather stubbornly human.

That is another reason why I think advanced accreditation should eventually concern itself not only with whether a candidate can reach an answer, but whether they understand enough to take responsibility for it.

So are the AFM and CFM still worth it?

I think they are.

But I am less convinced that the best argument for them is the traditional one.

The economic value of some mechanical modelling work is likely to come under pressure. Formula construction becomes easier, routine schedules become quicker and first-pass models become cheaper to produce. A skill that once derived some of its value from scarcity will naturally feel that change.

The GLC itself is strikingly divided on what this means for the profession. Forty-eight per cent see downward pressure on financial modelling as a career over the next three to five years, while 37% see upward opportunity. The report treats that disagreement as a finding rather than trying to manufacture a consensus that is not there.

That feels right to me. We do not know yet.

At the same time, another sort of scarcity may become more important. People who understand models deeply enough to challenge them. People who recognise the commercially absurd answer hiding inside technically correct calculations. People who know when a review has been sufficient, can explain what matters to a decision-maker and are prepared to stand behind the conclusion.

This is not an argument for freezing the AFM and CFM in their current form, nor for replacing them with something fashionable and AI-branded.

It is an argument for being clear about what each accreditation is supposed to prove.

FMI is already moving in that direction in how it talks about the AFM. The current positioning explicitly connects modelling competence with challenging assumptions, validating logic and applying judgement rather than treating spreadsheet construction as an end in itself.

The harder question is how far the assessment architecture itself should eventually follow.

So why am I still working towards the CFM?

There is a personal footnote to all of this. I am still working towards the CFM, and I still very much want to earn it.

That is not because I think AI is going away. Quite the opposite. I use it extensively already and expect to use it more. I have no romantic desire to return to a world in which every formula has to be coaxed into existence by hand while Excel occasionally responds with #VALUE! as though it has developed a principled objection to my career choices. Nor am I particularly interested in racing an AI to see which of us can build a debt schedule faster. I suspect I know how that ends.

What I do want is an independent test of what I can still do when those tools are taken away.

If I am presented with an unfamiliar modelling problem, I want to know that I can still understand what is really being asked, work through the commercial logic, design a sensible solution, build it and challenge the result. And if somebody else, human or artificial, has got it wrong, I want to retain enough of the underlying craft to recognise that too.

That is why the CFM does not feel to me like a rejection of AI. If anything, the opposite is true. The more work I delegate to capable tools, the more important it becomes that delegation remains a choice rather than a dependency.

There is probably something less rational in it as well. I have spent a large part of my professional life doing this work, and I still want to be good at it. Not merely good at instructing something else to do it for me.

There is a slightly bittersweet side to watching machines reproduce in minutes parts of a craft that took years to learn. That does not make the technology bad. It is extraordinary, useful and, every now and then, mildly irritating. There are moments when an AI produces in thirty seconds something that once occupied a meaningful part of my afternoon and my reaction is a mixture of professional admiration and a rather less sophisticated:

Oh, come on.

But expertise was never only the finished spreadsheet. Some of it is what happens to you while learning to produce one: the mistakes, the corrections, the strange inherited models and the gradual accumulation of judgement after seeing enough things go wrong. Eventually you start to recognise when a number feels wrong before you have fully worked out why. That instinct has value, even if the mechanics that helped create it become increasingly automated.

I have no interest in preserving unnecessary manual work simply because earlier generations had to endure it. That would be hazing, not professional development. But neither do I want efficiency to remove important learning without noticing what disappeared with it.

As the tools become more capable, I do not want to quietly become less capable myself.

That is probably the simplest explanation for why I still want the CFM. I am not trying to prove that I can beat AI. I would simply like to know that I understand the work well enough to use AI without surrendering the judgement that makes the work worth doing.

Keep the craft

I do not think the future of financial modelling requires us to choose between humans and AI. AI is entering the workflow and, in many organisations, is already firmly inside it. The more useful question is what kind of modeller we want working alongside it.

Clearly, that person should know how to use the tools. I also think they should understand enough of the underlying craft to work without them when necessary, challenge them when appropriate, explain their outputs and take responsibility when the answer matters.

That is why I still believe the AFM and CFM have an important role. They should evolve with the profession, preserving technical foundations while placing increasing weight on judgement, review, explanation, accountability and, eventually, our ability to supervise AI responsibly.

The answer to automation cannot be to stop understanding the thing being automated.

Keep the craft. Expand the judgement. Teach the supervision.

Sources and further reading

1. Financial Modeling Global Leaders Council / Financial Modeling Institute — The Human Financial Modeller, GLC White Paper No. 1, 2026.
2. Financial Modeling Institute — Advanced Financial Modeler (AFM).
3. Financial Modeling Institute — Chartered Financial Modeler (CFM).
4. ICAEW — How to identify AI errors in financial models, 5 June 2026.
5. CFA Institute — What employers want: A new skills blueprint, 6 March 2026.
6. CFA Institute — Finance Skills Pulse Survey, 1 May 2026.
7. UNESCO — What’s worth measuring? The future of assessment in the AI age, 19 May 2025.