Treating a probabilistic engine like a calculator risks ruin
Why in consulting, finance, and law, 70% accurate is worse than useless: leverage comes from unleashing AI in ambiguity.
I have spent two decades inside boardrooms and war rooms, advising banks and insurance companies across four continents.
If there is one universal truth in those environments, it is this: credibility is binary. A single wrong number can undo months of effort, derail billion-dollar programs, and permanently stain reputations.
This is why I am baffled when I see so many GenAI products trying to prove themselves in areas where precision is non-negotiable.
Drafting legal clauses.
Building financial models.
Producing compliance artefacts.
They pitch the promise of speed, but the hidden cost is trust.
70% accuracy is not “better than nothing” in domains that demand perfection. It is useless. In fact, it is worse than useless because it lulls you into a false sense of security.
A financial model that is 70% correct is like a bridge where 30% of the beams are missing: it might look impressive from a distance, but you would never dare walk across it. And yet, I see executives trying to walk across those bridges every day, convinced that GenAI has given them an edge.
The reality is that if I cannot predict where the errors are, I am forced to check everything.
Fix one formula, and another hallucination (a more-or-less fancy word for “invented number”) pops up. I end up shifting my human effort from construction to auditing, while the actual modelling muscles of the team slowly atrophy.
To anchor this scientifically: large language models are statistical sequence predictors. They optimize for plausibility, not verifiability. Their accuracy distribution follows a long‑tail error curve: most outputs are “good enough,” but the unpredictable outliers are catastrophic. In probabilistic terms, variance is high even when mean performance looks acceptable.
In deterministic domains like finance or law, these rare but severe errors dominate the risk profile. This is akin to Nassim Taleb’s concept of “ruin risk”: one failure, no matter how infrequent, can erase the entire upside.
In this LinkedIn post I discussed the case of a Big4 consultancy that delivered a LLM-generated $439K artefact to a government agency in Australia, with significant reputational repercussions.
Where AI should play
We said that GenAI is a probabilistic machine. It thrives in the fog. Its native domain is ambiguity, not arithmetic.
When I sit with a client and face a blank slide, the hardest part is not perfecting the numbers, but framing the problem, surfacing the unspoken assumptions, testing hypotheses, exploring scenarios.
This is where “70% right” is gold. Because 70% is enough to expand thinking, provoke discussion, and unlock new angles.
The smart play is to redeploy GenAI into ambiguity, creativity and iteration speed. Let me explain:
Ambiguity: In practical terms, this means helping consultants detect the subtle coalitions, unspoken fears, and implicit trade‑offs that often derail programs more than any technical issue. GenAI can process thousands of emails, meeting transcripts, or feedback threads and cluster themes we would never spot manually. The output is not deterministic, I get it, but it acts like a cognitive radar, showing where the fog is thickest and where attention should go.
Creativity: Human teams often converge too quickly on the first idea that sounds credible. A probabilistic engine resists that premature closure: it generates divergent options, metaphors, and framings that stretch the conversation. Neuroscience shows that creativity emerges from recombination more than invention; LLMs are powerful at recombining, surfacing analogies across domains in seconds. This accelerates the “adjacent possible”, solutions just one step beyond what we already know.
Iteration speed: Strategy lives in iterations, not in single shots. With GenAI, we can explore multiple narrative arcs (eg, a cost story, a risk story, a growth story) and stress‑test them with stakeholders before committing. This is some sort of rapid prototyping of mental models. By burning through cycles faster, we compress the learning loop, which is the essence of competitive advantage in fast‑moving environments.
In these domains, being “almost there” is a catalyst.
About 15 years ago, I remember working on an investment accounting transformation in Asia, surrounded by twenty consultants who were burning 80 hours a week for almost two months straight.
The modelling had to be flawless. Every journal entry had to reconcile, every test had to tie out, every decimal point mattered. In that crucible, a 70% solution would have been laughed out of the building.
But later, when we needed to present options to the executive board (should we replatform?, should we buy a vendor?, should we build in-house?) that was where ambiguity reigned. That was where we needed to explore confusing “what ifs.”
An AI tool that could spin scenarios and reframe arguments quickly would have been worth its weight in gold. It would not have replaced the accountants and the finance experts, but it could have accelerated the consultants.
My takeaway
Executives who jam GenAI into deterministic work are committing a category error: they are treating a probability machine as if it were a calculator.
But a calculator is deterministic (2+2 is always 4) whereas an LLM generates outputs by sampling from probability distributions. This fundamental mismatch means leaders are betting deterministic credibility on stochastic processes. And in doing so, they risk both productivity and trust, because every unnoticed error compounds into reputational fragility.
The smarter use of GenAI is not to force it into the world of zero tolerance but to unleash it in the world of multiple possibilities.
What differentiates a consultant is, after all, not how quickly they can crunch numbers: machines will always outpace us there! Our edge is turning political minefields into solvable equations, or crafting a story so crispy it forces a decision.
Numbers only carry weight when someone has the guts to interpret them, defend them, and weaponize them into a tangible Outcome. That is the work that bends billion‑dollar outcomes.
And that, I believe, is where GenAI can finally shine… not a modern shortcut to a broken spreadsheet.
As usual, if you enjoy reading Consulting Intel, please do me a favor: spread the word and share this post.
👋
👀 Links of interest
A few corners of the internet you may find interesting:
I have started to be more active on LinkedIn in the last month. Would be awesome to connect with you over there as well. Feel free to follow or connect with me!
The Leaders Toolkit is a deck of 52 tools, frameworks and mental models to make you a better leader (use code CONSULTANT10 for 10% off);
The Consulting Intel private Discord group with 250+ global members is where consultants meet to discuss and support each other (it’s free).
