Two conversations are doing the rounds at present:
How do we deal with the trebuchet of seductive verbiage generated by AI, andHow do ‘seasoned’ professionals make the most of their knowledge when using AI? This has a generation-spanning subquestion: how do we develop the thinking of junior team members as they shortcut decades of experience and start cranking AI-powered business shizzle?
Recent reading from other people’s better brains has helped shape this question.
Kelly Goetsch posted on Linkedin about the value of experience (or rather ‘whether it’s valued’).
Harvard Business Review talked getting “trendslop” from LLMs instead of strategic advice.
Friend of the blog, Neil Perkin talked of “premature conversion” that focuses on getting a good-looking answer, while not holding multiple perspectives and tensions in balance as long as possible. ie the drive to a conformant ‘quick answer’ comes at the cost of deeper thinking about the question and nuance.
Shopify’s CEO, Tobias Lütke, surely one of the most ardent and committed exponents of AI use within the organisation, laments the “Slop Grenades” that are quick to create but waste everyone’s time to process and - yes - question.
Taken together, we’re seeing a world of slick answers where the value of experience, breadth of thinking and checking that all considerations have been, er, considered is getting stampeded.
Experience as structured thinking
How do experienced people avoid sounding like ‘grampa’ with lots of ‘in my day we did this’ or ‘that’ll never work because…’? How can we take the benefits of AI’s speed and facility with creation, but bring our own expertise to bear? How can we help a junior colleague to learn the thinking approaches that will help get better answers?
In short, it’s about developing a considered, honed approach to questioning. This can come from many areas.
Turning learning into judgement
Confession time: I’m a Chartered Accountant
Last century, when Lotus1231 was the talk of the town, I trained as a Chartered Accountant in Ernst & Young’s audit practice. I brought all the intellectual tools from my English Literature degree and got rammed through a graduate conversion course, squeezing the basics of accounting, law, economics, and tax into my rapid-recall memory. After 6 months, we were released as bag-carriers at global clients, expected to act as if we were worth our hourly charge-out rate.
EY (and the other global peers) had created a structured approach to having an enquiring mind. Making the ‘habits of enquiry’ second nature. Then they invested in knowledge transfer to layer “good judgement” on top of intellectually-derived facts. EY was a scaled machine at turning raw brains into evaluating, billable entities.
In the first year I was exposed to global publishing companies, Lloyds syndicates, a pickle-packer in Peckham, the Retail Trust’s cottage homes, music royalties and manufacturers. Of course we researched the companies, but I always had the advice and experience of the managers and Partners to call upon.
When I asked “how do you just know that?” the answer would be…
Cake. Cumulative Audit Knowledge and Experience. CAKE
It was a mix of the seen-it-before, done-it-before battle scars, broad experience in similar situations and a questioning approach. All honed into a way to smell the way to a better question or sniff out a dud answer.
The thinking tools
EY invested enormous amounts of time in training and evaluation, and there were three layers of tools drilled into us:
Questioning what’s presented to you - does it stack up (COVED, below)
Assessing the quality of information you use to reach a decision
Does it matter? Balancing the perspectives of legality, importance, purpose, strategy, perception, propriety - is the juice worth the squeeze? it something material?
COVED
Whenever we were presented with information, we had a mnemonic to question it: COVED. Completeness, Ownership, Valuation, Existence, Disclosure. Developments in standards since my time have added in ‘timing’ (to give a less memorable COVEDT?). If we strip out the accounting language2 then we could describe these assessments of AI output as:
Completeness - do we have the whole picture? did you start with a complete picture? What wasn’t asked, what alternative wasn’t generated (Perkin’s push-back-on-convergence, operationalised)
Ownership - whose framing, perspective or position is this; is it germane to the specific situation, or does it dissolve into generic best practice or waffle
Valuation - is confidence proportionate to evidence, or is fluency doing the polishing a turd (the trendslop effect). Have we weighted and tested?
Existence - is the cited precedent real and checkable, or a plausible fakery? Hallucination or fact? And is the fact the right one (cf. Ownership above)
Disclosure - has it shown assumptions and sources, or handed over a tidy answer with the workings hidden
Timing - is this still true, or was it true when the pattern was formed?
GIGO
So that’s our questioning approach, but even the best analysis is flawed if the supporting data is not right.
There are approaches to assess the “enoughness” of information. Is it of the right quality to support the inferences?
International accounting standard (ISA 500) sets out the requirements for “good information” to be used. We’re to have regard to three things:
Sufficiency. Is there enough data upon which to base our opinion
Relevance. Is the information you’re looking at relevant to the question?
Reliability. Can the information you’re accessing be relied upon?
Is the data you’re working with sufficient, relevant and reliable? No additional amount of shonky data can create a better answer!
Other paradigms, other training methodologies, other pedagogical approaches are available - just ask anyone with a professional qualification (from architects and engineers to lawyers or doctors, builders, plumbers, software engineers) - they all attempt to codify judgement and assurance at the input stage.
CAKE
Where then does CAKE come in? That layer of judgement applied on top?
It’s about “Does it matter?” Auditors call it ‘materiality’. This is the threshold value at which a misstatement, error, or omission would reasonably affect the user of the accounts. Users could be the shareholders, tax authorities, suppliers, etc.
Some things are clearly trivial. Take a global oil company generating over $1 trillion in turnover across 6 years. A small sum of $28 million is unlikely to affect the picture of their financial performance. It’s a quarter of one per cent of their turnover in that period.
However, that amount was very significant because it was money paid to officials in five countries to secure contracts. Bribery. Corruption3.
So the amount itself can be small while the materiality is significant. Bribery and illegality are material things, no matter the size. Meanwhile, simply wasting millions on being wrong, inefficient, unlucky, or inept may not be worth noting.
So this is where CAKE comes in. The knowledge of the market, of the drivers of behaviour and compliance, the linked web of “so whats” and “if then” connections. It’s knowing which pebbles cause which ripples, and which ripples form interference peaks.
Commercial leaders aren’t born with this knowledge and experience, and they must continually renew it. An open and enquiring mind, eyes to see and ears to hear, and a knack for correlating cause and effect - these are the lifelong habits that distinguish commercial leaders from face-in-the-spreadsheet followers.
Cutting through slop: the layer cake
This then is the layer-cake of cutting back on slop:
An evidential challenge - are you feeding the engine with information that is sufficient, relevant and reliable to allow it to consider the question
A thinking approach that checks whether you’re “asking the right question”. Is your approach COVED(T)?
Are you checking the output (however glossy) for the connections, the so whats, the implications? This is where risks, resonances and reward probabilities are investigated
The cherry on this analytical cake is whether it matters. What is your action? Does it support purpose, values, commerce?
An LLM has no soul. It has no moral compass. It doesn’t understand risk, pain or reputation. It has no desire to create joy, make the world a better place, improve your wealth, health or retirement options. It converts patterns into words that sound as if they come from an eloquent, privately educated charmer. CAKE is what helps us identify whether this charmer is a cad or not!
Upside down
Now, all of the above looks very pyramidal and processy. The real fun comes because the experienced nose helps sniff out what’s “material” at the outset, thereby directing the work, trimming or extending scope, challenging or ignoring. In this way, there’s a balance between deploying experience and insight alongside data crunching and argument development.
By working together, the skills learned “the old way” can accelerate value from the AI-charged capabilities today, while the neophytes with the LLMs can understand the questioning, scoping and shaping structures that will make them more insightful, more commercial analysts.
When should I think?
To answer my own question, it should be before you get the Slop Grenade, before you ask for a Slop Grenade, and while you’re framing your inquiry.
If you ask a good question, discuss what’s material in your thinking, and scope the data sources for sufficiency, relevance and reliability, then you’ll have a fighting chance that the beautiful initial output will add something to your life, rather than sap your will to live!
What’s your approach to killing slop at the outset?
How do you balance experience with the speed and accessibility of sensible-looking AI slop?
I’ve shared my accounting path, but what path(s) have you taken or combined?
AI will of course be on the Agenda at CustomerX on 14 and 15 October. Cut through the slop to the commercial opportunities - register here.
Our dedicated event, CommerceAI Festival, is on 12 May, 2027 at Olympia. Register your interest on the site.
Sources
HBR, “Researchers Asked LLMs for Strategic Advice. They Got ‘Trendslop’ in Return.”, Romasanta/Thomas/Levina, 16 March 2026: https://hbr.org/2026/03/researchers-asked-llms-for-strategic-advice-they-got-trendslop-in-return
Neil Perkin, LinkedIn post on premature convergence in AI-assisted strategy: https://www.linkedin.com/feed/update/urn:li:activity:7498287725373014016/
Neil Perkin profile: https://www.linkedin.com/in/neilperkin/ | website: https://onlydeadfish.co.uk/
Launched in 1983, this was the killer app of its age, until unseated by Microsoft Excel. We students stared at 6cm cathode ray screens while holding a spreadsheet in our brains… https://en.wikipedia.org/wiki/Lotus_1-2-3
ISA 315 covers the audit assertions (existence, completeness, rights & obligations, valuation & allocation, presentation & disclosure, accuracy/cut-off), if you want to open the bathrobe and see the details - https://www.frc.org.uk/library/standards-codes-policy/audit-assurance-and-ethics/auditing-standards/isa-uk-315/
Search for Glencore’s history for more, or see the US Justice Department here: https://www.justice.gov/archives/opa/pr/glencore-entered-guilty-pleas-foreign-bribery-and-market-manipulation-schemes