Too Easy to Believe
AI, consulting and the price of professional trust
Half the working day is probably already behind you. If there is coffee within reach, now might be a good time to pick it up: this is not going to fit into three paragraphs and an instructive final sentence.
https://unsplash.com/@cmilan1
It is becoming difficult to talk about the next few years in consulting without AI appearing within the first few sentences. There is a fairly tangible reason for that. If parts of analysis, research and presentable output can be produced substantially faster, the position of three different players changes at once.
It is a little like moving on from candles, losing our fear of electricity and finally daring to switch on the light. It does help, of course, if a competent electrician has checked the wiring first and made reasonably sure that using the switch will not electrocute us.
With AI, we are roughly at that stage. The light is unquestionably brighter, the switch is convenient, and everyone would quite like to use it. Someone still needs to understand what is running through the wall. The user also needs to know that if a little thing is sticking out of it, with yellow metal inside and plastic around it, picking it up and examining it while wondering hmm, what could this possibly be? is probably best avoided.
Consulting firms need to recalculate where the value sits for which they can continue charging a premium. The people working inside them need to work out which parts of their knowledge and labour will remain scarce. Clients would do well to become rather more precise about what they are buying: information, analysis, access, judgement, responsibility — and which of those they actually received.
An AUD 440,000 Warning
There is already a fairly tangible example.
In 2025, Deloitte Australia produced a 237-page assurance review for the Australian employment department for AUD 440,000. The report was later found to contain non-existent academic references and a quotation supposedly taken from a Federal Court judgment that had never actually appeared in it. The revised version disclosed that an Azure OpenAI GPT-4o-based tool had been used for part of the work. Deloitte ultimately refunded the final instalment of the contract.
The easy conclusion would be: there we are, AI hallucinated again. Except that for an AUD 440,000 expert report, the client was hardly buying advice from a language model. It was buying it from Deloitte.
A consulting firm can, of course, use AI. It can use search engines, databases, modelling software, Excel spreadsheets or anything else that makes the work faster and better. For the client, however, the name on the finished product carries the promise that somebody checked the wiring before the light was switched on.
In internal correspondence, the Australian department explicitly raised the possibility that the citation problems affected confidence in the report’s findings. At that point, AI has already moved beyond a technology discussion.
A model can make mistakes, as the accompanying exhibit rather helpfully demonstrates, and humans are hardly strangers to the activity either. At the end of the day, somebody still has to check the machine’s work, notice the error, put their own name to the result, and the client still buys somebody’s professional credibility.
So What Is a Partner For?
In a recent post, Ben Appleton lands on much the same issue through McKinsey’s new “distinguished partner” role. He argues that as AI reshapes consulting, domain expertise, commercial judgement, senior relationships and the ability to turn complex specialist knowledge into organisational change will become more valuable. His word for the next generation of partners is “translator”.
Rather than the approving I too think something impressively sophisticated comments that so often flood a thread, thoughtful professional objections and additions made consulting’s still-living reason for existence unusually visible.
Julius Hollmann thought much of this sounded rather familiar. A good McKinsey, Bain or BCG partner has always been expected to understand the client, grasp the problem, connect different disciplines and carry a decision through. He also argued that T-shaped, lateral and out-of-the-box thinking may become more valuable, while some specialist roles narrow. Samik Das approached the change from a similar direction: in his view, the partner’s role has essentially always involved change management, while the more immediate organisational consequences of AI may emerge closer to engagement-manager level.
That is an awkwardly good question for McKinsey’s new label. If the most important capabilities of the “partner of the future” turn out to be the same ones that made a partner good yesterday, there are at least two possibilities. Either the relative weight of those capabilities is genuinely changing, or the industry is repackaging an old expectation in AI-compatible language.
Here, “translator” already means rather more than rendering technical language comprehensible to a board.
Rolled into a single “future partner” profile, all of this could quite easily become another carefully wrapped competency list, lacking only “strategic mindset”.
The comments pulled the package apart. Luca De Angeli suggested “orchestrator” rather than “translator”: somebody has to connect domain expertise, strategy, technology and the operating reality of the business. Olivier Burnouf, referring to a McKinsey interview with Sven Smit, added the data constraint: much of the most interesting information used in larger research exercises simply is not sitting there waiting for AI. It may be proprietary data, newly created information, or knowledge held by people. Sridhar Peddisetty separated the production of intelligence from the decision itself. Vinod Venugopal put CFO-quality, traceable data underneath “qualified subjectivity”.
It becomes harder to describe the partner’s job as simply translating whatever the machine says into understandable human language. Someone first needs to know what the machine worked from, what was missing, where professional judgement entered the process, and how far the final product placed in front of the client can be traced back.
What Trust Does
Trust performs a very concrete function in this chain. A board receives an analysis, listens to the people behind it, asks questions, argues and eventually decides. It does not reconstruct every data source, rerun every model or spend three days checking footnote 184. If it had to, very few decisions would ever get made.
At some point, it accepts that other people have done work it is not going to repeat. A Deloitte, McKinsey or BCG logo forms part of that. So does the partner’s name, the method, the origin of the data, the checking process and the fact that, if something goes wrong, there is somebody available to answer the question: how exactly did you manage this?
Within such a system, trust saves time, permits division of labour and allows decisions to be made even where nobody can see the entire chain alone. Which is also why discovering that it was placed badly can become expensive.
In the Deloitte case, that assumption was damaged: that somebody had actually looked through what went into the finished product.
The operational consequences of trust become particularly visible when the issue is no longer the credibility of a report or a client relationship, but whether an institution gets money again the next morning.
Monday Morning, 2008
If you are experiencing a slight sense of déjà vu: yes, the time machine is taking us back to 2008.
For an entire weekend that September, bankers, regulators and political decision-makers tried — to put it more sophisticatedly — to assess the chances of the system surviving. Less sophisticatedly: who would still be standing on Monday morning. And, as a minor footnote, that particular Monday morning went on to shake rather a large part of the world.
Without taking another trip through the entire mechanics of Lehman’s collapse, I covered that in more detail in Too Big to Believe: The behavioural logic behind Lehman’s collapse and the role of confidence. Here, I am pulling out only the thread of trust. The length of this article is already making a reasonable attempt on the reader’s patience.
Lehman’s short-term funding continued to function for as long as enough counterparties kept saying yes to financing it for another day.
The bank relied heavily on short-term funding. In the repo market, it repeatedly had to convince the other side that the collateral was acceptable, the money would come back, and Lehman would still be there tomorrow. While that continued, Lehman could maintain its short-term funding. Once increasing numbers of counterparties no longer wanted to take the same risk for another day, Lehman ran out of time very quickly.
Nobody needed a perfect view of Lehman’s entire balance sheet for that to happen. It was enough for growing numbers of counterparties to become uncertain about assumptions they had previously been willing to make. Trust here was not a mood. It had funding consequences.
None of this makes AI the next financial crisis, nor does a consulting report operate like the repo market. Even my imagination would have to work rather hard to sustain that comparison.
The narrow common point is that in both cases somebody has to make a decision without being able to verify everything independently. Somewhere in the process, information, a person or an institution has to be judged reliable enough for the next step to be taken.
That bleak weekend in September 2008 shaped the financial world for years. Eventually, the system recovered. Would the consulting industry have as much time if its own ominous Monday morning really were approaching?
If the friend walking beside us has just stepped into a brown, suspiciously well-formed pile, and a few metres ahead there is another one directly in our path, analysis still has a certain practical use: we can walk around it.
On this still-hypothetical Monday morning, the trouble would look rather different. Clients would simply start calculating differently: what they are paying for, what they can produce themselves, and which parts still genuinely require the person or firm whose name, judgement and professional responsibility they have been buying.
When the Client Discovers How Easy It Is to Press Enter
One of consulting’s longstanding advantages has been information asymmetry. More markets, more benchmarks, more experts and more hours of human work sat on the other side of the table. Clients paid for that access and processing too.
The Enter key has already shaved a surprisingly large piece off it. You still have to ask a question first, of course, and how you ask it is not entirely irrelevant. But even a less tech-savvy executive can reach a usable first analysis within minutes that, only a few years ago, might have set at least a small project team in motion.
That does not make the executive a McKinsey partner. It does, however, give them something against which to compare what the McKinsey partner puts in front of them.
The client will probably continue not to recheck every reference in a 237-page report. That is part of what external expertise is bought for. The material arrives, questions are asked, a decision is made, and everyone moves on. If something later goes seriously wrong, the reconstruction can quite easily reach the lawyers’ desks.
At that point, the origin of the data suddenly becomes interesting. So does which model touched it, who checked the output, who approved it, what the contract promised and whose name appeared at the end.
In the Deloitte case, the name carried the weight: Deloitte delivered the report as a professional product, and the client was paying for that name as well.
AI is beginning to pull at the same business model from two directions. Expert-looking output can be produced increasingly quickly and cheaply. Checking it, standing behind it professionally, defending it, implementing it and still being there months later has not become a cheap pastime.
Clients, meanwhile, have more and more tools with which to reach a usable first answer themselves. A growing share of the consulting premium will therefore depend on what happens after Enter: whether somebody spots the missing data, stops the weak inference, stands behind their own professional judgement, and remains there when things go wrong.
If those elements are missing, an AUD 440,000 report eventually prompts a rather prosaic question:
Why, exactly, did it require a consultant?