A competent consultant can now turn interview notes into a polished market map before lunch. The deck may look expensive. The production cost was not.
That does not make AI consulting fake. It changes what the client should be buying.
Traditional consulting can finish with a recommendation. AI consulting should finish with a working decision path, a measured result and somebody who owns the failures. When the deliverable is still a deck, generative AI mainly makes the old product cheaper to manufacture.
The deliverable has changed
Classic consulting is good at framing an ambiguous problem, collecting evidence, comparing options and helping senior people agree on a direction. Those skills still matter.
AI work adds a second obligation. The recommendation has to survive contact with data, systems and users.
A useful AI consulting engagement therefore includes things a strategy project can leave to somebody else.
- A working integration with the systems already in use
- A test set drawn from real cases
- Defined permissions and audit records
- A human route for exceptions
- Monitoring after release
- A named owner for maintenance and change
The centre of gravity moves from presentation to operation.
Advice is no longer scarce
Research, synthesis and first draft analysis used to consume a large share of a consulting team's time. Models compress that work.
The most useful public evidence comes from Harvard Business School and Boston Consulting Group. Their field experiment with 758 BCG consultants found that GPT-4 increased speed by more than 25 percent, human-rated performance by more than 40 percent and task completion by more than 12 percent for work inside the model's capability boundary.
The same experiment included a task outside that boundary. The control group reached the correct answer 84.5 percent of the time. The two AI-assisted groups reached 60 percent and 70.6 percent. Combined, AI reduced correctness by an average of 19 percentage points.
That result is a better description of AI consulting than most vendor pages. The tool creates leverage where the task fits. It creates confident damage where it does not. Expertise is the ability to know the difference and design the work around it.
The funny failures are assurance failures
Large consultancies have supplied several useful case studies, although not the kind their marketing teams would choose.
Deloitte sold assurance and had to refund part of the fee
Deloitte Australia produced a report for the Australian Department of Employment and Workplace Relations. The government paid AUD 440,000. The published report contained nonexistent academic references and a fabricated quotation attributed to a Federal Court judgment.
Deloitte later confirmed that generative AI had been used and agreed to a partial refund. The department published a revised version of the report. The Guardian's account documents the errors and refund.
The expensive part was not that a consultant used AI. The expensive part was that elementary source checking did not catch the output before delivery.
PwC left a ChatGPT trail in thought leadership
A 2026 investigation by GPTZero examined four PwC Middle East reports published between 2024 and 2026. According to City AM's report on the findings, the documents included unverifiable claims, inconsistent citations and one citation URL that pointed to ChatGPT.
GPTZero estimated that one report had an 84 percent probability of being entirely AI-generated, rising to 100 percent when its reference section was excluded. AI detection scores are not proof by themselves. The broken references and unsupported claims are the more important evidence.
PwC said it was updating a limited number of supporting citations and that staff were expected to follow its quality controls.
There is a small comic precision to leaving ChatGPT in the bibliography. There is a larger commercial lesson in publishing claims that the named sources do not support.
KPMG published an AI trust lesson by accident
KPMG withdrew a report on agentic AI after UBS, NHS Greater Manchester, Swiss Federal Railways and Transport for London challenged descriptions of their work. Inc reported that the organisations considered information in the report incorrect or misleading.
Whether generative AI produced the errors has not been proven. That distinction matters. The confirmed failure was editorial control. A firm selling AI trust published case studies that the organisations named in them did not accept.
Big Three firms face the same boundary
There is less public evidence of an MBB firm shipping a hallucinated client report and being forced into the same kind of correction. That absence should not be filled with a made-up scandal.
The BCG experiment shows the underlying risk cleanly. Smart consultants with a capable model still became less accurate when the problem sat outside the model's reliable frontier. Intelligence, brand and confidence did not remove the failure mode.
The safe conclusion is modest. Large firms are not uniquely careless, and they are not protected by their review hierarchy. Any delivery model that rewards speed without preserving source checks will eventually publish something embarrassing.
How AI consulting differs in practice
The difference appears in the unit of work.
Traditional consulting often organises around a workstream. AI consulting should organise around a decision or action that can be observed end to end.
The evidence changes too. A polished benchmark is weaker than a test on the client's own cases. A model evaluation is weaker than an operational evaluation that includes permissions, latency, escalation and failure cost.
The team shape changes. Research and slide production need fewer hours. Integration, evaluation, product ownership and domain review need more attention.
The commercial model should change with it. Hourly billing rewards the amount of work produced. A narrow fixed milestone or an outcome-linked fee makes the working result harder to avoid.
What a buyer should demand
Before buying AI consulting, ask to see the artefacts that make the work accountable.
- The source register behind important claims
- The real cases used for evaluation
- The decision boundary
- The permissions granted to the system
- The exception queue and its owner
- The monitoring view
- The rollback procedure
- The acceptance criteria
- The production metric tied to value
A provider that cannot show these items may still be useful for exploration. It should not be trusted with autonomous action.
Where the market goes next
AI consulting will not replace classical consulting. It will split it.
Some work will become cheaper because research and presentation production compress. Some work will become more valuable because system integration, governance and domain judgment are hard to automate. Firms that keep selling hours and slides will face price pressure. Firms that own a working outcome can charge for the risk they remove.
The strongest consultants will still frame the right problem and handle the politics around change. They will also know how the software works, how it fails and how to measure it after launch.
The joke is not that consultants use AI. Everyone serious will use it. The joke is charging for certainty while skipping the checks that create certainty.
AI consulting earns its name when the advice runs, the evidence holds and the result survives Monday.