In September 2022, two months before OpenAI introduced ChatGPT, I proposed a different operating model for consulting.
The original post imagined a self-service platform through which clients could deploy a firm's research, frameworks, analytical models, diagnostics, benchmarking tools, bots and experts. Capabilities could be bought individually or in bundles. Revenue would combine access, usage and expert-support fees. Human involvement would fall as far as the client wanted.
The phrase I used was:
with as little human intervention as desired
Nearly four years later, the direction looks recognisable.
The operating model looks incomplete.
That gap is more useful than being proved right.
What held
The central call was not simply that consultants would use more technology. AI sat on both sides of the model.
It would reduce the labour required to produce consulting work. It would also become part of what consulting firms packaged and sold.
Both are now visible. HBR describes research, modelling and analysis as junior consulting work exposed to automation. The firms themselves are turning accumulated knowledge and delivery methods into reusable technology:
- McKinsey says Lilli searches and synthesises its knowledge base, and that versions of the underlying architecture are being adapted for clients.
- EY launched EY.ai after US$1.4 billion of investment.
- Deloitte describes Zora AI as role-based agents that can operate with client and third-party agents.
- IBM describes Enterprise Advantage as a consulting service built from reusable AI assets and expertise.
These are not identical products. Some are internal platforms. Some are services assembled around software. Some help the client build its own environment.
But the movement is clear: consulting delivery is becoming more modular, repeatable and software-mediated.
The 2022 warning that this would be a "big lift" also held. Turning a people business into a hybrid product-and-services business forces decisions about what can be standardised, how tools connect to client systems, who approves automated output, how software revenue interacts with project revenue and what the firm is prepared to stand behind.
The technology was only one part of the job.
What has not held, at least not yet
The original post went further. It proposed a consulting app store with subscription, usage and expert-support pricing.
Major firms do not publicly break out material revenue from that model. Their product descriptions still combine technology with implementation, engineering, governance, change and expert support. Deloitte says a Zora AI implementation is generally measured in weeks or months. IBM calls Enterprise Advantage an asset-based consulting service, not a stand-alone product.
The modularity arrived faster than the self-service economics.
That may change. But there is an important difference between making expertise accessible through software and building a low-touch software business from expertise. The first is well under way. The second remains much harder to see at scale.
What I overstated
The 2022 post called the challenge existential.
That was too broad if it meant the large firms themselves were likely to disappear.
Accenture reported fiscal 2025 revenue of US$69.7 billion, including US$2.7 billion of generative- and agentic-AI revenue and US$5.9 billion of related bookings. One firm does not settle the future of an industry. It does show why "AI destroys consulting" is the wrong frame.
AI creates demand for strategy, implementation, workflow redesign, engineering, governance and change at the same time as it compresses parts of traditional delivery.
The sharper threat is to the assumptions underneath the leverage model:
- That substantial junior labour is required to produce high-quality analysis
- That team size remains a reasonable proxy for value
- That generic frameworks stay proprietary
- That clients need to buy the whole engagement
- That the consulting firm keeps control of the client interface
The firms may grow while the model that built them changes underneath them.
What the original model missed
"Human intervention" is not one thing.
There is the labour required to produce an answer.
Then there is the judgement and accountability required to act on it.
AI is reducing the first much faster than the second.
A client can generate a market scan, acquisition screen, operating-model proposal or transformation plan in minutes. In low-consequence work, a useful first answer may be enough.
A board recommendation, restructuring plan, regulatory remediation strategy or transaction thesis carries a different burden. Someone still has to decide whether the evidence is sound, whether the assumptions survive challenge, whether the recommendation fits the institution and whether the conclusion can be defended.
The document was only the visible part.
Behind it sat interviews, evidence, challenge, judgement, alignment and an institution prepared to put its name on the conclusion.
Generative AI makes that distinction easier to see because it makes a plausible artifact cheap. Once the artifact can be produced in minutes, the remaining value has nowhere to hide.
Where accountability matters
Accountability is not the product in every consulting purchase.
A systems migration may be a capacity and engineering problem. A benchmark may be valuable because the data are proprietary. An enterprise-platform implementation may depend on scarce technical skill.
The accountability argument is strongest where an executive must make and defend a consequential decision to a board, regulator, investor, court or workforce.
In those situations, the client may be buying:
- A documented decision process
- Independent challenge to internal assumptions
- Verification by experienced practitioners
- Evidence and traceability
- Institutional credibility
- A named person prepared to defend the reasoning
Law and audit make the separation explicit. ABA guidance says lawyers using generative AI retain their duties of competence, confidentiality, supervision and candour. PCAOB standards leave the auditor responsible for forming and expressing an opinion.
The tool can help produce the work. It does not inherit the professional duty attached to it.
The revised model
The 2022 logic was:
Analysis is becoming cheaper, so firms should distribute analysis more efficiently.
The better logic is:
Analysis is becoming cheaper, so firms must identify which surrounding capabilities remain scarce.
There are five places to look.
Proprietary inputs. Data, benchmarks, case patterns and domain knowledge that a general-purpose model cannot readily reproduce from public information.
Reusable execution. Software, agents and workflows that perform defined work at lower marginal cost.
Client integration. The engineering and operating work required to connect the capability to the client's systems, controls, incentives and roles.
Verification. The evidence, evaluation, traceability and expert review required before the output can be trusted.
Accountable judgement. A person or institution prepared to interpret, endorse and defend the conclusion when the stakes justify it.
Different services will rely on different layers. A benchmark may live mainly in proprietary inputs. A managed workflow may depend on execution and integration. A regulatory assessment may depend on verification. A board recommendation may place most of its value in judgement.
The opportunity is not to remove people everywhere. It is to remove production work where it adds little value and make human involvement deliberate where it does.
The second thing I missed
The 2022 model assumed that the consulting firm would own the platform.
IBM now explicitly helps clients build their own tailored internal AI platforms. Model providers and enterprise software companies are moving closer to deployment. The consulting firm can productise its expertise and still become one capability provider inside somebody else's environment.
That changes the strategic question.
It is no longer:
How many applications can the firm assemble?
It is:
Which capabilities remain differentiated when the client or technology provider controls the platform?
Nearly four years later, the original prediction looks directionally right and structurally incomplete.
AI became both the disruptor and the product. Consulting knowledge is becoming software. Delivery is becoming more modular.
But low-touch production, low-touch accountability and consulting-firm ownership of the interface are three different bets.
The artifact can become software. The workflow can become an agent. The platform may belong to the client.
When the decision matters, somebody may still need to put a name behind it.