Adoption Under Pressure: Challenges, Risks & Pushback on AI in Consulting
This series has covered two things so far. When Algorithms Advise set out how AI is changing the advisory value chain, from internal tooling to clients building analytical capability of their own. Beyond the Pyramid described what the firm looks like when the pyramid's two conditions no longer hold, covering pricing, trust, roles and governance.
Both articles described a destination. Neither asked why so few firms are reaching it.
The gap is wide and well documented. Access to AI is close to universal across professional services, budgets have been committed, and leadership mandates are in place. What has not followed is confidence. Firms deploy tools and stop short of changing how work is organised. Governance frameworks are published and then not enforced. Teams are given systems and no clear answer about who is accountable for what those systems produce.
The reasons are organisational, ethical and cultural rather than technical. That distinction matters, because it changes what a firm should do about them. A technical constraint is solved by better tooling. A constraint rooted in accountability, incentives or professional risk is not, and treating it as one is why so much adoption produces cost without return.
This article examines those barriers: what actually blocks adoption at scale, why firms adopt without redesigning, what happens when governance exists on paper only, and why some of what firms describe as resistance is better read as accurate risk assessment.
The Confidence Gap
Most accounts of AI adoption in professional services treat the problem as a maturity curve. Firms that have not scaled are assumed to be earlier on it, and the implied remedy is time, investment or better tooling. The barrier data does not support that reading.
What Actually Blocks Adoption at Scale
McKinsey's State of Organizations 2026, drawing on a survey of more than 10,000 senior executives across 15 countries and 16 industries, asked what prevents organisations from adopting AI at scale. The leading barrier, cited by 46% of respondents, was concerns about AI itself, covering bias, intellectual property risk and the perceived threat to jobs. Regulatory, ethical and legal concerns followed at 44%, and were more pronounced among European leaders.
Neither answer describes a capability problem. Both describe organisations that understand what the technology can do and are unresolved about what happens when it does it badly.
That distinction changes the remedy. A capability gap closes through procurement and training. A gap rooted in unassigned liability does not close through either, and firms that respond to the second with the tools appropriate to the first will keep finding that adoption stalls at the same point for reasons they have misdiagnosed.
Why Agentic Systems Widen the Gap
Delegation to a system that produces output for human review is one kind of exposure. Delegation to a system that plans and executes a sequence of connected tasks is another.
McKinsey's State of AI Trust in 2026 found that nearly two-thirds of respondents cite security and risk concerns as the leading barrier to scaling agentic AI, well ahead of regulatory uncertainty or technical limitations. The same research found that 74% identify inaccuracy and 72% cybersecurity as highly relevant risks.
The ranking is the useful part. Organisations are not saying they cannot build these systems. They are saying they are not confident they can deploy them safely at scale, which is a statement about controls and accountability rather than engineering.
For advisory work this lands harder than in most sectors. When a system contributes to a recommendation a client acts on, the question of who validated it is not procedural. It determines who carries the consequence when the recommendation turns out to be wrong, and most firms have not answered it with enough precision to satisfy the people being asked to rely on the output.
Adopting Without Redesigning
The most common adoption failure in professional services is not refusal. It is a firm that buys the tools, trains the staff, announces the programme, and changes nothing about how work moves through the organisation.
The Three-Way Split
Deloitte's research, referenced in the previous article, found roughly a third of organisations using AI to transform products, services, processes or business models, another 30% redesigning important processes, and 37% using AI with little change to how work was organised.
That last group is the one worth examining. These are not firms that failed to adopt. They adopted, and then routed the technology through an operating model built for a different kind of work. Tools sit alongside existing workflows rather than replacing them, so effort is added rather than removed, and the return does not appear.
The conclusion drawn internally is usually that the technology underdelivered.
Why Redesign Meets Resistance
The reason firms stop at deployment is that deployment is the part that asks nothing of anyone.
Redesign is different. It asks people to give up ways of working that currently define what they are worth. A senior consultant whose reputation rests on producing a particular kind of analysis is being told that analysis is now partly automated. A manager whose standing comes from running a large team is being told the team should be smaller. A specialist whose value lay in knowing where to find something is being told the search is trivial now.
None of that is irrational resistance. It is people correctly reading what a change costs them, and responding to the incentives actually in front of them rather than the ones in the transformation plan.
Firms tend to treat this as a communication problem. It is a compensation and progression problem. If promotion still rewards headcount managed and utilisation still rewards hours logged, then a redesign that reduces both is asking people to act against their own interest and hope the system catches up.
Most do not take that bet, and they are not wrong to decline it.
The Version Firms Choose Instead
What happens instead is a version of transformation that leaves the incentive structure untouched.
Tools are procured and access is granted. Usage is measured, because usage is easy to measure. A centre of excellence is established. Pilots run in functions where the risk is low and the political cost of failure is lower. None of it requires anyone to change what they are rewarded for.
This produces activity that looks like progress and economics that do not move. Cost rises through licensing, training and internal programme overhead. Efficiency gains stay trapped inside individual tasks rather than reaching the process. And because the programme was visible and expensive, its failure to deliver becomes evidence for the sceptics, which makes the next attempt harder to fund.
The barrier here is not that firms will not adopt AI. It is that adoption without redesign is the path of least organisational resistance, and it is the version that does not work.
When Governance Exists on Paper
The first article in this series opened on Deloitte Australia refunding part of a government fee after a report was found to contain fabricated references. At the time it read as a cautionary case, the kind of early incident a maturing industry absorbs and corrects.
It has not been corrected.
A Recurring Failure, Not an Isolated One
In July 2026 the Financial Times reported that four thought-leadership reports published by PwC in the Middle East between 2024 and 2026 contained fabricated footnotes, misattributed claims and sources that could not be verified. One citation linked to a media report that made no mention of the survey it was supposedly evidencing. Another carried a URL with a ChatGPT source tag still attached. PwC said it was correcting a limited number of citations and pointed to its quality control processes.
A separate KPMG report drew disputes from several named organisations over findings they said were false or misleading. Analysis of that report found that of 45 citations, five pointed accurately to real sources.
Three large firms, across roughly a year, publishing work with the same category of defect. At that point the pattern is not individual error. It is a control that is not operating.
The Distance Between Principles and Controls
What makes these cases instructive is that none of the firms involved lacked governance on paper.
All of them have published responsible AI principles. All have internal review processes, quality assurance functions and professional standards that long predate generative AI. Several have proprietary assistants built specifically so that staff would not be pasting client work into consumer tools. The infrastructure existed. It did not bind.
What separates a stated control from a binding one comes down to three practical things:
Ownership. Verification is either a defined step with a named owner, or an expectation that somebody will check. The second is not a control.
Standing. The reviewer needs the time to do the work and the authority to send it back. A reviewer who can raise a concern but not delay a deadline is a formality.
Scope. The review has to cover what the system actually produced, including citations and source claims, rather than whether the document reads well.
Published output makes these failures visible. A thought-leadership report is checkable by anyone, so the gap surfaces. Client deliverables are not public, which means the same control gap can persist in delivery work for a long time without anyone outside the engagement knowing.
That asymmetry is worth sitting with. The incidents that reach the press are the ones where verification was easy.
What Clients Take From It
The reputational cost does not stop at the firm.
A consulting report enters circulation. Clients cite it internally, journalists reference it, and other analysis builds on its figures. When the underlying citations turn out to be invented, everyone downstream has repeated something that was never true, and some of them made decisions on it.
For a client weighing whether to accept AI-supported delivery, these cases answer a question that governance frameworks cannot. The question was never whether firms have controls. It is whether the controls hold when the work is under deadline and the output looks right.
Resistance as Risk Assessment
The word most often used for what has been described so far is resistance. It carries an implication: that the objection is emotional, that the objector has misunderstood something, and that the remedy is persuasion.
Some of it is. Most of it is not.
Reading Objections as Signal
Consider what the common objections actually say when taken at face value rather than as attitudes to be managed.
"I do not know who is accountable if this is wrong." A statement of fact about the firm's governance, not reluctance. If nobody can answer it, the objection is correct.
"I cannot put this client's material into that system." Usually accurate. Confidentiality undertakings and sector rules do constrain what can be processed where, and the person raising it is often the one who signed the undertaking.
"The output looks right and I cannot verify it quickly." The precise failure mode behind every incident described earlier.
"This does not save me time once I have checked it properly." Frequently true for work where verification is the expensive part rather than production.
Each of these is a risk assessment made by someone closer to the exposure than the person running the adoption programme. Treated as resistance, they get overridden, and the firm loses its most useful early warning about where its controls are thin.
The barrier data supports reading them this way. When the leading obstacles cited by senior executives are concerns about AI itself and about regulatory and ethical exposure, the objection and the executive assessment are the same assessment, arriving from different levels of the organisation.
Answering Rather Than Overriding
The distinction that matters is between objections that dissolve when answered and objections that hold.
An objection dissolves when the firm can state who owns the system, who validates its output, what happens when it fails, and which categories of work it may not touch. Most confidentiality and accountability objections are of this kind. They are questions the firm has not answered yet, and the answer is governance rather than persuasion.
An objection holds when the honest answer is that the risk is real and unmitigated. Work where verification costs more than production. Advice where a plausible but wrong output would not be caught before it reached a client. Situations where the regulatory position is genuinely unsettled. In those cases the correct response is to narrow the use case, not to press ahead and manage the culture.
Firms that make progress here tend to do two things. They answer the dissolvable objections properly, in writing, before asking anyone to adopt anything. And they accept the ones that hold, which means publicly declining to use AI somewhere, and treating that as a governance decision rather than a failure of ambition.
The alternative is a firm where the people who understood the risk earliest learned that raising it was unwelcome.
Adoption Under Pressure
The barriers described here are not variations on a single problem.
The organisational barrier is that adoption without redesign is the path of least resistance, and it is the version that does not work. The cultural barrier is that redesign asks people to act against incentives the firm has not changed. The ethical barrier is that controls written down are not controls that bind, and the gap only becomes visible when the output is public.
What connects them is that each one is a governance question wearing a different costume. Who owns this. Who validates it. Who carries the consequence. What are people actually rewarded for. Firms that have answered those questions adopt without much drama. Firms that have not keep running programmes that produce cost, activity and very little else.
This series has argued that AI changes what consulting is worth, how firms are built, and what clients can reasonably expect. None of it happens automatically. The technology is available to everyone in the industry on similar terms. What separates outcomes is whether a firm was willing to change how it works, and honest about what it could not yet safely do.
Frequently Asked Questions
Why does AI adoption stall in professional services firms?
Adoption stalls at the point where deployment ends and redesign begins. Buying tools and granting access asks nothing of anyone, so firms complete that stage and stop. Changing how work moves through the organisation asks people to give up ways of working that currently determine their standing and their pay, which no communication programme can resolve on its own. The result is a firm with high tool usage, unchanged workflows and no measurable return, and the conclusion drawn internally is usually that the technology underdelivered.
What are the main barriers to scaling AI in an organisation?
They are not technical. McKinsey's State of Organizations 2026, based on a survey of more than 10,000 senior executives across 15 countries, found the leading barrier was concerns about AI itself, covering bias, intellectual property risk and the perceived threat to jobs, cited by 46% of respondents. Regulatory, ethical and legal concerns followed at 44% and were more pronounced among European leaders. Both describe organisations that understand what the technology does and remain unresolved about accountability when it goes wrong.
Why do employees resist AI at work?
Because most of what gets labelled resistance is a risk assessment made by someone closer to the exposure than the adoption team. Not knowing who is accountable for a wrong output is a statement about governance. Being unable to process client material in a given system is often a contractual fact, raised by the person who signed the undertaking. Separately, redesign asks people to act against incentives the firm has not changed, so if promotion still rewards headcount managed and utilisation still rewards hours logged, declining to adopt is the rational choice.
Why do AI generated reports contain fake sources?
Because verification is a control that firms write down and then do not enforce. In 2026 the Financial Times reported that four PwC Middle East thought-leadership reports contained fabricated footnotes and unverifiable sources, one carrying a URL with a ChatGPT source tag still attached. A separate KPMG report was found to have five accurate citations out of 45. All the firms involved had responsible AI principles and quality assurance functions. What was missing was a named owner for verification, a reviewer with authority to delay publication, and a review that covered citations rather than readability.
Why is agentic AI harder to scale than other AI tools?
Because autonomy changes the accountability question rather than the technical one. McKinsey's State of AI Trust in 2026 found that nearly two-thirds of respondents cite security and risk concerns as the leading barrier to scaling agentic AI, ahead of regulatory uncertainty or technical limitations, while 74% identify inaccuracy and 72% cybersecurity as highly relevant risks. A system producing output for review is one kind of exposure. A system planning and executing connected tasks is another, and most firms have not defined who validates the result with enough precision to satisfy the people asked to rely on it.
When should a company not use AI?
When the risk is real and unmitigated rather than merely unfamiliar. Three cases recur: work where verification costs more than production, advice where a plausible but wrong output would not be caught before it reached a client, and situations where the regulatory position is genuinely unsettled. The practical test is whether an objection dissolves once the firm can state who owns the system, who validates its output and what happens when it fails. If it does, the answer is governance. If it does not, the use case should be narrowed or declined, and that decision stated openly rather than managed as a cultural problem.
How Metheus Can Help
The pattern in this article is not specific to consulting. Any organisation adopting AI faces the same question: whether the tools have been added to an unchanged way of working, or whether the work itself has been redesigned around them.
We work with growth-stage and scaling businesses on market expansion, and that work has become more AI-supported over the past two years. Research moves faster. Segmentation and scenario modelling cover more ground. What has not changed is the part that carries the risk. Someone has to decide which assumption to test hardest, which market signal is noise, and whether an analysis that reads well is actually right.
That is where we spend our attention. We define what needs proving before capital is committed, verify what the analysis is resting on, and put our name to the recommendation that follows. Where we are not confident, we say so rather than presenting a plausible answer.
If you are weighing an expansion decision and want the reasoning stress-tested rather than accelerated, we are here to assist you.