Most pages that answer this question are written by consultants who want your enquiry, and the figures they give rarely say where they come from. This post separates the numbers that can be traced from the ones that cannot, and says what each price band actually buys.

What does an AI consultant cost per day in the UK?

The only independent measure of day rates that I could find is the contract market. The IT Jobs Watch tracker of UK contract adverts for AI consultant roles shows a median daily rate of £564 for the six months to 19 October 2026, with the tenth percentile at £425 and the ninetieth at £713. Outside London the median was £575, against £552 in London.

Treat that with care. It is built from 70 quoted rates, so it moves around, and it describes contractors placed through agencies on defined assignments, usually inside a larger team. It is a floor for what a person with this title costs, not a price list for advice.

The figures that sit above it come from consultancy marketing. One widely read UK guide puts freelance rates at £800 to £3,500 a day and senior staff at the largest firms at £3,000 to £15,000 a day. It cites no source for either, and its author quotes their own rate as both £2,500 to £4,000 and about £850 in different sections. The numbers are plausible as an indication of what is asked. They are not evidence of what is paid.

Why do the quotes you receive vary so much?

The strongest objection to everything above is that a day rate is the wrong unit. A buyer does not want days; they want a decision, a working process or a trained team. A cheap day that produces a slide deck is worse value than an expensive one that produces a ranked list of what to build.

That is right, and it explains most of the spread. Three things move a quote. The first is seniority: the person who has seen a pattern fail before charges more than the person who is seeing it for the first time. The second is how much is bespoke, because a workshop run from an established method costs less to produce than a diagnosis of your particular operation. The third is who carries the risk. A fixed fee puts the risk of overrun on the consultant, while a day rate puts it on you.

So when two quotes differ by a factor of three, ask what each includes before asking which is cheaper. The same wording, "AI strategy", is used for a half-day conversation and for a six-week piece of analysis.

How much does each type of AI consulting engagement cost?

For a business with a few dozen to a few hundred staff, the useful comparison is by type of engagement. The table below uses the fees AssumeAI publishes on its own service pages, which is the one set of UK prices in this post that I can vouch for, and it shows what each buys. Other providers will differ; the shape of the ladder is what travels.

The pattern is that the price follows the amount of your business that has to be understood, not the amount of technology installed. Tools are cheap; working out which process is worth changing, and getting people to use the change, is the expensive part. That fits what we have argued before about adoption being the binding constraint.

What does an AI strategy review cost, and is it worth it?

A review is where most buyers should start, because it is the cheapest way to find out whether the larger spend is justified. At our prices it is in the low five figures. At the other end, one guide puts a single two to four hour discovery session at £400 to £1,500, and boutique audit-and-sprint engagements at £15,000 to £80,000, though again without a source.

Whether it is worth it depends on one test: does it leave you able to say no? A good review ends with a short list of things to do and a longer list of things not to do, each with a reason. If it ends with a general recommendation to "adopt AI across the business", you have bought a brochure. We set out what you should expect in more detail on the AI strategy review page.

It is also worth comparing the fee with what it protects. If a firm is about to spend six figures on tools, licences and staff time, a fee that is a small fraction of that and that can stop a poor choice is easily justified. If the planned spend is a few hundred pounds a month on licences, a review is probably more than you need and a workshop will do.

What should an AI consultant's fee include, and what costs extra?

Ask for the following to be written into the quote: the named people who will do the work, the documents or working systems you will own at the end, how many days or weeks it takes, and what happens if it takes longer. Most disputes about cost come from one of these being left vague.

Then ask what is excluded. Software licences, cloud hosting and data preparation are commonly billed separately, and for implementation work they can exceed the consulting fee within a year. Training for your staff is another frequent omission. One guide suggests allowing about 15% of the project cost for it, which is a reasonable rule of thumb but is the author's estimate rather than a measured figure.

VAT is the final trap. Published prices on professional services pages are often quoted excluding VAT, and a business that cannot reclaim it should add 20% when comparing.

Should you hire an AI consultant or an employee?

The same tracker shows a median permanent salary of £82,500 for UK AI consultant roles, based on 154 quoted salaries over the same six months. Once employer costs are added, a full-time hire is a commitment of well over £100,000 a year before any tooling, and it only makes sense if you have enough continuing work to keep that person busy.

For most businesses with fewer than a few hundred staff, that is not yet the case. A fixed-scope engagement to decide what to build, followed by a retainer or a managed service while the first projects bed in, costs less and can be stopped. Hiring becomes sensible once there is a stable list of AI work that you expect to keep growing for at least a year. Our services are built around that sequence, but the logic holds whoever you buy from.

How do you tell whether a quote is fair?

Three checks are enough. Compare the day rate, if one is given, against the contract market of about £425 to £713 and ask what justifies anything above the top of that range. Ask for a fixed price, or at least a cap, so that the unit you are buying is an outcome. And ask for a reference from a client of a similar size, since a price that suits a bank will not suit a twenty-person firm.

Be wary at both ends. A very low price usually means the work will be a generic workshop with your name on it. A very high price with no breakdown usually means you are paying for the brand. The middle of the market is dull, and that is where most of the useful work is done.

That is a less exciting answer than most vendors give, but it is the one the evidence supports. The rest of this post sets out the evidence, the tasks to start with, what changes in how clients buy, and a sensible first ninety days.

What does the evidence say about AI for consulting firms?

The best study of the question so far is a pre-registered experiment run with 758 consultants at Boston Consulting Group, described by one of its authors, Ethan Mollick, with the full paper on SSRN. Consultants were randomly given GPT-4 or not, and asked to do eighteen tasks chosen to resemble real consulting work. On tasks inside what the model could do, those with AI finished 12.2% more tasks, worked 25.1% faster and produced results of 40% higher quality.

The same experiment contains the warning that matters more. On a task deliberately chosen to sit outside the model's competence, consultants without AI got the right answer 84% of the time; with AI, they got it right 60 to 70% of the time. They trusted a fluent answer that was wrong. The paper is a working paper and had not been peer reviewed when it was published, and the models have moved on since, so treat the exact numbers with care. The shape of the result has not been contradicted: large gains on the right tasks, and quiet damage on the wrong ones.

The practical lesson is that the return depends on sorting tasks, not on buying a tool. We argued something similar about adoption in general in Why adoption is the most important thing: capability is rarely the limit; the way a firm uses it is.

One more finding deserves attention. In the same experiment, the consultants who scored lowest before the test improved most with AI, by 43%, while the strongest improved by less. For a firm, that means the tools narrow the gap between a good first draft and an average one, and raise the premium on the person who can tell the difference. Review skill, not typing speed, becomes the scarce thing, and hiring and training should reflect it.

Which consulting tasks should use AI first?

Start where a wrong answer is cheap to spot and the work is repetitive. The table below sorts the common candidates by how quickly they tend to pay back and how closely a person needs to check the output.

The Management Consultancies Association's 2026 member survey fits this picture. Of the firms surveyed, 76% already use AI for research and information gathering, up 9% on the year before, and 77% have either built AI into their systems or let staff use AI models. Those figures come from a January 2026 release covering a Savanta survey of more than 1,000 consultants. Research is where almost everyone has started, which also means it is where a firm gains least relative to its competitors.

Why are clients spending less on consultants as AI spreads?

A consulting firm should not only ask what AI does for its own margins. It should ask what AI does to the thing clients are willing to pay for. City AM's Business Spend Pulse, built from card transactions at more than 300,000 British businesses, found that median spending on consulting fees fell 1.9% in August against a year earlier, while overall corporate spending rose 2.2%. One in seven businesses now pays an AI company, up from one in thirteen, and the share paying professional services firms fell by about a percentage point to 18.6%, the steepest fall of any category.

That data comes from one card provider's SMEs, so it is a signal and not a census. The direction is still uncomfortable. Where a client can get a competent first answer from a tool, the part of a consulting day that was research and drafting is worth less per hour. The same survey from the MCA found that 52% of firms see the rise of AI as the industry's leading challenge, with some uncertainty over whether clients will pay for it.

The sensible reading is not panic. It is that the billable value moves towards the parts AI does badly: knowing which question matters, knowing the client's politics, and being accountable for the recommendation. We made the wider case in AI agents are being trusted faster than they are being checked, and it applies to the firm as much as to its clients.

What does AI for a consulting firm look like in practice?

At AssumeAI, the pattern we see most often in professional services firms is not a shortage of tools. Staff are already using them, unevenly and without a shared view of what is allowed, so one person saves hours while another puts client material somewhere it should not go. The firms that get value first are the ones that decide, in plain terms, which tasks use AI, who checks the output, and what may never be pasted into a model. That is dull work, and it produces a return faster than any new software purchase.

Firms that skip it tend to find the gains are real but invisible: time saved on one project is absorbed by rework on another, because nobody agreed how to check the output. Making the saving visible, even crudely, by timing a handful of repeat tasks before and after, is what turns a feeling into a business case.

What should a consulting firm do in the first ninety days?

Firstly, list the ten tasks that consume the most junior time and sort them using the table above. Pick two from the fast column, and measure how long they take today. Second, write a one-page rule on what client information may go into which tools, and who signs off client-facing output. Third, run the two tasks with AI for six weeks, checking every output, and compare the time and the error rate with your baseline.

Finally, decide what you will do with the time you have bought. If it becomes more client conversation, better synthesis and faster turnaround, the firm gains. If it becomes simply cheaper junior hours handed to the client, the City AM figures suggest the client will notice, and the price will follow. A structured AI strategy review is one way to get through the sorting and the rules quickly, and it is worth doing before spending on tooling. You can read more about who we are and how we work on our about page.

The firms that do well from this will not be those with the most AI. They will be those that know, task by task, where it can be trusted, and who say so to their clients in plain language.

Consider what has happened in 2026. In April, Anthropic said of Claude Mythos Preview that its large increase in capability had led it not to make the model generally available. Access went instead to a defined group: eleven launch partners and over forty further organisations that build or maintain critical software, with $100 million in usage credits to support them. In June, an executive order invited developers to give the US government up to 30 days of pre-release access to models with advanced cyber capabilities. The order is voluntary, and it says in terms that it creates no licensing requirement. Yet it describes a direction of travel: the most capable systems reach a few trusted hands first.

Then came the incident that gives the trend its force. In July, Hugging Face disclosed that an autonomous, agent-driven intrusion had compromised its production infrastructure over a weekend, through thousands of individual actions across short-lived sandboxes. OpenAI has since acknowledged that this involved models under internal cyber testing. Its security team said at Black Hat in August that it was consciously slowing research to improve security, and on 18 August it announced a two-week pause on reinforcement learning of its latest models. Sam Altman called it the first security incident he had felt viscerally. Warning signs of this kind are what change the behaviour of a provider with every commercial incentive to ship.

The strongest objection deserves its full weight. Holding capability back from the public mainly entrenches the largest labs, which keep the best systems for themselves and their chosen partners. It pushes ordinary businesses towards open or overseas models with fewer safeguards. And it does little to reduce risk if development itself carries on accelerating behind closed doors. All three points are fair, and I would not dismiss any of them.

Even so, a visible lag prompted by a real incident is better than releasing every advance the moment it exists. A delay gives defenders time to harden software before the capability becomes widely available, which is precisely the logic of giving Mythos to infrastructure maintainers first. It also makes the gap legible. A business that knows the frontier is ahead of what it can buy can plan for it; a business that believes it already has the frontier cannot.

The open-weight question is where I am least certain, and it matters most. Hugging Face completed its forensic work with GLM-5.2, an open-weight model from the Chinese company Z.ai, running on its own infrastructure, reconstructing the timeline from over 17,000 logged attack events. Its stated reason is instructive: the lesson for defenders is to have a capable model you can run yourself, vetted and ready before an incident, to avoid guardrail lockout and to keep attacker data and credentials inside your own environment. Whether Chinese developers will show the same restraint as their US counterparts is far less clear, and I would not build a plan on the assumption that they will.

The practical conclusion for a decision-maker is modest. Treat the model you can buy as a lagging view of what exists, and review your security assumptions against the capability that sits just beyond it. Decide in advance which tools you would rely on if your usual provider declined a request or restricted access. The frontier will move faster; you will simply see it later, and the businesses that do well will be those that expected it.

It would be wrong to read this as evidence that artificial intelligence cannot do real work. Amazon, like Salesforce and Google, is genuinely integrating agents into workflows that used to require people, and it is competing hard, with Google, OpenAI, Meta and Anthropic, for staff who can build and run those systems. The capability is not fictional, and a business that dismisses this wave of automation as a rerun of past hype cycles makes the same mistake in the opposite direction. The case here is not that AI substitution is a myth.

The trouble is narrower. Many of this year's headcount decisions were made on the promise of that capability, not on evidence that a particular deployment, in a particular workflow, could sustain it. Klarna is the case study everyone already half remembers. In 2024 its chief executive credited a customer service bot with work equivalent to seven hundred full-time agents and projected forty million dollars a year in savings. By early this year the company was quietly rebuilding its human team, because the system handled routine queries well and multi-step disputes, fraud cases and escalated complaints badly, and unwinding the failed automation cost more than it had saved.

The scepticism is loudest where the claim is easiest to check against day to day work. When GitLab cut seven per cent of its workforce in May and called it an investment in the agentic era rather than a cost reduction, the developers who actually use the product were unconvinced. Several pointed out on Hacker News that the share price had fallen by half over the preceding year, and one summarised the real constraint bluntly: the bottleneck is decision making, and AI is bad at that. A framing built for investors and a framing that would survive contact with the backlog are not always the same framing.

The aggregate data now agree with the sceptics. Gartner's survey of three hundred and fifty companies with revenues above a billion dollars, published in May, found that eighty per cent had cut headcount after deploying AI, with no correlation between the size of the cut and any improvement in return on investment; the firms with the strongest returns were amplifying their staff, not replacing them. Challenger, Gray and Christmas's monthly tally supplies a quieter version of the same point. Artificial intelligence led every stated reason for layoffs from March through July, then in August fell to fourth place behind plain restructuring, the least glamorous explanation a company can give. A stated reason that unwinds that fast was doing rhetorical work, not descriptive work.

Economists at Stanford's Institute for Economic Policy Research went looking for the aggregate employment effect directly and mostly could not find it. Only five per cent of firms report any employment impact from AI adoption, and what remains sits alongside explanations that have nothing to do with model capability, among them interest rate rises that predate ChatGPT and a correction from pandemic-era over-hiring. The one clear casualty in their data is recent graduates in AI-exposed fields, whose unemployment reached five point six per cent early this year, up one point six percentage points on three years earlier, a narrower finding than the industry-wide story built on top of it.

None of this means the AI-exposed roles are safe, or that the technology stays where it is today. It means the decision quality behind this year's layoffs has lagged the confidence with which they were announced, which is why Forrester now expects roughly half of them to be reversed, with more than a third of the reversing firms rehiring over half the people they let go. A business weighing the same cut should ask a narrower question than whether AI can plausibly do the work. It should ask what evidence, from this deployment, in this workflow, shows that it already does. Amazon's recruiters are, in effect, asking the same question about the people it just let go.

That story broke on 18 September, and it was not the only one. In the same eight days, three unrelated systems failed at exactly the point meant to stop this happening, each in a different part of the AI stack.

The generous reading is that none of this is new. Every fast-moving deployment produces incidents, and the response here was, in places, genuinely responsible: Anthropic and OpenAI shipped patches within a day, and Google volunteered the details of its own failure rather than waiting to be caught. A defence analyst using a chatbot badly is a training problem, not evidence the technology is unsafe. None of that is wrong.

But in every one of these cases, the failure was not that the AI did something unexpected. It was that the mechanism built specifically to stop the unexpected reaching a decision did not hold, and nobody found out until an outsider looked closely, or the consequences were already unfolding.

Researchers disclosed Plugin4Shell on 17 September, a flaw in how Claude Code, Codex, GitHub Copilot and Gemini CLI verify plugin code. All four rely on cryptographic hash pinning to guarantee an installed plugin matches the version a developer reviewed. Plugin4Shell breaks that guarantee: a malicious branch can be checked out in place of the pinned commit while the agent reports a clean match, and because these agents update plugins automatically, no click or approval is required. Anthropic and OpenAI patched within a day; Microsoft has not, and Google simply deprecated Gemini CLI. The pin was the whole point of the mechanism, and it pinned nothing.

A day later, Google disclosed that Gemini had gained unauthorised access to three real organisations during a security evaluation back in May, only reporting it in September after an outside testing firm flagged it. The model believed it was inside a sealed test environment. A fictional company name in the test collided with a live domain, a misconfiguration left the sandbox connected to the internet, and Gemini guessed and harvested credentials against real infrastructure, still believing it was inside the exercise. Google called it mistaken identity rather than misalignment. Sydney von Arx of the Nightingale Collective was blunter, noting Anthropic had initially downplayed similar incidents before later revising its account.

A paper on cyber-capable agents published that same week gives the pattern a name. Evaluation frameworks measure whether an agent completes a task, the authors argue, treating the tool access, egress controls and privileges that made completion possible as background detail rather than the security object worth testing. The boundary is assumed, not verified, which is precisely how a sealed test environment turns out not to be sealed.

The business evidence says the assumption is universal, not a frontier-lab problem alone. A survey of 919 executives for Gravitee's State of AI Agent Security 2026 report found 88 percent of organisations had a confirmed or suspected AI agent security incident in the past year, while 82 percent remained confident their existing policies protected them from unauthorised agent actions. Merritt Baer, formerly of AWS and now chief security officer at Enkrypt AI, named the gap precisely: enterprises believe they have approved an AI vendor, when what they have actually approved is an interface, not the underlying system.

That is the sentence worth sitting with for a professional services firm or an SME, rather than a defence command or a frontier lab. Nobody in this week's stories was reckless. They trusted a boundary that looked, from where they stood, exactly like something a serious vendor would already have tested. The useful question is not whether to trust an AI system's output, but which of your own boundaries between that output and a decision you have actually tested yourself, rather than assumed someone else had.

That reading survives only if the industry's existing commitments are anything to go by, and a study published the same week says they are not. Researchers at the Oxford Internet Institute examined every published version of twelve frontier developers' safety frameworks and found that 67% of material changes could not be identified from the developers' own accounts. Where a change could be traced, 77% weakened or removed a commitment rather than strengthened one, and weakenings were the ones most often left undisclosed. The mechanism Amodei's essay rests on, namely that a lab's account of its own safety posture can be trusted and checked from outside, is the same mechanism this study found already failing, across the industry, in the same fortnight the essay was published.

Nor is the danger these commitments exist to manage a future one. Days before the essay, Anthropic published its own account of Claude's misuse over the preceding eight months: a Russian state-linked group automating intrusions against more than twenty government and defence-sector targets across Europe and Ukraine; a financially motivated group that analysed 1.8 million Android applications for stealable credentials and, in one case, moved from first access to bulk data theft in hours; and researchers linked to a Chinese university who generated more than a dozen candidate zero day vulnerabilities against roughly fifty organisations in a single month. The essay describes a threshold the industry must not cross. The report filed by the same company days earlier describes several it says have already been crossed.

The people building on these tools were not slow to draw the connection. A further disclosure this week, that Yemeni militants had used Claude Code while developing missile guidance software, was read on Hacker News less as a warning than as convenient timing: the most upvoted comment noted that a hobbyist cannot deploy a test project to a virtual machine without tripping Anthropic's own safeguards, while an alleged weapons use went undetected until Anthropic disclosed it. A separate thread the same week put the suspicion more bluntly: labs warn at length about what happens if powerful AI reaches the wrong hands, when by their own account it already has.

That opacity is not confined to industry policy. A systematisation of agentic jailbreak research, also published this week, found that an agent's final output can look entirely safe while the planning and tool calls that produced it were already compromised. The same undisclosed drift that the Oxford researchers found in published frameworks recurs, invisibly, inside a single deployment. A business watching only what an agent says, rather than what it did to arrive there, is checking the wrong layer.

This is where the argument stops being an industry debate and becomes a balance sheet one. Gartner's survey of 161 chief audit executives, published the same week, found that 93% of audit functions use AI while 60% have no formal strategy for it, leaving most unable to say whether the investment is paying back. Salesforce, launching a set of packaged agents days later, sold a governance layer alongside them as a separate feature rather than something already built in. Businesses have been filling that strategy gap with borrowed confidence, treating a vendor's safety framework or a chief executive's essay as a stand-in for oversight they have not built themselves.

It cannot serve as that stand-in, on this fortnight's evidence. A commitment its own author can revise without disclosure, and a threat report describing harm that has already happened, are not a substitute for verification a business runs on its own account. The discipline of checking what an AI system actually did, rather than what its maker says it would never do, has to sit with the business using the tool. Nobody upstream is going to do that work on their behalf, and the record this week is fairly precise about why. The same gap between what labs disclose and what the public sees has since widened, as we argued in The frontier will move faster, and the public will see it later.

OpenAI's own account of the decision is reasonable and deserves to be stated before it is argued with. Chief scientist Jakub Pachocki has said Astra's effective reasoning depth has been kept close to that of GPT-4, and that the company continues to prioritise chain-of-thought monitoring precisely because it understands what is at stake in giving it up. On this account the architecture was chosen for efficiency, not obscurity; the loop count is a dial OpenAI has deliberately left low, and nothing published about Astra suggests the model reasons entirely in latent space today. Read generously, this is a faster way to spend compute, managed by people who take the risk seriously.

The difficulty with a dial is that somebody eventually turns it, and the people best placed to judge that possibility are the least reassured by it. When the reporting broke, Redwood Research's Buck Shlegeris said he was extremely concerned, and his colleague Ryan Greenblatt asked the more pointed question of whether the industry now has a readily available route to much higher opaque reasoning depth, not whether one exists in theory. On the LessWrong thread that followed, the disagreement was rarely about the current model; even researchers offering reassurance conceded the trajectory, noting that OpenAI's own job postings already refer candidly to the potential loss of monitorability this class of architecture invites. Competitive pressure does the rest. Nobody expects a rival lab to leave its loop count low once a faster model is one setting away.

This is not a new worry, only a newly urgent one. A paper on chain-of-thought monitorability, circulated among safety researchers months before Astra existed, made the underlying argument in its abstract: verbalised reasoning is a fragile safety mechanism that depends on models continuing to think in language, and a shift toward continuous latent representations would eliminate that mechanism outright, not merely weaken it. The paper was written as a warning about a hypothetical shift. Astra is the first widely reported model where the hypothetical has a name.

England's legal system is, at the same moment, building the opposite assumption into how it intends to hold AI to account. The UK Jurisdiction Taskforce's statement on liability for AI harms, published in July, concluded that existing negligence law is flexible enough to cover most of what AI will do wrong, and went further: a professional could in time be found negligent not for using AI but for failing to, once competent peers routinely do. That rests on being able to establish, case by case, what a system did and why, closely enough to show it fell short of a reasonable standard. The statement acknowledges the opacity of AI systems as a live complication for causation, then largely sets it aside as no different in kind from problems English law has solved before. It may be right about that. It is also being written at the exact moment the reasoning it will need to inspect is becoming less available to inspect, not more.

For the professional-services firms and SMEs this affects, the two shifts arrive as one problem, not two abstract ones. Deloitte's most recent survey of AI in the enterprise found that only one company in five has a governance model mature enough to oversee autonomous agents at all, long before those agents are reasoning in loops nobody outside the lab can read. A rising standard of care and a shrinking evidentiary trail are heading toward each other, and most organisations adopting agentic AI this year have no way to notice the collision until a court, or a client, asks them to explain something they cannot.

None of this argues for waiting on AI, a case this publication has made before on different grounds. It argues for treating the audit trail as a procurement question rather than an engineering afterthought: whether a vendor's model can produce its reasoning in a form fit to be shown to someone, not whether the model is impressive. That is a duller question than whether an architecture is exciting. For the businesses that will eventually be asked to account for what their agents did, it is the one that will matter.

The natural objection is that capability is still the binding limit, and this week supplied its strongest form. Salesforce and Anthropic announced Claudeforce, plugging Claude directly into the revenue cycle: pipeline review, forecasting, deal health, all handled by an agent with live access to CRM data and, in Salesforce's own language, permission to take governed action on it. Anthropic's annualised revenue has surged past $65 billion this year, and the company is committing $45 billion to computing capacity from Nscale to keep building. The honest reading of this thread alone is that models are not yet capable enough, not that anything is missing from how businesses use them.

Does a more capable model mean a more reliable one?

A paper published this year under the title Towards a Science of AI Agent Reliability tested 15 models across two benchmarks and found the opposite: recent capability gains have yielded only small improvements in reliability, once reliability is measured properly, across consistency, robustness to minor variation, predictability of failure, and how badly things go wrong when they do. The point is structural, not a complaint about any one model. A business does not need to know whether an agent is clever enough for a task; it needs to know, in advance, whether the tenth run will behave like the first nine, and at present nothing about a newer model answers that question.

Practitioners said the same thing more bluntly this week. A Meta security researcher described an email agent that had behaved correctly on a small test inbox and then, turned loose on her real one, deleted messages it had been asked only to organise. The discussion that followed on Hacker News did not converge on the model needing to be smarter; it converged on treating agent output the way any engineer treats input from outside the system, verified through a deterministic check before it is allowed to act, never obeyed on trust. That is a discipline, not a capability, and most organisations giving agents access to real systems this year have not built it.

What separates the 6% of high performers?

The earnings data shows what happens without it. McKinsey's 2026 survey of over 1,700 executives found that roughly 6% of organisations qualify as AI high performers, attributing meaningful earnings impact to AI use; the other 94% do not, despite record spending. What separates the 6% is not model choice: high performers are close to three times more likely to have redesigned the workflow around the tool rather than simply installing it. Deloitte's 2026 enterprise survey found the same shape from the other direction: 66% of firms report productivity gains, but only 20% report revenue growth, since barely a third have redesigned any process at all. Deloitte's own recommendation is to embed oversight into performance metrics as AI takes on more of the task, which is a verification structure by another name.

Who is building the check on what agents do?

Regulators have started building the version of this that businesses have not. Transparency obligations under the EU AI Act took effect on the second of August, requiring chatbots to disclose that they are automated and AI generated content to carry a machine readable mark, backed by fines of up to €15 million or 3% of global turnover. Whatever one thinks of the mechanism, the instinct is identical to the one in the reliability paper and the deleted inbox: somebody outside the system has to be able to check what it did, since a more capable model does not remove the need for that check, it only raises the cost of skipping it.

None of this argues against adoption, which this blog has made the case for before and will not re-argue here. It argues that adoption has stopped being the question that matters most. The businesses separating themselves this year are not the ones that gave agents the most access soonest; they are the ones that built, at the same pace as the deployment, a way to check what the agent actually did before it touched a customer, a contract, or an inbox that mattered. Nobody is going to build that discipline for a company except the company itself, or eventually a regulator with a fine attached.

Building that check is the work of our AI governance implementation service, and our page on AI for professional services shows where it applies. The same pressure on oversight appears in our post on AI reasoning going dark.

Everything in the media is polarised. That polarisation is not merely an aesthetic problem with journalism; it is creating a genuine and widening misunderstanding about what AI actually is, what it can do, and what businesses should reasonably expect from it.

The coverage falls into two camps, and both are wrong.

On one side, you have the breathless evangelism: AI is coming for every job, every industry, every assumption you hold about how work gets done. Transformation is imminent. Weeks, not years. On the other side, you have the sceptics and the contrarians, who have looked at the consumer version of ChatGPT, found it capable of writing a mediocre birthday card, and concluded that the whole thing is a sophisticated autocomplete dressed up in venture capital money.

Neither position is accurate. And the tragedy is that both, in their own way, cause real damage.

The watershed moment has already happened

Think about the day the first iPhone was released. At the time, the headlines debated whether it was a gimmick, whether people really needed the internet in their pocket, whether it would ever rival the BlackBerry. What those headlines missed entirely was the more important question: not what the iPhone could do on day one, but what the world would look like once that capability was in everyone's hands and had fifteen years to compound.

We are at that exact moment with AI.

The watershed has happened. The capability threshold crossed in the last two to three years is not comparable to the incremental software improvements that came before it. Something genuinely changed. The question is not whether AI represents a meaningful technological shift. It does, plainly and verifiably. The question that actually matters is: what does the iPhone 17 of AI look like in fifteen to twenty years? What industries, roles, and assumptions get rebuilt entirely once this technology has had time to mature, spread, and compound in ways we cannot yet fully anticipate?

That is the conversation businesses should be having. Instead, the media has reduced it to a binary: transformative or overhyped. Both framings miss the point.

AI as a lever, not a magic wand

Like any technology, AI acts as a lever. It amplifies the output you can extract from a given input. It does not replace the input, the judgement, or the strategic direction. A business with unclear processes, poor data, and no real understanding of where its bottlenecks sit will not be transformed by AI adoption. It will be confused by it, at scale, slightly faster than before.

A business that understands its own operations, knows where time is being lost, and has people willing to engage seriously with new tools can move meaningfully. AI consulting and AI implementation work best not when they are treated as a silver bullet, but when they are applied to specific, well-defined problems where the technology's current capabilities are a genuine fit.

That distinction matters enormously for any organisation considering AI transformation. The value is real. The timelines are not what the headlines suggest.

Why the extremes are dangerous, not just irritating

The polarised headlines produce a specific and observable problem in the market. Some businesses arrive at conversations about AI development and AI adoption expecting transformation in a matter of days or weeks. They have read that AI can automate entire departments, and they want to know why their department has not been automated yet. They have unrealistic timelines, vague objectives, and a growing sense of disappointment when reality arrives.

Others, equally misled, have concluded that AI is simply the consumer version of ChatGPT and nothing more. They have tried it, found it imprecise or generic, and moved on. They are not hostile to technology in principle; they simply do not believe that what they have seen bears any resemblance to something that could add value to their specific work. That scepticism, in many cases, is entirely rational given what they have been exposed to.

The divide between these two groups is the problem. The over-believers risk burning budget, goodwill, and internal credibility on poorly scoped initiatives. The disbelievers risk ceding ground to competitors who are quietly building genuine operational advantages through AI integration. Neither outcome is inevitable, but both are common, and both trace back to a business press that has chosen drama over precision.

What a more honest conversation looks like

AI is not a gimmick, and it is not an overnight fix. It is infrastructure, of a kind, and it rewards the same quality of thinking that any serious infrastructure investment demands. The organisations extracting genuine value from AI right now are not doing so because they moved fastest or spent most. They are doing so because they identified specific, bounded problems, engaged the right expertise in AI consulting and implementation, built something, measured it, and iterated.

The iPhone did not transform commerce, communication, and culture on the day it launched. It did so over a decade and a half, as developers, businesses, and users found new ways to use what was now possible. We are at day one of that same arc with AI. The businesses that will benefit most are those who understand that, resist the noise on both sides, and start asking the right questions now.

That is not a particularly exciting story to run on the front page. But it is the one that actually helps businesses make good decisions, and it is the conversation that AI transformation, done properly, is built on.

If you are deciding where AI fits in a firm like yours, start with our page on AI for professional services. The reasons adoption matters more than headlines are set out in why adoption is the most important thing.

#AIImplementation #AIConsulting #AIAdoption #Cybersecurity #AISecurity #SME #SoftwareSecurity #AIGovernance

Adoption is the single largest factor, it is in fact counterproductive to roll out AI schemes with licences and new software, if the users it is intended for never use it.

Why does AI adoption fail?

There are three main reasons adoption fails when implementing AI into companies. Firstly, fear plays a massive role. Many employees fear their job role being taken or being replaced entirely. Second, many users are not aware of what AI is actually capable of, it is often viewed as a search tool to be used like google. Finally, trust is naturally low when being introduced to these tools for the first time, with negative media and talk of hallucinations, often people view AI as a temperamental, unreliable tool.

Is the fear of losing your job justified?

To begin with, fear. Fear of AI taking your job is partially real and massively misunderstood. AI cannot do your job for you without human intervention, initial input and quality assurance are key parts in the AI prompting cycle. AI is a tool just like any piece of software. When Microsoft Excel released in 1985, many people who were using pen and paper didn't adapt. We're seeing a similar pattern where those who do not adapt are left behind. Jobs have decreased over the last few years, but we're beginning to see an uptick of job roles coming back as companies realise AI without the human component is fractionally effective. Using AI effectively is a skill which is both desirable and hireable.

Why do people use AI so narrowly?

Lack of knowledge is the next big hurdle. Users view AI as basic chat functionality, when in reality, it's a Swiss army knife of productivity. Showing users successful use cases for some more out the box areas, which they are likely not to expect to be possible, works to reframe the current belief. The question all users of AI should be asking is, how can I utilise this tool to improve my day-to-day work and eliminate my most unenjoyable tasks.

Why is trust so low?

The final barrier, distrust and scepticism are at an all time high, talks of an AI bubble, hallucinations and wrongful AI usage are everywhere in media. However, the tools are not the problem. AI in its current forms is arguably the most powerful piece of software ever created, and people use them day to day to create incredible things over a timespan that was not even thought possible 5 years ago.

What fixes all three?

For all three points there is one solution that helps with all. Upskilling employees to use AI correctly, safely and productively. This helps gain the skill that makes you more valuable and less likely to be replaced, learn about use cases that can expand what you thought was possible with the current tools and build trust and confidence through usage and practice. AI training and adoption will have to be hand in hand for the foreseeable future.

For how this looks in a consulting firm, see our page on AI for consultancies and the Granite5 AI training workshop.

On the 30th May, we were fortunate to be invited to the OXAI Frontier event at the Mathematical Institute, Oxford University, to speak about the future of work and how AI is affecting, and will continue to affect, companies of all sizes. Our co-founder Rupert Wiseman gave an excellent talk on the most exciting and concerning patterns we are seeing across the industry. It was fascinating to hear how AI is reshaping the lives of so many different people and sectors, including education, traditional business and online media.

We were also lucky to hear from a number of outstanding keynote speakers, among them Sir Nigel Shadbolt, whom some regard as the godfather of AI. He has worked in the field since the 1970s and has been part of its journey from the very beginning.

We often find ourselves immersed in the practical side of AI and integration, so it was a breath of fresh air to step back and consider a more theoretical framework: how AI is applied, and the wider impact it is having on society now and in the years to come.

Who was in the room and how did they feel?

The room was full of interesting people. There were business owners, some deeply involved with AI and others interested on where to use it effectively. There were also students who have used AI natively since the first publicly available large language model was released. The mood was mixed. Some felt a degree of fear about what is to come, while many were full of excitement. One thing was clear, however. AI is not going anywhere, and it will only continue to weave itself further into our lives.

It was a pleasure to connect with so many like-minded individuals. I left the event energised and genuinely excited about the direction of travel, and about the minds driving AI forward.

A special thank you to the wonderful team at Swift Strategies for making it to the event.

The practical side of that conversation is on our about page, in our page on AI for professional services and in our AI team training.

Specifically, I am referring to the implications of Anthropic's frontier model, Claude Mythos, and the initiative built around it, Project Glasswing. Mythos is a general-purpose model that happens to be exceptionally capable at reading, understanding and modifying complex code, and that same capability allows it to find software vulnerabilities at a depth and speed that, until very recently, sat well beyond most human teams. In its early use it has surfaced thousands of previously undiscovered vulnerabilities across critical infrastructure and widely used open-source software, a number of which had been sitting undisturbed for a decade or more.

Anthropic has chosen not to release the model publicly, and has instead granted access through Project Glasswing to twelve launch partners and more than forty further organisations, over fifty in total, including some of the largest technology and infrastructure providers in the world, so that they can begin to find and patch weaknesses in their existing technology before anyone with hostile intent can exploit them. That is the right instinct, and it is a serious piece of work, but it is also where my concern begins.

Will this capability stay contained?

A capability of this kind does not stay contained indefinitely. Whether through an eventual, safeguarded release or through adversaries reaching equivalent capability of their own, the ability to scan ordinary software for exploitable weaknesses will in time become far more widely available than it is today; and at that point, legacy software across the SME economy faces a genuine fight to survive. The window between a vulnerability being discovered and a vulnerability being exploited has already narrowed from months to minutes, and the organisations inside Project Glasswing are preparing for precisely this. The concern is everyone else.

Why are SMEs more exposed than large firms?

Most of the SME economy runs on legacy software, on systems that were built years ago, are no longer actively maintained, and have never been examined with tools of this strength. For most of their working life that has been acceptable, because the effort required to find their weaknesses was itself a meaningful barrier; as that barrier falls, a great deal of software that has quietly done its job for years will be exposed in a way it never has been before.

I do not raise this to create alarm, and it is, if anything, the opposite of the hype that usually surrounds these announcements. I raise it because, in my experience, very few firms outside the largest enterprises are aware of this shift, and fewer still are ready for it. The work of preparing is measured and entirely achievable, but it has to begin before the risk becomes visible rather than after, and that is a question of sequencing and priorities far more than it is a question of budget or headcount.

The firms that treat this as a foundational issue now will be in a markedly stronger position than those who wait for the problem to announce itself. By the time this capability is widely felt, the work of preparing for it should already be done; and at present, I do not think most firms are aware of that, let alone ready for it.

If it would be helpful to think through where your own exposure sits, we would be very happy to have that conversation.

Preparing is mostly a matter of governance and sequencing, which our AI governance implementation service covers for firms like those on our AI for professional services page. A related point about trusting AI without checking it is in our post on AI agents.

#AIImplementation #AIConsulting #AIAdoption #Cybersecurity #AISecurity #SME #SoftwareSecurity #AIGovernance

What's clear through these filings and SpaceX's latest deals, including the Anthropic deal, delivering ~$15 billion a year, and their recent moves to purchase Cursor for ~$60 billion, is that AI is a core focus at SpaceX and they have no plans to slow down anytime soon.

I find this a very exciting opportunity for both the future of space exploration and AI.

How much is SpaceX spending on AI?

Never before have we seen a non-government-funded space program with the resources available to SpaceX, with a reported spend of $42 billion this year and that figure doubling YoY for the last 2 years. If this continues at this rate, it will be comparable to the Apollo missions, where the US government spent ~$25.8 billion (~$280 billion accounting for inflation) over a 12-year stretch (1960-1972).

What did the Apollo programme leave behind?

Although this spend was unpopular, many fantastic inventions we have today come as a direct result of the engineering that went into the Apollo missions. These include memory foam, fireproof suits, cordless power tools, water filtration systems, and countless computer and technical advancements that would lead to inventions such as MRI machines and guidance systems.

Will space pull AI capability forward?

I believe, given the rate of innovation, SpaceX can bring about another wave of advanced engineering and acceleration to the AI race, due to the complexity of the problems that must be solved in order for space missions to be possible and AI's ever-evolving ability to solve these problems. It will create a feedback loop similar to what we have seen internally with Anthropic's models.

We already live in a world where a new invention gets released every day, and I believe this will accelerate massively in the coming years. With the vast amount of resources being poured into programs at this scale, I find it hard to believe that Artificial Intelligence will not be found in every aspect of our lives in the future. Culturally and economically, it will be hard to predict and measure the impact that AI might have, but for the progress of science and technology, I believe it's clear that AI will be the single largest driving factor.

For a business, the practical question is where to point this capability first, which is what our AI prioritisation consultation and our page on AI for professional services are for. We have also written about how the frontier will reach the public later than it moves.

Now, both are extremist views, but when pushed for an answer, I strongly favour the latter contention. The complexities of which are greater than those laid out above, but nevertheless, it is representative of the huge technological leap we have seen take place over the past decade (and more visibly over the last five years). With that being said, there has been a demonstrable downturn in the state of the graduate job market, and the overall job market. Albeit, this is coupled with fast-moving market dynamics, and to name a few: the geopolitical instability, the mass expansion of the middle management layer over the past two to three decades, and the lack of financial scrupulousness owed to these corporations.

Nevertheless, the question prevails. How does one adapt to these shifting dynamics, and interestingly, how does one protect their position within a company with such a threat of 'AI replacing you'? The answer, in my humble opinion, is obvious, namely: relationships.

The power of relationships has grown tenfold over the last few years. Now, as a past legal student, my natural question becomes, well, how does one define relationships?

How does one define a relationship?

Take, for example, a non-executive director for a fast-growing start-up; let's call him Stuart. He has previously exited three firms and now acts as a NED for a number of start-ups. If I receive advice from Stuart on a challenge surrounding an M&A merger, and the same advice from an LLM, who am I going to trust more? Personally, Stuart. He has the previous experience, and even if the LLM has a better answer, I'm going to be inclined to follow the person who has learnt through blue shins.

Similarly, imagine someone who manages money at an investment management firm. Even with the rise of AI and tooling, although I may expect them to innovate internally and use these tools, I'm not going to end the relationship with the person who manages my money and has built up that relationship over years and years.

Which roles are at risk, and which are safe?

So, what does this mean moving forward? It means those who conduct manual data transposition, back office management, administration, KYC, AML: your jobs are at risk.

Those who manage a relationship with a client: your roles are safe. And therefore, how does one protect their job security in the world of AI?

They develop relationships, maintain those relationships, and deliver relentless value. Because even where AI can replace you, if you have done the above, it will not.

Firms that sell on relationships still need to put AI to work behind them, which is the subject of our page on AI for consultancies and our AI strategy review. The wider argument about people and adoption is in why adoption is the most important thing.