Clients are already using Claude to challenge agency decisions. The real question isn't whether AI can do the work, it's what an agency should charge for once part of the job takes minutes instead of days.
A client sends over a chunk of code. Claude wrote it, and the brief attached is simple: this is how it could be done. There's just one problem. The client has no idea how our theme is actually set up, or how the CSS is structured underneath. We see this occasionally rather than constantly, but often enough that it's become a familiar shape of conversation rather than a one-off surprise. Fixing it doesn't involve touching a single line of the code. It means walking the client through why their version won't work as shown. Once that's explained, clients generally understand, and the exchange usually still surfaces a clear enough sense of what they actually wanted in the first place.
This isn't an isolated habit on one account. Across both development and performance marketing clients, the same pattern keeps showing up: a client uses Claude to question an estimate, an approach or a cost, sometimes proposing an AI-generated alternative that conflicts with how we actually work. Clients aren't being difficult here. They're informed differently now, and that's a genuine shift in the balance of information between agency and client, not a discipline problem to be managed away.
We've watched this play out as a repeating cycle rather than a single incident. A client arrives with a Claude-generated idea, missing the context that would have shaped it properly, and we end up doing the work twice: once to understand what was actually being asked for, and again to build it correctly. It's a slower route to the same destination, and it's frustrating on both sides of the table. Multiply that cycle across several accounts running at once, and what looks on any single call like a minor detour starts to show up as a genuine drag on delivery capacity.
The instinct in that moment is usually to correct the client, and correcting the technical detail is necessary. But the more useful move, and the one that actually closes the loop, is treating the Claude-generated proposal as a rough translation of something real the client wants, rather than a document to be argued with line by line. Underneath most of these exchanges is a legitimate question the client couldn't otherwise articulate as clearly. Claude gave them a vocabulary for it, even if the specific implementation it produced doesn't survive contact with our actual codebase.
There's an upside buried in the friction, too. Being questioned by a client armed with Claude forces us to check that our own solution is actually the optimal one, not just the familiar one. That's a genuinely useful discipline, even when the conversation that produces it is a nuisance to have. It's easy to build the same kind of flow the same way for years and lose sight of whether that's still the best way to do it, and a client's AI-generated challenge, however impractical the specific suggestion, is sometimes the only thing that forces the question back onto the table.
AI compresses timelines on build and technical work in ways that are hard to argue with. It doesn't compress judgement about what to build, or why. Our work with Bruusta is the clearest example: using Claude Code and Claude Design, we rebuilt their site in around two months, at roughly a tenth of what a traditional rebuild would have cost, in a fraction of the time a full retainer-led rebrand would normally take. But the client had already tried building sections themselves with Shopify's AI tools, and the results were inconsistent fonts, inconsistent spacing, and code that was difficult to modify afterwards. The speed only produced the right outcome because there was still someone defining the requirements, the design direction and the best-practice considerations, and reviewing the code before anything went live. Take that away and the same speed produces the wrong thing faster, not the right thing sooner.
We've compared this moment to what Shopify did for ecommerce: a genuine democratisation of technical capability, where expert deployment stayed just as important even as the baseline task got easier. AI is following a similar shape here. What matters is what an agency does with the capacity that gets freed up. Reinvested into creative and strategic work, the parts of the job that still need someone who has done this before, that freed-up time makes the agency more valuable. Left unaccounted for, it just shows up as margin the agency is no longer earning.
That distinction doesn't land the same way across every part of the business, either. Performance marketing is largely insulated from the AI time-compression problem in a way that build work isn't, because it's billed on commission rather than time. When the unit being sold is a percentage of ad spend rather than hours worked, a faster workflow doesn't automatically shrink the invoice the way it does on a retainer built around time-based build and design work. A media buyer using AI to build campaigns faster is still being paid on the outcome those campaigns produce, and clients can see that outcome directly in profitable, growing revenue built on genuine understanding of the ad ecosystem, not just faster execution. A developer using AI to ship a feature faster is, under a lot of existing contracts, technically being paid less for doing the same job better.
So far, that hasn't shown up as clients trying to negotiate the monthly fee down. Most seem to understand that the invoice reflects something specific: senior decision-making and genuine hands-on involvement, not hours logged. That's what the fee has always paid for. But it's a gap that's easier to explain today, while build timelines are the ones visibly compressing, than it might be once AI starts doing the same to other parts of the business.
We keep coming back to a comparison borrowed from Picasso: the story of him sketching something in a couple of minutes on a napkin, then naming an outrageous price for it. When the person objected that it only took him a few minutes, he pointed out it took him forty years to be able to do it in a few minutes. That logic always applied to expert services. It just wasn't tested very often, because expert time and elapsed time used to move at roughly the same pace. AI breaks that link. The work can now take minutes while the judgement behind it still took years to build, and a billing model built entirely around hours spent will increasingly undercharge for exactly the expertise that makes the fast version possible.
This isn't a hypothetical problem sitting a year or two out. It's already showing up in how contracts get written and how disputes get resolved, and agencies that don't address it directly will end up relitigating the same argument on every account, client by client, rather than settling it once at the commercial level. That's a conversation worth having plainly with a client: a task taking less time doesn't mean it should cost proportionally less, in the same way a doctor's diagnosis isn't priced by the number of minutes spent looking at the scan.
Our answer is to offer both: a lower-cost, AI-driven option and a higher-value, expert-led option, letting the client choose which one fits the decision in front of them. That turns clients realising some tasks are now cheap into a legitimate menu, with the agency setting out plainly what each option does and doesn't include. A brand that genuinely just needs a fast, competent first pass can have that, at a price that reflects how little judgement the task actually required. A brand making a decision it can't afford to get wrong can pay for the version with a person attached to it, and know exactly why that version costs more.
The useful question for a brand to ask a prospective agency partner isn't whether it uses AI. Most now do, in some form. The better question is what that agency still needs a person for, and why. An agency that can answer that clearly, specifically, and without defensiveness is one that has actually thought through where its value sits. One that can't is probably still working it out in real time, on the client's account, at the client's expense.
It's worth listening for a specific kind of answer. A vague reassurance that "we combine AI with human expertise" says nothing, because every agency will say it. A useful answer names the actual boundary: this part is fast because a machine can do it well, this part still needs someone who has shipped this kind of project before and knows where it tends to go wrong. That level of specificity is hard to fake, and it's usually the clearest signal of whether an agency has actually done the thinking or is simply reaching for the phrase that sounds current.
The instinct to feel kept honest by client scrutiny applies here too, just from the other direction. A brand pushing back with a Claude-generated alternative isn't necessarily trying to catch the agency out. It's a reasonable response to having more visibility into what's technically possible than a client had two years ago. Agencies that treat that scrutiny as an attack will lose the relationship. Agencies that treat it as a prompt to justify their own thinking will usually come out of the conversation with a better answer than they started with, and a client who trusts the answer more for having tested it.
Claude is a faster way to get an opinion. It isn't a substitute for someone who has to live with the consequences of being wrong, and it doesn't know how a specific theme is built, what a specific client's customers actually respond to, or which shortcut will look fine in a demo and fall over in production. Expertise still matters exactly as much as it did before. It just has to earn its price every time now, in a conversation the client is better equipped to have than they used to be.
If you're working out where that line sits for your own brand, that's exactly the conversation worth having with us.