Your Agency Built Its AI Workflow Around Sycophancy — And Didn't Notice

The Accidental Audit

Something revealing happened in February 2026. Anthropic told the Pentagon no, the White House retaliated, and the American public — in a fit of contrarianism that would make any marketer weep with envy — made Claude the number one app in the country. Millions of downloads in days. Most of those people had never heard of Anthropic a fortnight earlier.

That's the headline. Here's the story underneath it: the vast majority of those new users are going to open Claude, type the exact same prompts they use in ChatGPT, get shorter and occasionally uncomfortable responses, and conclude that the app is worse. They'll be wrong, but they'll be wrong in a way that reveals something important about how we've all been using AI — particularly in agencies.

Because when you switch from a model that's been optimised to agree with you to one that's been trained to be honest with you, you don't just get different outputs. You get a mirror held up to your entire workflow. And what most agencies are going to see in that mirror isn't flattering.

The Sycophancy Problem Nobody Talks About

Let's be precise about what sycophancy means in this context, because it's not a personality quirk. It's an architectural outcome.

ChatGPT is trained primarily through reinforcement learning with human feedback (RLHF). Humans rate responses. Responses that feel helpful, thorough, and agreeable get higher ratings. The model optimises for those ratings. The result — documented by OpenAI's own researchers and made spectacularly visible when a GPT-4o update in April 2025 had to be rolled back within days because the sycophancy became so extreme — is a model with a structural tendency to tell you what you want to hear.

OpenAI has done serious work to address this. The current ChatGPT is measurably less sycophantic than the version that triggered the rollback. But the underlying orientation hasn't disappeared, because it's baked into the training methodology. When your reward signal is human satisfaction, you optimise for human satisfaction — and humans are reliably more satisfied by agreement than by challenge.

Claude takes a different approach. Anthropic's constitutional AI trains the model against explicit principles — be helpful, be honest, avoid harm — rather than purely optimising for what feels like a good response. The practical effect: Claude is more likely to flag a problem than smooth it over. More likely to ask what you're actually trying to achieve. More likely to tell you your plan has a hole in it.

Now here's where it gets uncomfortable for agencies. Over the past two years, thousands of ecommerce agencies have built their AI-assisted workflows around ChatGPT's defaults. Content production pipelines. Strategy documents. Client deliverables. Audit frameworks. All of them shaped by a tool that, by design, tends to validate rather than challenge.

That's not a tools problem. That's a quality control problem wearing a tools costume.

What Sycophancy Looks Like in Agency Work

Let me make this concrete, because abstract concerns about AI training methodologies don't pay invoices.

A typical mid-market ecommerce agency in 2026 is using AI for content production, technical auditing, strategic recommendations, and client reporting. In most of these agencies, the AI workflow looks roughly like this: brief goes in, output comes out, a human gives it a quick scan, maybe edits for tone, and it ships.

The problem is what happens during that "quick scan." When your AI consistently produces outputs that sound confident, thorough, and aligned with whatever direction you've already decided on, the human review step atrophies. Why would you scrutinise something that keeps confirming your assumptions? The tool has become a validation machine, and validation machines are the enemy of quality work.

I've seen this in practice. An agency produces a "comprehensive" SEO audit for a client using ChatGPT-assisted analysis. The audit is 4,000 words, professionally formatted, contains specific recommendations. It also fails to mention that the client's site has 340 orphan pages, because the prompt didn't explicitly ask about orphan pages, and the model — optimised to satisfy the stated request — delivered exactly what was asked for rather than what was needed.

A strategy deck recommends expanding into three new markets. The AI helped research each market, produced compelling slides for each, and at no point flagged that the client's logistics infrastructure can't support delivery to two of the three regions within an acceptable timeframe. The plan was plausible. The plan was also wrong. But it sounded right, and that was enough to get through internal review.

These aren't hypothetical scenarios. They're the kind of mistakes that emerge when your primary AI tool is trained to complete the assignment rather than question it. Pixelpeaks' 500-task comparison found Claude hit 94% exact instruction compliance versus ChatGPT's 87%. That sounds like a minor difference until you multiply it across 200 client deliverables a quarter. Across an entire year of output, 7% more drift means dozens of deliverables that technically answered the prompt but missed the point entirely.

The Real Cost Is in What Didn't Get Flagged

Here's the number that should worry every agency principal: the most expensive AI mistakes are not factual errors. They're plans that should never have been executed.

A factual error — wrong revenue figure, incorrect market size — gets caught in review. It's visible. It's correctable. The mistakes that actually cost money are structural: strategies built on unchallenged assumptions, campaigns launched without stress-testing the premise, recommendations that optimise for the wrong metric because nobody asked whether the metric was right in the first place.

In the pre-AI era, these mistakes got caught (sometimes) by the expensive humans in the room. The strategy director who'd say "hang on, have we actually validated that assumption?" The senior developer who'd ask "what happens when this hits 10x the expected traffic?" The experienced account manager who'd push back on a timeline because they'd seen enough projects to know the difference between an optimistic estimate and a realistic one.

AI was supposed to augment that scrutiny. Instead, in many agencies, it's replaced it. Not because the humans stopped being capable of critical thinking, but because the tool's default behaviour — confident, thorough, agreeable — created a workflow where critical thinking became optional. When every draft sounds good, you stop interrogating whether it is good.

This is what I mean when I say agencies built their workflows around sycophancy. It wasn't intentional. Nobody sat down and said "let's design a process where AI validates our assumptions without challenging them." It happened gradually, because the tool's default behaviour aligned perfectly with the path of least resistance in a busy agency: get it done, make it sound right, move on to the next client.

The Claude Test Your Agency Should Run

Here's a genuinely useful exercise. Take five recent client deliverables — strategy documents, audits, content pieces, whatever your agency ships most often — and run them through Claude with a single prompt: "What's the weakest argument in this document, and what's missing that should be here?"

Then run the same documents through ChatGPT with the same prompt.

The differences will be instructive. Not because Claude is infallible (it absolutely isn't), but because a model trained to evaluate framing rather than satisfy requests will interact with your work differently. Access Intelligence's independent comparison found that in blind testing with over 100 voters per round across eight prompts, Claude won four of eight rounds — and users consistently rated Claude's outputs as more natural and publishable. More relevantly, Claude scored 85% on structural coherence of analysis versus ChatGPT's 78%.

That structural coherence matters enormously for agency work. An audit that identifies surface-level issues but misses the structural problem is worse than no audit at all, because it creates a false sense of thoroughness. A strategy that reads well but doesn't hold together under scrutiny is actively dangerous, because it's convincing enough to get approved and implemented.

I'm not suggesting agencies should switch wholesale to Claude. That would be idiotic advice — ChatGPT genuinely does some things better, including image generation, real-time web research, and the sheer breadth of its ecosystem with custom GPTs and the emerging app store. But I am suggesting that if the only AI in your workflow is one that tends to agree with you, you've built a system with a critical vulnerability that you probably can't see from the inside.

Multi-Model Isn't a Luxury. It's Quality Assurance.

The mature position — and I recognise that "mature" is doing a lot of work in that sentence — is that agencies in 2026 should be running multi-model workflows. Not because it's trendy. Because it's the only reliable way to build challenge into an AI-assisted process.

Think about it from a quality control perspective. Every serious engineering team runs multiple testing approaches. Unit tests, integration tests, end-to-end tests, code review by a different human. The entire point is that a single perspective — no matter how capable — will miss things that a second perspective catches. This is not a controversial idea in software development. It should not be a controversial idea in agency work.

A practical multi-model workflow doesn't require doubling your AI spend or redesigning your entire process. It requires adding a challenge layer. Use whatever model you prefer for generation. Then use a differently-trained model to stress-test the output. Claude's extended thinking capability — where the model allocates additional processing to work through problems step-by-step and shows you the reasoning chain — is particularly useful for this, because you can actually watch the model's reasoning unfold and intervene when it goes off track. Anthropic reports up to 54% improvement on hard reasoning tasks when extended thinking is engaged.

The specific workflow matters less than the principle: your AI process needs a built-in adversary. If every tool in your pipeline is optimised to make the output sound good, then nobody — human or artificial — is optimised to make the output actually be good.

And this is where the ecommerce angle gets particularly sharp. In agency-land, the client is paying for your judgement, not your ability to produce confident-sounding documents. If your AI workflow has quietly replaced judgement with validation, you're not delivering a service. You're delivering a liability with nice formatting.

The Habit-Shaping Problem

There's a subtler issue here that deserves attention, because it affects not just the quality of agency output but the quality of agency thinking.

When you work with a tool that agrees with you by default, you stop expecting challenge. You stop framing requests in ways that invite pushback. You stop noticing when an output confirms your existing assumption without evidence. Over months of daily use, this reshapes how you think about problems — not just how you use AI, but how you approach analysis, strategy, and decision-making generally.

Conversely, when you work with a tool that's somewhat more likely to say "I'm not sure the way you've framed this is optimal," you develop different habits. You start pre-emptively strengthening your arguments because you expect challenge. You start including more context because you've learned the model does more with richer inputs. You start treating AI interaction as a thinking process rather than a production process.

This is the difference between using AI as a content factory and using AI as a thinking partner. Both are valid uses. But agencies that have spent two years in content-factory mode may find that their strategic muscles have quietly atrophied. The Claude moment isn't really about Claude. It's about what happens when millions of people suddenly encounter a tool that treats them as someone who might benefit from being challenged rather than someone who needs to be pleased.

For agencies specifically, this matters because your clients can increasingly do the content-factory work themselves. ChatGPT is free. Claude is free. Gemini is free at the basic tier. The commodity layer of AI-assisted content production is already being consumed directly by brands. What clients are still willing to pay agencies for is judgement, strategy, and the kind of rigorous analysis that requires someone to say "your assumption is wrong" rather than "here's a beautifully formatted version of your wrong assumption."

If your AI workflow has been quietly optimising for the latter, the Claude surge is your wake-up call. Not because Claude is the answer — it has real limitations, including a notable weakness in mathematical reasoning, no image generation, limited real-time web search, and less persistent memory than ChatGPT. But because encountering a model with different defaults forces you to confront the defaults you've been living with.

What Actually Changes

Let me be specific about what a sycophancy-aware agency workflow looks like, because vague advice about "being more critical" helps nobody.

First, every AI-assisted deliverable gets a challenge pass before it ships. This can be Claude, it can be a different model, it can be a senior human — but something in the process is explicitly tasked with finding weaknesses rather than polishing strengths. The prompt isn't "review this for quality." It's "what's wrong with this, what's missing, and what assumption hasn't been validated?"

Second, project-level context becomes mandatory rather than optional. Claude's Projects feature — and equivalents in other tools — allows you to set operating rules that persist across every conversation in a workspace. Not "help me with marketing" but "I'm working on a DTC brand in home furnishings targeting 25-40-year-olds in the UK, average order value £180, primary competitors are [X, Y, Z], our client's differentiator is [specific thing]." Every output inherits that context. Every output is therefore more likely to be relevant and less likely to be generic.

Third, agencies need to stop treating AI outputs as first drafts and start treating them as hypotheses. A first draft gets edited for tone and shipped. A hypothesis gets tested against evidence and either strengthened or discarded. The difference is enormous, and it's the difference between an agency that uses AI to produce more stuff faster and an agency that uses AI to produce better thinking.

Fourth — and this is the uncomfortable one — agencies need to accept that a tool which pushes back will slow things down. Not dramatically, but noticeably. The brief that used to go from prompt to polished output in twenty minutes now takes thirty because the model flagged a concern that required additional research. That extra ten minutes isn't inefficiency. It's the cost of quality that was previously being skipped.

The agencies that thrive in the next eighteen months won't be the ones with the most sophisticated AI tooling. They'll be the ones that figured out that AI's most valuable capability isn't generating content — it's stress-testing thinking. And stress-testing, by definition, requires a tool that's willing to tell you something you didn't want to hear.

Millions of people just downloaded that tool. Most of them will give up within a week because it doesn't generate images. The ones who stick around — and the agencies that build workflows around the capability rather than running from the discomfort — will have a genuine competitive advantage.

Not because Claude is better. Because challenge is better than agreement. And we've spent two years forgetting that.