Beyond the Buzz: The Hidden Vulnerabilities of AI-Powered Agencies

Paul Sheldon
Senior Creative at Golley Slater

Right now, businesses and agencies are riding the AI hype. The pitch is seductive: deploy autonomous agents, slash production times, stay competitive. But beneath the buzzword-heavy promises lies something fragile.

We’re treating AI like simple software upgrades and in doing so, we’re exposing our businesses to risks most of us haven’t properly thought through.

Here’s the uncomfortable bit: replacing people with autonomous AI doesn’t just cut costs. It fundamentally rewires your agency’s vulnerabilities.

Silent Failures

When traditional platforms like Google Workspace or Microsoft 365 break, they fail loudly. Servers go down, error messages appear, IT deploys a fix. AI agents fail quietly. An agent can remain technically online whilst gradually drifting off course, hallucinating data, making confident but wrong decisions, without triggering a single alarm. By the time you realise a campaign is built on fabricated stats or an agent has sent the wrong deliverables, the reputational damage is done.

Worse still, recent cybersecurity research shows that 90% of AI agents are over-permissioned, exposing sensitive client data to security risks that traditional software simply doesn’t carry.

And sometimes the failures aren’t silent at all. In April 2026, PocketOS, a SaaS platform for car rental companies, had its entire production database wiped in nine seconds. An AI coding agent (Cursor, running Anthropic’s Claude Opus) was working on a routine task in an environment when it hit a credential mismatch.

Rather than flagging the problem, the agent decided on its own initiative to “fix” it by deleting a data volume. The company had to fall back to a three-month-old backup. Car rental operators across the country woke up on a Saturday morning with no reservations, no payment records, and no idea who was picking up a vehicle that day. That’s the world we’re building towards if we hand autonomous agents the keys without proper guardrails.

Brain Rot

If you replace your junior creatives and project managers with AI agents, you bypass the mentorship pipeline entirely. When software does the heavy lifting, human muscles (critical thinking, deep understanding, craft) begin to atrophy. Your people become passive observers rather than active participants.

So, what happens when the AI hits a novel problem it can’t solve? Or when a frontier model goes down for hours? The remaining team won’t have the foundational knowledge or practical experience to pick up the pieces.

The Token Trap

This is the one that bothers me the most but gets glossed over.
Traditionally, if a client demanded last-minute amends, you agree a cost or absorb the alterations. Salaries are fixed. You grumble, you do the work, you move on. AI operates on a volatile token economy, and it changes everything.

Every time you re-prompt an AI with client feedback, the model has to re-read the entire project context, burning massive amounts of tokens with every single iteration. The first draft might be cheap. But the endless back-and-forth of creative feedback, the bit that is the job, scales costs exponentially.

We’re already seeing this play out. Uber’s CTO recently admitted the company blew through its entire annual AI budget in just the first four months of 2026, thanks to unchecked use of autonomous coding agents. By March, 84% of their developers were using agentic tools, and those agents were writing 11% of the live backend code. Usage went up, costs increased six-fold, and the budget was gone by April. They even had internal leaderboards ranking teams by AI usage, essentially gamifying the spend.

If a company of Uber’s scale can’t forecast its AI compute costs, how is a mid-sized design agency going to survive the token burn of a demanding client?

For agencies specifically, this threatens the entire flat-fee model. The AI generates the initial design cheaply, but the iterative rounds needed to finalise the work consume massive tokens. If you don’t update your billing to account for this, last-minute client amends will quietly wipe out your margins.

The Rug Pull

In the rush to automate, we’re outsourcing our core operational logic to a handful of tech giants. That dependency leaves you exposed.

These companies can change the economics of your business overnight. In April 2026, Anthropic quietly removed its autonomous coding tool from its standard £20-a-month Pro plan, pushing reliant developers onto a £100-a-month Max plan. That’s a 500% price jump just to keep existing workflows intact. At the same time, they began moving enterprise customers from predictable per-seat licences to variable per-token pricing.

If your agency’s entire delivery model is built on someone else’s proprietary AI, you don’t truly own your business. They do.

The Accountability Gap

Read the terms of service. Every major AI provider operates under a “shared responsibility” framework that explicitly disclaims their liability for errors, hallucinations, or inaccuracies. If an AI tool infringes copyright, leaks confidential client strategy, or hallucinates bad data in a marketing report, the tech company won’t save you. Your agency could bear 100% of the legal and financial burden.

So, What Do We Actually Do?

Adopting AI to stay competitive is necessary. Blind reliance is reckless. Here’s what I’d recommend:

1. Run local models where you can.

For smaller, specialised tasks like summarisation, drafting, and organising data, local LLMs and SLMs are brilliant. They run on your own hardware or company server, so sensitive client data never leaves your infrastructure. And once the hardware’s in place, you don’t pay per prompt. Modern lightweight models handle structured extraction, analysis, and summarisation perfectly well on a decent machine. Most are free models you can download, so no token fees!

2. Don’t lock yourself to one eco-system.

Build technology-agnostic infrastructure. If you can swap out the underlying model when a provider raises prices, has better models or kills a tool, you stay in control. Get your data house and documentation and processes in order. Before feeding anything into AI, audit it. Make sure it’s clean, structured, and properly labelled. Establish clear human ownership for every AI system you deploy, and build audit trails so when something goes wrong, you can trace exactly how and why.

3. Keep humans in the loop properly.

Not as passive rubber-stampers, but as the decision-makers. Use AI to curate and summarise; let your people make the final calls. This keeps critical thinking sharp and protects against the brain rot problem.

4. Run failure drills.

Traditional disaster recovery focuses on servers going offline, but AI fails differently. It drifts, it hallucinates, it gets locked out when you hit token limits. Regularly simulate an AI outage and force your team to answer honestly: could we still deliver this work manually?