AI has made execution cheap: drafts, ad variants, report assembly and routine code take a fraction of the time they did. What has gained value is judgment, taste and accountability for results. Agencies selling hours are exposed; agencies selling outcomes are getting faster.
AI is deflating the price of execution at agencies. First drafts, ad variants, monthly report assembly and routine code now take a fraction of the time they did, and the ad platforms themselves increasingly generate creative and find audiences. What has gained value is everything execution depended on: knowing which problem matters, the taste to tell competent work from compelling work, the willingness to be measured on results, and the experience to catch a confident mistake before it reaches a client's customers. Agencies that sell hours are exposed. Agencies that sell outcomes are getting faster.
We use these tools in our own work every day, so this is written from inside the change. Here is what the evidence says, and how to buy agency services when a deliverable on its own is no longer scarce.
- drop in the cost of running GPT-3.5-level AI, November 2022 to October 2024
- 280x
- drop in the cost of running GPT-3.5-level AI, November 2022 to October 2024
- of organisations reported using AI in 2024, up from 55% a year earlier
- 78%
- of organisations reported using AI in 2024, up from 55% a year earlier
- faster: developers finishing a coding task with GitHub Copilot, in a 2022 trial
- 55%
- faster: developers finishing a coding task with GitHub Copilot, in a 2022 trial
- slower: experienced developers using AI tools on their own projects, in a 2025 trial
- 19%
- slower: experienced developers using AI tools on their own projects, in a 2025 trial
01
Execution got cheap, and the platforms now do some of it
The Stanford AI Index 2025 found that the cost of running a system at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024, and that 78% of organisations reported using AI in 2024, up from 55% the year before. The tools are cheap and widely used, so having them is no longer an advantage on its own.
Speed gains on routine work are real. In GitHub's 2022 trial with 95 professional developers, those using Copilot finished a set task, writing an HTTP server in JavaScript, 55% faster: an average of 1 hour 11 minutes against 2 hours 41 minutes.
The ad platforms are absorbing work agencies once billed for. In May 2025 Google introduced AI Max for Search campaigns, which finds new search queries and writes ad text from your landing page and keywords. At Google Marketing Live later that month, it said creative tools built on its Veo video and Imagen image models were coming to Google Ads and Merchant Center, and unveiled agentic capabilities for advertisers. When the platform can produce the variants itself, a retainer that charges days for them is charging for friction.
02
What got cheaper, and what did not
| Work | What AI changed | What still needs a person |
|---|---|---|
| First drafts of copy and content | Fast, cheap first versions | Deciding what to say, and checking it is true |
| Ad variants and creative | Platforms increasingly generate them | Choosing the offer, the audience and the budget |
| Monthly reporting | Assembly is largely automatic | Explaining what changed and what to do next |
| Routine code and site changes | Faster on simple, well-defined tasks | Architecture, security and review |
| Strategy | Faster research and analysis | Picking the problem worth solving |
03
The judgment premium is real, and measurable
The same tools that speed up simple work can slow down expert work. In a trial METR published in July 2025, 16 experienced open-source developers worked on 246 real issues in projects they knew well. With AI tools they took 19% longer. Beforehand they expected a 24% speed-up, and afterwards they still believed they had been about 20% faster. METR calls it a snapshot of early-2025 tools in one setting, not proof that AI slows everyone down. The lesson for buyers is narrower and more useful: feeling productive and being productive are different things, and only checking tells them apart.
Checking is where unaided AI fails publicly. This week Fortune reported that Deloitte's Australian firm will partially refund the government for a report that cited academic papers that do not exist and included a fabricated quote from a federal court judgment. A University of Sydney researcher, Chris Rudge, flagged the errors. A revised version disclosed that Azure OpenAI had been used, and Deloitte will refund its final instalment.
04
Selling hours versus selling outcomes
Hourly billing made sense when hours were a fair proxy for effort. When a task that took a day takes an hour, an agency billing by the hour must either charge you for time it no longer spends or cut its own revenue. Both pressures push towards pricing on outcomes, or on fixed, clearly scoped deliverables.
Outcomes are the scarce good now: qualified leads, sales, calls from search, and visibility in the places customers look, including AI answers, which still depend on the basics we covered in SEO is not dead. An agency willing to be measured on those has to understand your business, not just operate the tools. That understanding is what you are really paying for.
Take a hypothetical jeweller in Thrissur whose monthly retainer lists twenty social posts, four blog articles and a report. Drafting that list is now the cheap part. The parts worth paying for are the ones not on it: which collections to push before the wedding season, which enquiries turned into sales, and what to stop doing.
05
How to buy agency work now
Ask every agency how it uses AI, and distrust both easy answers. One that says it does not use AI is probably overpriced. One that is only a wrapper around AI adds little you could not prompt yourself. Google's position on content is a useful reference point: it rewards quality however content is produced, and treats automation aimed mainly at manipulating rankings as spam.
- Where exactly do you use AI, and where do you deliberately not?
- Who checks the output, against which sources, before anything is published or sent to customers?
- What will you be measured on? Agree two or three business outcomes, not a count of deliverables.
- Which tools see our customer data, and on what terms?
- Do we own the accounts, the content and the data if we part ways?
- What have you stopped charging for now that it takes a fraction of the time?
Using AI doesn't give content any special gains. It's just content.
What to do with your current agency
- Go through your retainer line by line and mark which items are execution that AI has made cheap.
- Put the six questions above to your agency, and ask for the answers in writing.
- Replace deliverable counts with outcome measures you can track yourself, using analytics you own. Our note on privacy-first analytics is a good place to start.
- Ask for AI use to be disclosed on anything published under your name.
- Keep admin access to your ad accounts, analytics and website, whoever runs them.
Deliverables are cheap now. Accountability for leads, sales and visibility is not. If you want to see how we frame that accountability, each of our services names the measures we work to.
Sources
Facts checked against these sources, as of 10 October 2025.