The productivity case for AI assistance is real and it comes with a footnote nobody quotes. A controlled experiment with 453 professionals published in Science found writing tasks completed roughly 40 percent faster with output rated 18 percent higher, and the researchers noted plainly that accuracy was not assessed, because the tasks did not require precise factual accuracy. That footnote is the entire problem for agencies, since accuracy is the deliverable. Google’s position points the same way: it judges content on quality rather than how it was produced, while its scaled content abuse policy targets mass-produced material regardless of whether automation was involved. The agencies that keep quality do not rely on a human glancing over the output, because human factors research shows reviewers of automated output systematically miss errors. They assign a named person accountable for every factual claim.
There is a question clients have started asking on sales calls, sometimes directly and more often sideways: am I paying you to run this through a chatbot. It is a fair question, the honest answer at most agencies is partly yes, and the interesting part is not whether AI is used but what happens between the model producing something and a client receiving it.
The two loud positions on AI-assisted workflows are both wrong. Refusing to use these tools means charging clients for hours that no longer need to exist. Using them without changing the process means shipping work that reads fine and cannot be defended. What follows is what the evidence actually supports about AI-assisted workflows, where quality really fails, and the controls that hold.
The productivity case is real, and it has a footnote
Start with the strongest evidence in favour of AI-assisted workflows, because pretending the tools do not work is not a defensible position either.
A controlled experiment with 453 college-educated professionals, published in Science, gave marketers, consultants, analysts, grant writers and managers two occupational writing tasks. Those with access to the tool finished around eleven minutes faster, roughly a 40 percent reduction, and independent evaluators rated their output 18 percent higher. Performance inequality between workers also narrowed, with the weaker writers gaining the most.
Those are meaningful numbers and they are not the whole story. The researchers were explicit about what the experiment did not measure, and the limitation is precisely the thing an agency is paid for.
The study did not assess accuracy, because the tasks did not require “precise factual accuracy or context about things like a company’s goals or a customer’s preferences.”
Read that alongside the headline figures and the correct conclusion is narrow rather than sweeping. These tools reliably improve speed and perceived quality on tasks where being right is assumed rather than tested. Agency work is the opposite: the client’s goals, the market’s constraints, and the factual claims are the substance. That distinction is the one we build the process around, and it is why the people who do the work here are named on it rather than hidden behind a house style.
What Google actually cares about, and what it does not
A lot of anxiety about AI-assisted workflows rests on a misunderstanding of the search side, so it is worth settling before discussing process.
Google’s published guidance on AI-generated content, issued in February 2023, takes the position that it rewards high-quality content regardless of how it was produced, while using automation to generate content primarily to manipulate rankings violates its spam policies. The method is not the issue. The intent and the result are.
The March 2024 core update sharpened that with the scaled content abuse policy, which targets producing many pages primarily to manipulate rankings rather than to help people, and applies whether the pages were made by automation, by humans, or by some combination. The policy deliberately removed automation as the deciding factor, which means an agency churning out volume with human writers is in exactly the same position as one doing it with a model. Getting that distinction right is the foundation of the position we have taken publicly on AI in search, because it moves the question from whether you used a tool to whether the output deserved to exist.
Why “a human reviewed it” is not a quality control
Ask how agencies use AI responsibly and almost every one answers the same way: a human reviews everything before it goes out. That sounds like a control and mostly is not one, for a reason documented decades before these tools existed.
Human factors researchers studying automated decision aids identified automation bias, the tendency of people to over-trust automated output and reduce their own vigilance. It produces two distinct failures: omission errors, where the reviewer misses a problem the system did not flag, and commission errors, where the reviewer follows an incorrect recommendation against evidence available to them. Both increase precisely when the automation is usually right, because a tool that is correct nine times out of ten trains its reviewer to stop looking.
This is the central problem with human in the loop content as most agencies practise it. The loop exists, the human is present, and the review has quietly degraded into a readability check because the drafts are usually decent. The same body of research found that accountability changes this: participants who believed they were personally answerable for accuracy showed measurably less automation bias. Naming an owner is not a formality, it is the intervention, and it is why review responsibilities are assigned by name rather than by role in the methodology we work through on every engagement.
Which tasks survive automation and which do not
Designing AI-assisted workflows well comes down to one distinction, and it is not between writing and not writing. It is between tasks where an error is visible and tasks where an error is invisible until a client repeats it in a meeting.
| Task | Suitability | Why |
|---|---|---|
| Structuring, outlining, reformatting | Safe | Errors are structural and immediately visible |
| Summarising a source you supplied | Safe with the source open | Checkable in seconds against the original |
| Variants of copy you already approved | Safe | The claim was already verified once |
| Drafting from a brief | Conditional | Fine for prose, dangerous wherever a number appears |
| Sourcing statistics or citations | Unsafe | A plausible fabricated citation is indistinguishable from a real one |
| Claims about a client’s business | Unsafe | The model has no access to the truth and will invent a confident version |
The bottom two rows account for nearly every genuine failure. A fabricated statistic carries no signal that it is fabricated, reads more fluently than a real one because it was written to fit the sentence, and survives review precisely because nothing about it looks wrong. Everything we publish, including the analysis on our own blog, runs on the rule that a number without a link to its source does not ship, whoever drafted the sentence around it.
Two production lines drawn in parallel. The upper line runs brief, draft, review, publish, with the review drawn as a single wide gate labelled “looks fine” that everything passes through at the same speed regardless of what it contains. The lower line runs the same first two stages, then splits the review into two narrower gates: a prose gate handling structure, tone and clarity, and a separate claims gate where every number, citation, date and statement about the client is checked against a source and initialled by a named person. The lower line is slower by one step. It is the only one of the two where a fabricated statistic has anywhere to be caught.
The controls that actually hold at scale
Four practices separate AI content quality control that works from the kind that only looks like it does, and they separate agencies whose quality survives AI adoption from agencies whose quality quietly erodes, and none of them is a tool.
All four are process changes rather than tooling decisions, which is why they survive a change of model, a change of vendor, and a change of staff. Anything that depends on a particular tool being good will need redoing the next time the tool changes.
Separate the prose review from the claims review. They use different attention and merging them means the fluent-sounding half absorbs all of it. Reading for clarity and reading to verify a figure are not the same task and should not happen in the same pass.
Require a source link for every factual claim, at draft time. Not at review, at draft. A claim that arrives without a source is either verified before it enters the document or removed, which converts fact-checking from a search exercise into a confirmation one.
Name the accountable person per deliverable rather than per role. This is the intervention the automation bias research supports directly, and it costs nothing except the discomfort of being the person whose name is attached when something is wrong.
Keep humans on anything requiring context the model cannot have. Strategy, client-specific claims, competitive positioning, and anything touching a regulated category. Governing that consistently across many people and properties is a policy problem rather than a skill problem, and it is a defining feature of the enterprise programs we manage, where one loose process replicates across thousands of pages before anyone notices.
Two places quality dies quietly
Beyond drafting, two applications of AI-assisted workflows cause more damage than the rest combined, and both are attractive precisely because they scale so well.
The first is translation and localisation. Machine translation has become genuinely good, which is exactly what makes it dangerous: the output reads fluently to someone who does not speak the language, so nobody can tell the difference between a good translation and a confidently wrong one. Idiom, regulatory phrasing, and category terminology are where it fails, and a native reviewer per market is the only control that works. We treat that as non-negotiable on the multi-market programs we run, because a fluent mistranslation is invisible to everyone on the account except the customer.
The pattern in both cases is the same and worth naming, because it predicts where the next failure will appear. Risk concentrates wherever the output is fluent, the volume is high, and the person reviewing it lacks the specific knowledge required to spot an error. Any task matching all three deserves a control before it scales, not after.
The second is scaled page generation. Producing hundreds of near-identical location or service pages became trivial, and that is the precise pattern the scaled content abuse policy describes. The test is not whether a person or a model wrote them but whether each page would exist if the ranking benefit disappeared. If it would not, volume is the strategy and no amount of editing repairs that.
The technical side has the same shape
AI-assisted workflows on the technical side behave the same way as they do on the prose side, with one useful difference: much of it can be verified automatically.
Schema markup, redirect maps, and configuration files are all fast to produce and all fail silently when wrong, since nothing about a plausible-looking JSON block announces that a property has been invented or a rule points at a path that no longer exists. The advantage over prose is that validators, tests, and staging environments exist, so the correct process is to generate freely and verify mechanically rather than by reading. That is how it is handled in the development work we take on, where the review gate is a test suite rather than a second pair of eyes.
What to ask an agency about this
If you are buying AI-assisted workflows rather than running them, the useful questions are procedural rather than philosophical, because everyone will say they use AI responsibly.
Ask which specific tasks in your engagement involve AI assistance and which do not. Ask who is accountable by name for the factual accuracy of a deliverable. Ask what happens to a statistic between a draft and your inbox, and expect a description of a process rather than a reassurance. Ask whether anything is produced in volume, and why each item would exist without the ranking benefit. An agency with real AI-assisted workflows answers all four immediately, because it has had to decide the answers in order to operate, and the ones treating AI as an efficiency secret tend to go vague at the second question. That transparency is part of what we bring to the Generative Engine Optimization work we deliver, where the irony of using these systems to earn visibility inside them is not lost on anyone here.
Frequently Asked Questions
Does Google penalise AI-generated content?
Not for being AI-generated. Google’s February 2023 guidance takes the position that it rewards high-quality content regardless of how it was produced, while using automation to generate content primarily to manipulate rankings breaches its spam policies. The March 2024 scaled content abuse policy went further by applying whether or not automation was involved, which removed production method as the deciding factor entirely.
Does AI assistance actually improve output, or just speed?
Both, on the tasks that have been tested. A Science study of 453 professionals found writing tasks completed roughly 40 percent faster with output rated 18 percent higher by independent evaluators, and the gap between stronger and weaker writers narrowed. The critical caveat is that the study did not assess accuracy, because the tasks did not require precise factual accuracy.
Is human review enough to catch AI errors?
Not on its own. Human factors research on automation bias documents that people over-trust automated output and reduce their vigilance, producing omission errors where a problem goes unnoticed and commission errors where an incorrect recommendation is followed. The effect strengthens when the automation is usually right. Accountability reduces it, which is why naming the responsible person matters more than adding another reviewer.
Which tasks should never be automated in agency work?
Sourcing statistics and citations, and any claim about a client’s business. Both fail invisibly: a fabricated citation looks exactly like a real one and often reads better, because it was written to fit the sentence. Strategy, competitive positioning, and anything in a regulated category also need a human who holds the context the model cannot have.
What is scaled content abuse and does AI cause it?
It is Google’s policy against producing many pages primarily to manipulate rankings rather than to help people, introduced alongside the March 2024 core update, and it applies whether the pages were made by automation, by humans, or by a combination. AI makes the pattern cheaper rather than causing it. The honest test is whether each page would still exist if the ranking benefit vanished.
Is machine translation safe for international content?
Only with a native reviewer per market. Machine translation has become fluent enough that nobody on the account can distinguish a good translation from a confidently wrong one without speaking the language, and the failures cluster in idiom, regulatory phrasing, and category terminology. Fluency is what makes it risky rather than what makes it safe.
How should AI-assisted workflows handle generated code and schema?
Generate freely and verify mechanically. Schema blocks, redirect maps, and configuration files fail silently when wrong, but unlike prose they can be checked by validators, tests, and staging environments. Reading generated code for correctness is the weakest available control; running it against a test suite is the strongest.
Should agencies disclose their use of AI to clients?
Describing the process is more useful than a blanket disclosure. Clients rarely object to a tool being used for structuring or drafting; they object to discovering it after the fact, and to not knowing who verified the facts. Naming which tasks involve assistance and who is accountable for accuracy answers the real concern.
Does using AI mean an agency should charge less?
It means the value should sit somewhere defensible. If the deliverable is words on a page, the price of words has fallen and pretending otherwise is untenable. If the value is judgment, verification, strategy, and accountability for being right, those got no cheaper and arguably got scarcer. Agencies billing for typing time have a real problem; agencies billing for correctness do not.
Skyfield Digital will walk you through which tasks involve AI assistance, who is accountable by name for every factual claim, and what happens to a number between a draft and your inbox.
Sources
| Science | Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence |
| MIT News | Study Finds ChatGPT Boosts Worker Productivity for Some Writing Tasks |
| Google Search Central Blog | Google Search’s Guidance About AI-Generated Content |
| Google Search Central Blog | What Web Creators Should Know About the March 2024 Core Update and New Spam Policies |
| Google Search Central | Spam Policies for Google Web Search |
| The International Journal of Aviation Psychology | Automation Bias: Decision Making and Performance in High-Tech Cockpits |
