Process before prompt: why your AI workflow matters more than which tool you pick

There is a lot of noise right now about which AI tool you should be using. Viktor pitches itself as "not a tool, a hire." Perplexity wants to be your research department. Claude, ChatGPT, Gemini, and a dozen others are all competing for the same spot in your browser tabs. The conversation in most marketing teams has shifted from "should we use AI?" to "which one?"

That is the wrong question. Or at least, it is not the first question.

The lawyers got there first

In the first two weeks of August 2025, three separate federal courts sanctioned lawyers for submitting AI-generated hallucinations in their filings. One attorney used a well-known legal research database that produced fabricated case citations, which he submitted to court without checking. A California appellate court handed down a $10,000 sanction for citing non-existent legal authority. A US District Court in Oregon went further, fining an attorney $15,500 for citing fake cases and not being sufficiently candid about it.

The throughline across all of these cases is the same: the professional trusted the output. They skipped the verification step. And the court's position was consistent: it does not matter whether the AI mistake was intentional. The professional is responsible for the accuracy of what they submit.

That principle applies in digital marketing just as cleanly.

The platform is not the process

Tools like Viktor, which connects to 3,000+ platforms and promises to produce reports, dashboards, and campaigns on your behalf, represent a genuine step forward in how AI can be integrated into a workflow. But they also represent a new version of a very old problem: the output is only as good as the context you give it, and the context is only as good as the process you have built around it.

If you connect an AI layer to your Google Search Console without first understanding which queries actually matter, what your cannibalisation issues look like, and what the gap between impressions and clicks is telling you, you will get a very confident-sounding answer to a question nobody has properly asked.

The same applies to Ahrefs or Semrush exports. A raw keyword export handed to an LLM will produce output. Whether that output reflects your actual competitive position, your client's priorities, or anything resembling a coherent strategy depends entirely on what you curated and labelled before the AI touched it.

What good input actually looks like

Before you involve AI in any analytical task, the useful question is: what does this data need to say before the AI can do anything sensible with it?

From GSC, that means filtering to the queries you actually care about, grouping by intent, flagging pages where CTR is substantially below impression-weighted expectations, and noting where multiple URLs are competing for the same terms. A raw export of 5,000 queries tells an LLM very little. A curated set of 80 queries with annotations tells it a great deal.

From Ahrefs or Semrush, it means being clear about what you are asking AI to do with the data. Gap analysis is different from content opportunity identification, which is different again from a backlink audit. Each needs different inputs structured in different ways. Treating them as interchangeable is how you end up with generic output that could apply to anyone.

There are real differences between the major models. Some handle structured data more reliably. Some are better at maintaining context across a long document. Some are more prone to hallucinating specific types of content. But in practice, for the tasks most digital teams are using AI for, the model is rarely the limiting factor. The inputs are.

A well-structured prompt with properly curated data will outperform a brilliant prompt with a raw data dump, regardless of which model you are using.

The corollary: you cannot evaluate AI tools fairly without a consistent input process. If your GSC data is a mess, switching from ChatGPT to Claude will not fix the output. It will give you a different flavoured version of the same problem.

The correction loop is not optional

This is the part that tends to get underplayed in the AI efficiency conversation. Yes, AI can compress the time it takes to produce a first draft of a report, a content brief, or a competitive analysis. What it cannot do is remove the need for a human to review and correct that output before it goes anywhere.

The lawyers who were sanctioned were not doing something fundamentally different from what happens in marketing teams every week: they needed something done quickly, used AI to do it, and skipped the verification step. The difference is that their version of publishing unverified AI output came with a five-figure fine.

The marketing version tends to be quieter. A brief that mischaracterises the competitive landscape. A content plan built on keyword data that was never properly filtered. A report that sounds authoritative and is mostly right, except for the parts that are confidently wrong.

A practical starting point

Before you evaluate any AI tool for your workflow, document four things: what data you are feeding it, how that data is prepared, what specific question you are asking, and what the human review step looks like before the output is used. That process will serve you regardless of which model or platform you land on.

The wild west of AI tools is real. Viktor is not alone: there are dozens of platforms making similar promises about connecting your stack and doing the work. Some of them are genuinely useful. All of them require you to have done the thinking first.

The question is not which AI you hire. It is whether your process is ready for any of them.

 If you want to talk through what a practical AI-ready workflow looks like for your team or your clients, get in touch.