A post from Pradeep Sanyal went around this week describing something anyone who has worked inside a marketing team recognizes immediately: “many workflows were never designed. They accumulated.” A spreadsheet here. An approval added after one incident. A meeting nobody remembers the reason for. And now that AI shows up, the question everyone asks is whether it can automate that. Sanyal answers with the question that actually matters: shouldn’t we first ask why the thing exists at all?

The example that answers it completely

In the comments on that same post, Kaushik Birmiwal shared a real case: a five-step identity verification (KYC) process at a UAE organization. Digging in, they found three of those five steps existed only because the organization didn’t trust the authenticity of the identity documents it received. Once UAE Pass arrived, government-verified digital identity, those three steps stopped making sense. All that was left to automate was checking name mismatches, expired documents, and a few basic sanity checks.

The real transformation wasn’t automating the five-step process. It was asking whether those steps were still necessary now that the assumption they were built on had changed.

That’s exactly the problem every marketing team has with generative AI right now. A four-signature campaign approval chain. A lead-scoring rule that subtracts points for a criterion nobody can explain anymore. A weekly report still assembled for a director who left the company. None of these were designed on purpose. They piled up, layer by layer, each layer added to solve a specific problem that probably doesn’t exist anymore.

Who depends on the sediment staying exactly where it is

One comment in that same thread cuts deeper than most AI-adoption analysis out there. David Arena, a CTO and transformation advisor, pointed out something uncomfortable: the person who could honestly answer which step is unnecessary is usually also the person whose role depends on that step staying put. Sachin Raj added the piece that completes the thought: almost no extra step got added for no reason. Every one was somebody’s fix for a real problem, which is exactly why it’s so hard to remove. The right question isn’t just what to stop honoring. It’s which old fix is still load-bearing, and which problem that justified it has quietly disappeared.

This matters especially in marketing, because the person managing that four-signature approval chain, or the analyst keeping a three-year-old lead-scoring spreadsheet alive, is often the same person you’d ask for help simplifying it. Nobody builds sediment with bad intentions. Every layer was a reasonable response to a real problem at the time. The mistake isn’t adding it. It’s never going back to ask whether it’s still needed.

A third voice in that thread shifts the conversation entirely. Akshay Kokane, who works implementing AI agents in production, made a point every marketing team should keep in view: teams often bolt an AI agent onto an inefficient process when fixing the root cause would eliminate the need for AI entirely, saving the cost of running it. His rule before automating anything is to ask why the problem wasn’t solved before. Real use-case analysis, not the novelty of adding AI, is what should decide where the investment goes.

What automating sediment looks like next to redesigning it

Picture asking an AI agent to speed up that four-signature approval chain. It’s an easy project to sell internally: what took three days now takes three hours, the team celebrates, the efficiency report looks great. But those four signatures had stopped making sense a year earlier. The only thing you changed was how fast that bottleneck repeats itself.

Take a typical lead-scoring model at a mid-size B2B company. Someone built it three or four years ago: points for downloading a case study, points for visiting the pricing page, negative points for a Gmail address instead of a corporate one, negative points for listing “student” as the job title. Every rule made sense the day it was added. But the market shifted. There are legitimate B2B buyers who use Gmail for everything now, founders who list themselves as students because they just finished an MBA, and the case study that used to score points isn’t the content that converts anymore. Nobody went back to audit those rules. They just kept running, quarter after quarter, without anyone asking whether they still measured real buying intent.

Now hand that same model, exactly as it is, to an AI agent and ask it to run faster across ten times the lead volume. You’ll get precisely that: the same outdated model running over ten times more leads, at the same error rate, except now sales complains ten times faster that the “qualified” leads don’t qualify for anything.

The same pattern shows up in content and brand review. Plenty of teams still route every piece of content through three or four reviewers because one post went out with a bad fact two years ago, and nobody since has had the authority to publish without that committee pass. An AI agent that checks spelling, tone, and brand compliance can replace one or two of those reviews perfectly well. But if the real reason for four signatures was that nobody wanted to be the sole person accountable for a mistake, no agent fixes that. That’s a question of who owns the decision, not how fast the review runs, and automating it without naming an owner just moves the bottleneck from content review to approving who gets to run the agent.

What we do at IQ Source

When we start the discovery phase of AI Maestro with a marketing team, we don’t begin by asking what can be automated. We start by mapping every workflow (campaign approvals, lead scoring, recurring reports, nurture cadences) and asking, for each step, whether it exists for a reason that still holds or for an incident nobody remembers. We come out of that phase with a Process Reality Map that separates what needs redesigning from what’s ready to automate as-is, and use that to decide together whether the next move is an automation layer, a tweak to the underlying process, or cutting the step entirely.

This isn’t an abstract exercise. It’s the same logic behind why I’ve argued for building the system instead of buying a chatbot: the difference between giving a bottleneck speed and removing it. Before you automate your marketing team’s next workflow, ask the question Kaushik asked: does this step still exist for a real reason, or because nobody went back to check?

Map your marketing team’s sediment before you automate it