AI Projects Fail When Your Process Is Already Broken

Too many AI projects start with the wrong question

Nothing says “transformation” quite like taking a broken 14-step process, adding AI, and completing the wrong thing in 30 seconds. Congratulations! You have expedited the dysfunction.

I made this observation on a podcast a while back, and I keep coming back to it. Too many organizations layer AI onto workflows designed for a different era, usually one involving email chains, mystery spreadsheets, and at least one person named Linda who somehow knows how everything works.

AI can make that workflow faster. It can cut a few clicks. It may even produce a very attractive dashboard announcing that the broken process is now operating 37 percent more efficiently. But it rarely changes the business.

Here’s my view: AI removes a constraint, and if the process behind that constraint was never sound, removing it gives you faster chaos. Real transformation starts with a harder question than which tool to buy. Ask whether the process should exist in its current form at all. That conversation is harder than buying another tool, which is probably why everyone would prefer to discuss the tool.

Faster chaos is still chaos

I see the same failure pattern across client engagements. The biggest mistake organizations make with AI is treating it as an accelerant for a process they haven’t examined in years. If your process is fragmented, undocumented, or held together by workarounds whose origins have been lost to history, AI will only show the dysfunction to a much larger audience, faster and with greater confidence. Apparently, what operational chaos was missing was scale.

Research is catching up to what anyone who has ever inherited a spreadsheet called “Final_Final_UseThisOne” already knows. Early findings from an eight-month study investigating how generative AI changed the work of 200 employees at a U.S. tech company were published in Harvard Business Review (HBR). The researchers found that generative AI didn’t reduce employees’ workloads. People worked at a faster pace, took on a broader scope of tasks, and stretched work into more hours of the day, often without anyone asking them to.

AI made doing more feel possible, so people did more. The researchers recommend shared norms for how teams use AI, including intentional pauses and sequencing work. I agree, and I’d push that one step further. The bigger gain comes from redesigning the work to account for what “more” will mean before AI brings the extra capacity.

Workslop is a management problem wearing an AI hat

Workslop has the same root cause. You know the kind: low-effort, AI-generated work that looks polished and hands the real thinking to whoever receives it. Twelve pages with beautiful formatting containing several confident recommendations, none of which address the questions the report was supposed to answer.

Research published in HBR traces workslop to two things in management’s control: unclear AI mandates and teams that are already overwhelmed. An unclear mandate usually means the organization hasn’t defined the work AI is supposed to support: what it’s for, what good output looks like, or who checks it.

That tells me the process was never clear enough for a tool, or frankly a person, to succeed within it. And because AI produces professional-looking output so quickly, the recipient now spends an hour reverse-engineering five minutes of someone else’s effort. So much for efficiency.

Agents cannot orchestrate their way out of organizational confusion

The stakes rise as AI moves from helping with a single task to coordinating decisions across functions. Field research at companies including Walmart, Amazon, and Medtronic notes that most companies use AI to improve individual tasks and still struggle with the cross-functional work that decides enterprise performance.

The researchers are bullish on agentic orchestration, and their conclusion supports the redesign-first argument: the payoff goes to organizations that redesign the decisions around how people and AI work together. It doesn’t come from handing people more powerful tools and. Bolt an agent onto the current structure and hope the workflow develops self-awareness, and you get the same stalled handoffs at machine speed, with more futuristic names.

An agent can’t resolve unclear authority, contradictory policies, competing incentives, or a governance model based on “we usually ask Bob.” That’s organizational design work, and there’s no prompt that lets leadership skip it.

Redesign before you automate

The organizations getting real value from AI are willing to ask a more uncomfortable question before they ask whether AI can make something faster: Should this process exist at all, in this form?

That question can reveal that half the steps exist because of an old system limitation, a former employee’s preference, a policy nobody can locate, or a control designed to address a problem that disappeared six years ago. Across industries and platforms, I keep finding the same 14-step approval process and the same heroic spreadsheet holding the enterprise together.

Output deserves more scrutiny than it used to get. AI has made code, drafts, and analysis much cheaper to produce, and a kitchen that suddenly gets twice the ingredients doesn’t automatically serve twice as many good meals. If nobody agreed on what to cook, you get more waste and one very unfortunate casserole.

When a team can build ten things instead of two, judgment becomes the bottleneck. If eight of those ten were never going to create value, AI just helps the organization reach the wrong answer faster and in several formats. AI can generate options, but it can’t rescue leaders from deciding which ones deserve investment, what problem they’re solving, or how to measure success. Tragic, I know. Management is still required.

Before your next AI investment, map the process from beginning to end and ask:

• What outcome is this process supposed to produce?
• Which steps create value, and which ones only move work between people?
• Where are employees compensating for unclear rules with manual workarounds?
• Are we measuring a better outcome, or are we merely celebrating greater output?
• Would a smart new hire with no institutional baggage design the process this way on purpose?

If a smart new hire wouldn’t design the process this way, the process needs to change before the AI does. Otherwise, you’re giving the broken process better technology, a larger budget, and the confidence to fail at scale.