Fifteen ways AI goes wrong inside a company.
Which ones sound like your company?
Fifteen ways AI goes wrong inside a company. Each one has its own page, with what it costs and a first move.
Owners and leaders who want to see what's slowing AI down at their company.
Want a second opinion on where to start? Book a 30-minute call.
Many of these patterns start with one of three setups: AI run by a committee, AI split across department heads, or AI handed to an interested volunteer.
High means the pattern blocks AI from working anywhere in the company or carries real risk. Medium means it wastes money or time and usually travels with a High one.
The lines under "How it sounds" are typical phrasing we wrote, not quotes from clients or interviews.
Each pattern pulls down one or two areas most, and touches others.
Strategy and ownership · People and AI fluency · Culture and change readiness · Data, security and rules · Workflows and know-how · Measurement and quality
Everyone owns a piece of AI, and nobody owns the whole.
“We set up an AI working group, and it meets every other Thursday.”
“Each department picked its own tool, and we'll compare notes.”
“We're waiting on the committee's recommendation.”
A group with one person from each department gives AI a couple of hours a week. Say 6 people meet for 2 hours: that's 6 × 2 = 12 hours of company time, or 12 ÷ 8 = 1.5 eight-hour days, in one meeting (an illustration, not a sourced figure). Then everyone goes back to the day job with a list of to-dos and works them piecemeal between other work. If the next meeting is weeks away, the list waits and gets done in a rush right before it. The meetings produce lists, demos and small pilots. The decisions that cost something, like which tools, what data stays out and which work changes first, get pushed to the next meeting. Each member reports back to their own boss, so each department's view wins inside its own area.
A committee feels fair and safe. Nobody has to give one person authority over other people's departments, and nobody carries the blame alone if it goes wrong. Each member's first team is their department, so the committee works for the departments and nobody works for the whole business. Our view, which no study measures directly, is that committees also tend to settle on what nobody objects to, and the options that matter most are often the ones somebody objects to.
This is the first of the three setups. AI only gets the time left over after the day job, so it never becomes a first-class priority.
Keep the group as a sounding board and name one owner with the time and authority to decide. Write down the three kinds of decisions the owner makes alone and the ones that come back to the leadership team.
The person who got AI still has a day job.
“Our IT director is handling AI on top of everything else.”
“She's the one who's into AI, so it's hers now.”
“We'll get to the AI plan after the busy season.”
AI gets added to the CTO's, CIO's or IT director's list, or to whoever showed interest first, like the person who built a quick app with an AI coding tool. The interested volunteer is the most common version: someone keen offers to lead, and they're often the best person for it. They're capable and willing, and they still have no time for both jobs, so the work stalls, takes much longer than it should and drifts without a clear focus. The AI work happens in the gaps between their real job, gets dropped every time the day job calls, and has to be picked back up from scratch.
The owner can't justify a full-time hire for a job nobody has defined yet, and handing it to a trusted person feels efficient. It's the Hero Culture loop: the more the company relies on one person, the busier they get, and the less room they have for the new work. The person is never the problem. The setup asks them to do two jobs in the hours of one.
Count the hours. Ask the person carrying AI how much time they gave it last month and what got dropped to make room. Then decide on purpose: protect that time in writing, or give AI to someone whose job it is.
One peer asked to change how everyone else works, with no authority to make it stick.
“Talk to Dave, he's our AI champion.”
“I can show them how it works, but I can't make them use it.”
“Why is he telling my team how to do their jobs?”
The owner names a respected peer as the AI champion. The champion runs demos, shares tips and builds a few tools. Use picks up in their own team and stalls in departments whose leaders didn't ask for it. The friction rarely shows as an argument. It shows as polite non-use, and the champion's relationships with peers get harder.
Change needs a sponsor with authority. A peer has influence and no authority, so every request to change how a department works is one peer asking another, and that puts both of them in an awkward spot. People also want to hear about change from two places, and a peer isn't either one. There's a fair point on the other side: MIT NANDA's 2025 report found that many strong deployments began with power users who became early champions, paired with "executive accountability." The champion helps. The champion without a sponsor is the pattern.
Often the interested volunteer, the person in AI as a Second Job, is asked to roll AI out to everyone else. AI as a Second Job is the time problem. This pattern is the authority problem.
Keep the champion as the teacher and give the change a sponsor. The owner says why at a team meeting, and each manager tells their own team what changes for them and when.
No decisions made on purpose, so whatever grows in the dark is what you get.
“Marketing uses one tool, sales uses another and finance won't touch it.”
“We'll let it develop and see what sticks.”
“I didn't know operations had built that.”
Each department picks its own tools, rules and uses. Nobody has a list of what's in use or what data it touches. Good ideas stay in the team that found them, and the same problem gets solved three different ways.
Leaders don't want to slow anyone down, and making decisions feels early when the tools change every month. Without a decision, the defaults decide: the free plan, the tool someone saw on LinkedIn, the department with the most energy. It's the Silo Republic with faster tools.
Here nobody was given AI, so nothing was decided. In AI by Department, each head was given AI for their area, and each area decides on its own.
Spend one hour making the list: every AI tool in use, who uses it, on which plan and touching what data. Making the list is the first decision the company has made on purpose.
Your people already use AI. The open questions are whose account it's on and where the data goes.
“I paste the client file into my ChatGPT and it writes the summary.”
“I built a little tool over the weekend that pulls from our customer list.”
“Which plan? The free one, I think.”
Staff use personal AI accounts for work and paste in client files, contracts, pricing and code. Some build their own apps with AI coding tools, connected to company data. Nobody knows which of those personal plans can use what's typed into them to train the provider's models.
People have real work, and the tools are a browser tab away. When the company hasn't provided approved tools or written rules, people fill the gap themselves, and most of them are trying to do good work faster.
Move it into the open instead of banning it. Ask what people use and why, put work use on business plans the company controls, and write one page on what data stays out. Then pick the best home-built tools and make them supported ones.
Every tool made sense when you bought it. Together they don't add up to a plan.
“We've got Copilot, ChatGPT Team and two AI add-ons in the CRM.”
“Who's using the meeting-notes tool?”
“The vendor said it would save ten hours a week.”
Subscriptions pile up: suite add-ons, chat tools and AI features inside every app the company already pays for. Each purchase answered a demo or a pitch. Few are tied to a task, an owner or a measure, and use drops off after the first month.
Buying is the easiest AI decision there is, and it feels like progress. Nobody inside was asked which friction the tool should remove first. These are solutions looking for problems, based on what worked elsewhere instead of what this company needs.
List every AI tool with its monthly cost, its owner and the one task it's for. Anything without a task gets a 30-day test with a named owner or gets cancelled. Start with the friction, then ask how AI can make it better.
So many people selling AI answers that it's hard to tell good help from bad.
“Every week there's a new tool I'm supposed to be behind on.”
“Three vendors told me their AI employee replaces a hire.”
“I watched a video that said we should automate everything.”
The owner gets AI pitches every day, reads advice that contradicts itself and sees peers posting about their wins. The leadership team can't agree, partly because each person is hearing different noise. Decisions get made on the last demo anyone saw, or get put off until things settle down.
The market moves fast and claims are cheap. With AI, the whole market can feel like a mystery to an owner, and there's rarely someone with no product to sell who can sort it for them.
Noise is what fills the plan when busy leaders own AI part time. In AI by Department, it turns into recency: the plan follows whatever each head heard about last.
Before the next demo, write down the three business problems you'd pay to solve this year. Judge every pitch against that list, and say no to anything that doesn't touch one of the three.
Lots of promising pilots, and none of them became how the work gets done.
“The pilot went great, and then everyone got busy.”
“We tried AI for that last year.”
“It works for Jen, but nobody else uses it.”
Pilots start with energy and show good results for one person or one team, and then they stop. Nobody owns the move from test to everyday process, nothing gets written down, and the next pilot starts from zero. Over time people stop volunteering, because they've seen how it ends.
Pilots are set up to start, not to finish. There's no owner for the rollout, no success measure agreed up front and no time set aside to change the process around the tool.
Before the next pilot starts, write three lines: who owns the rollout if it works, what number says it worked and which written process changes. If nobody can fill in the first line, don't start the pilot yet.
The hardest AI options never get said out loud inside the company.
“Nobody's going to suggest automating their own department.”
“I had an idea, but it felt too far out to bring up.”
“We all know what should happen, and nobody wants to be the one to say it.”
Leadership conversations about AI stay safe: small tools, no changes to roles, structure or what the business sells. The bigger options, like redesigning a role, combining two teams, stopping a service or rethinking pricing, come up in hallway talk or not at all. The owner ends up choosing from a shorter list of options than the business has.
Normal human reasons. People inside weigh how an idea could affect their own job, their peers' jobs and how they'll be seen for raising it. Their position makes some things hard to say, and that's a fact about the position, not a flaw in the person. Westrum's work on how information flows in organizations and Lencioni's work on fear of conflict both point to the same place.
Ask each member of the leadership team to write down, without their name, one AI option they think is worth discussing and expect nobody else to raise. Read the list together with no names and no judgment, and pick one to look at properly.
AI speeds up the process you already have, including the parts that don't work.
“Can AI fix our onboarding?”
“Let's have AI answer the tickets so we don't have to sort out the ticket process.”
“Only Maria knows how month-end works, but maybe AI can learn it.”
AI gets put on top of a process nobody has written down or agreed on. The output is inconsistent, exceptions pile up and people blame the tool. Meanwhile the knowledge that would make it work is still in one or two people's heads.
Fixing a process is slow and sometimes political, and buying AI feels like a way to skip that step. When one person holds all the context, the rest of the team can't use a tool to do that work. AI can't follow a process nobody has written down.
Pick the one process you most want AI to help with and write down how it's done today, with the people who do it. Fix the obvious breaks first, then try AI on the written version.
It looks finished, and someone downstream has to make it right.
“This report reads well, but half the numbers are wrong.”
“I sent it back three times.”
“AI wrote it, so I figured it was fine.”
AI-made work moves between people without anyone checking it. Polished drafts carry wrong facts, missing context or confident filler. The person who receives it spends time fixing it and starts trusting the sender's work a little less.
Companies push people to use AI without setting norms for checking it. Speed shows up where everyone can see it, and the fixing happens out of sight on someone else's desk. The researchers who named this "workslop" point to "indiscriminate organizational mandates" and too little guidance.
Set one rule: whoever sends AI-assisted work has checked it and can explain it. For anything going to a customer, name the person who reviews it before it goes.
Everyone got access. Nobody got time to learn it.
“We rolled out Copilot to everyone in March.”
“I tried it once and it gave me garbage.”
“There's no time to learn new stuff right now.”
Licenses get bought and announced, maybe with a lunch-and-learn. Then people are on their own. A few power users get real value, most use it for an email now and then, and leaders read low use as low interest. Tools arrive without habits.
Training feels optional when time is tight, and AI tools are sold as easy to pick up. Skill with AI grows person by person, through practice on each person's own work, and that takes protected time nobody scheduled.
Give people one protected hour a week for a month to use AI on their own work, and pair each person who's stuck with someone further along.
People use AI every day and don't say so.
“I use it, but I don't advertise it.”
“If they knew AI did half of this, they'd wonder what I do all day.”
“Is it even allowed?”
AI use is common and hidden. People don't share what works, don't ask about data rules and don't flag problems. Leadership underestimates how much AI is in use, which means it misjudges both the risk and the opportunity.
People worry about looking replaceable or breaking a rule nobody has written down. When the company hasn't stated a position, staying quiet feels safest. Westrum's research describes this kind of information flow, where people share only what feels safe to share, and it's a response to the setup, never a character flaw.
Say the company's position out loud: AI use is welcome, this is what stays out of it, and using AI won't count against anyone. Then ask each team to share one way they use it at the next meeting.
Everyone's using AI, and nobody can say what it changed.
“Usage is up, so it must be working.”
“I'm sure it saves time. I couldn't tell you how much.”
“We'll measure ROI once it's rolled out.”
Success gets reported as logins, seats or prompts. Nobody recorded how long the task took before, so nobody can show what changed. Budget conversations about AI turn into one opinion against another.
Activity is easy to count, and outcomes need a "before" number somebody remembered to take.
For each AI use you care about, write down one "before" number this week, like minutes per task, errors per month or days to close the books. Check it again in 90 days.
Every department head owns AI for their area, and nobody owns how it fits together.
“Each of our department heads is responsible for AI in their area.”
“Sales is trying a tool someone heard about on a podcast.”
“Finance ran a survey, and ops is comparing three tools.”
Each department head owns AI for their own area. They run meetings, send surveys and compare tools, and each team builds its own habits. Because the heads are busy, the plan follows whatever they tried last or heard about on the drive in. There's no thread between departments, and no continuity unless one head happens to be deeply interested.
It feels like the natural split. Department heads know their own work best, and giving each one AI for their area avoids naming one owner over everyone. But each head carries AI as a Second Job, so it gets the time left over, and recent ideas win because there's no time to weigh them against a plan. See The Noise Problem. This is the department heads setup. Recency as the driver is our reasoning. We didn't find a study that measures it for AI decisions.
Ask each department head for their top three AI uses and the tools behind them, and put the lists side by side. Then name one owner for how they fit together, with the department heads as partners.
Every organization is a unique combination of patterns. An assessment identifies which ones are present and builds a roadmap to address them.
An AI strategy decides who owns the work, which tools stay, and what changes first.
How we work: drag6.com/engagementBased in Plymouth, Michigan. We work with companies across Michigan and beyond, on site or remote.