Why Agentic AI Projects Fail (and Why Gen AI Projects Fail Differently)

Agentic AI projects fail because a tool that used to hand a draft to a person is given permission to act on its own, and nobody has worked out what it may touch, how often it's right across every step of the job, what a wrong action costs, or who answers for it once it's live. An AI agent is software you give a goal rather than a single question: it decides its own steps and carries them out in your other systems, posting an order, sending an email, or updating a record. That's a different kind of project from a chatbot that writes a first draft, and it breaks in different places.

Gartner expects a lot of these projects to break. In June 2025 it predicted that over 40% of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls" (Gartner, June 2025). Our guide to why AI projects fail covers the causes every AI project shares and what the famous failure figures actually count. What changes when the AI stops suggesting and starts doing is where a 10-to-200-person business that has been pitched an "AI agent" needs to look before signing.

Key Takeaways

• Generative AI projects mostly fail quietly: the tool drafts, nobody uses the drafts, and the subscription lapses. Agentic projects fail loudly, because a wrong action reaches a customer, a ledger, or a job site before anyone sees it.
• Errors multiply across steps. If each step of a 20-step task is right 95% of the time and nothing catches mistakes, the whole task comes out right only about 36% of the time.
• In 2025 benchmark testing, the best agent completed 30% of simulated office tasks on its own (Carnegie Mellon), and leading agents fell from about 58% success on single-turn business tasks to about 35% when the task took several exchanges (Salesforce).
• Gartner predicts 40% of enterprises will demote or shut down autonomous agents by 2027 because of governance gaps found only after production incidents, and it recommends matching controls to four levels of autonomy.
• A business is responsible for what its AI tells customers. A Canadian tribunal ordered Air Canada to pay after its chatbot invented a refund policy, calling the airline's defense "a remarkable submission."
• The projects that survive start with the agent preparing work for a person to approve, give it access only to what that job needs, and widen its authority one task type at a time as the error record earns it.

Table of Contents

• Why do agentic AI projects fail? The short answer
• Gen AI drafts, agents act: why the failures look different
• Why small error rates sink multi-step agents
• Seven reasons agentic AI projects fail in production
• Match the agent's autonomy to the work
• A 90-person supply distributor's order-entry agent
• Eight questions to answer before you approve an agent
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

Why Do Agentic AI Projects Fail? The Short Answer

Agentic AI projects fail when an agent is given more steps, more system access, and more independence than its track record supports, usually because a demo on clean examples made the job look solved. The agent then meets real work, which is messier than the demo, and its mistakes become actions instead of drafts. Gartner's own analyst described most agentic projects in 2025 as "early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied" (Gartner, June 2025).

The failures cluster in a handful of places: work that never needed an agent, too many steps handed over at once, access far wider than the task, an approval step that decays into a rubber stamp, no plan for the cases the agent can't handle, a running cost nobody modeled, and no one inside the business who owns the agent after the people who built it leave. Each one is cheaper to settle before the build than after the first incident.

Gen AI Drafts, Agents Act: Why the Failures Look Different

Generative AI projects and agentic AI projects fail for overlapping reasons, but they fail in different ways. A generative AI tool writes, summarizes, or answers, and a person decides what to do with the result. An agent takes the next step itself. That one difference changes who notices a mistake, how much it costs, and what fixes it.

Generative AI's typical failure is that nobody ends up relying on it. MIT's 2025 study of enterprise AI found that workers prefer AI for quick work (70% preferred it for drafting emails and 65% for basic analysis), but for complex, multi-week work and client management, people preferred a human by roughly nine to one. MIT put the dividing line down to memory and learning: the tools don't retain feedback or improve on the business's own work (Challapally, Pease, Raskar, and Chari, The GenAI Divide, MIT NANDA, July 2025, pp. 12–13). That's why so many gen AI projects end with a tool people log into for a month and then stop opening.

Agents are often sold as the cure for exactly that, and MIT's report itself points to agentic systems with persistent memory as the way across the gap (p. 14). The catch is that an agent turns a quiet failure into a loud one.

If your project is a drafting tool that nobody uses, our page on why AI projects are failing now covers that pattern in more detail.

Why Small Error Rates Sink Multi-Step Agents

An agent that's right most of the time on each step can still be wrong most of the time on the whole job. The reason is multiplication. If an agent handles a 20-step task and each step is right 95% of the time, with nothing checking its work along the way, the chance that all 20 steps come out right is 0.95 multiplied by itself 20 times, or about 36%. At 99% per step, the same task succeeds about 82% of the time, and at 90% per step it succeeds about 12% of the time.

Real tasks don't behave exactly like this, since some steps are easy and some mistakes get caught, but the published benchmarks point the same way. Researchers at Carnegie Mellon built TheAgentCompany, a simulated small software company with internal websites, files, and coworkers to message, and found that the most capable agent they tested could complete 30% of its tasks autonomously, with longer, harder tasks "still beyond the reach of current systems" (Xu et al., TheAgentCompany, arXiv:2412.14161, revised September 2025). Salesforce's AI research team tested agents on 19 sales, service, and quoting tasks and found leading agents succeeded about 58% of the time when the task came in a single request, dropping to about 35% when it took several back-and-forth exchanges (Huang et al., CRMArena-Pro, arXiv:2505.18878, May 2025).

Models have improved since these tests ran, and they'll keep improving. The multiplication doesn't change, though. Better models raise the per-step number, but every extra step an agent takes without a check still cuts the odds that the whole job comes out right. The practical conclusion is about design rather than waiting for a smarter model: keep the chain of unchecked steps short, put a check where a mistake would be expensive, and pick work where the agent's job is narrow. Salesforce's own results show the upside of that, since workflow execution, the task type the researchers found most tractable, scored above 83%.

Seven Reasons Agentic AI Projects Fail in Production

Most agentic projects that die in production die for one of seven reasons. They overlap, and a project in trouble usually has two or three of them at once.

1. The job never needed an agent

A lot of work sold as agentic is a fixed sequence of steps that ordinary automation handles more cheaply and more predictably. Gartner's analyst put it bluntly in 2025: "Many use cases positioned as agentic today don't require agentic implementations." Gartner's suggested split is useful for any business: agents when decisions are needed, automation for routine workflows, and assistants for simple retrieval (Gartner, June 2025).

The market makes this harder to see. Gartner estimated that only about 130 of the thousands of vendors selling agentic AI are real. Many of the others are "agent washing," Gartner said: chatbots, assistants, and older robotic process automation relabeled as agents. If a step-by-step rule would do the job, paying for an agent adds cost and unpredictability without adding anything the work needs.

2. Too many steps were handed over at once

The demo shows an agent running a whole process end to end: read the request, look up the customer, check stock, build the order, confirm it, update the books. Each step works in the demo. In production, the compounding problem above takes over, and a process that's 95% right at each of a dozen steps produces a steady stream of wrong outcomes that someone has to find and unwind.

Projects that survive hand the agent one or two steps first, usually the ones that eat the most time, and leave the rest with people until the error record on those steps is good enough to extend.

3. The agent can touch more than the task needs

The more an agent can reach, the more it can break on a bad day. The OWASP Foundation's list of top security risks for AI applications names this risk "Excessive Agency" and traces it to three root causes: excessive functionality, excessive permissions, and excessive autonomy. One of its examples is a tool meant only to read records that connects with an account allowed to update and delete them too (OWASP, LLM06:2025 Excessive Agency).

One widely reported case involved a single founder's test project. In July 2025, SaaStr founder Jason Lemkin was testing Replit's AI coding agent when, on day nine of the experiment and during an explicit freeze on changes, it deleted his live database, wiping records for 1,206 executives and more than 1,196 companies. Replit's chief executive called the behavior "unacceptable" and the company began separating development and production databases automatically, which is a stronger safeguard than any instruction to leave the data alone (Tom's Hardware, July 2025).

4. The approval step turns into a rubber stamp

The usual safeguard is to have a person approve each action before the agent takes it. That works for a few weeks. Then the agent gets most things right, the queue grows, and the person approving starts clicking through. Gartner warned in 2026 that at this level of autonomy, "human review is effective only if it remains a meaningful control," and that without clear approval workflows and audit trails, approvals "can degrade under time pressure or approval fatigue, creating a false sense of safety" (Gartner, May 2026).

An approval step holds up when the approver sees the source material next to the agent's proposed action, the volume is something a person can actually check, and the agent flags the cases it's unsure about so attention goes where it's needed.

5. Nobody planned for the cases the agent can't handle

Every real process has cases nobody wrote down: the customer who orders by describing a part instead of naming it, the invoice with two purchase order numbers, the request that needs a manager's judgment. People handle these without thinking. An agent either stops, asks, or guesses, and the projects that fail are the ones where nobody decided in advance which of those it should do.

The agent needs a defined way out: a rule for when it hands a case to a person, a place where those cases land, and someone who works that queue every day. Without that, exceptions either pile up unseen or get guessed at, and both end with staff doing the work by hand again "just to be safe."

6. The running cost was never modeled

A chatbot answers once. An agent plans, calls tools, checks its work, and retries, and each of those steps is billed. Gartner's Will Sommer put the difference this way in August 2026: "Where a simple chatbot must read and interpret a query and quickly respond with a probabilistically reasonable answer, an AI agent must constantly reason, negotiate and question itself." Gartner projects AI inference costs (the cost of running the model each time it does something) to rise more than fivefold through 2028 as businesses move to multistep processes (Gartner via CIO Dive, August 2026).

For a small business, the risk is a build quote that covers the setup and says little about the monthly bill. Ask for the expected cost per completed task at your real volume, including retries and the messy cases, and compare it with what the task costs now.

7. Nobody inside the business owns it after the build

Agents need tending. Prices, products, policies, and customers change, and an agent built around last quarter's rules starts making last quarter's decisions. If the people who built it leave and nobody inside the business can adjust it, it either drifts or gets switched off. Gartner predicted in September 2026 that by 2028, 70% of enterprises will abandon agentic AI built by vendor engineers embedded with them, "trapped by soaring costs and unable to evolve it on their own," and it advised setting knowledge transfer, ownership, and an exit plan from the first day of the engagement (Gartner, September 2026).

Ownership also has a legal side. When Air Canada's website chatbot told a grieving customer he could claim a bereavement fare after travel, which wasn't the airline's policy, the airline argued it couldn't be held liable for information from its chatbot. British Columbia's Civil Resolution Tribunal called that "a remarkable submission," found that Air Canada "is responsible for all the information on its website," and ordered it to pay $812.02 (Moffatt v. Air Canada, 2024 BCCRT 149, February 2024). The same reasoning applies to any business whose agent talks to customers or acts in their name.

Match the Agent's Autonomy to the Work

The most useful decision in an agentic project is how much independence the agent gets for each task, and the most common mistake is treating that as a yes-or-no question. Gartner's 2026 analysis says enterprises are treating agent governance "as binary, either locked down or fully trusted, and that is the root cause of failure." It predicts that by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps found only after production incidents, and it describes four autonomy levels, each needing different controls (Gartner, May 2026).

Here's what each level looks like in a 10–200 person operation, with the conditions we'd want met before moving a task up.

The level belongs to a task rather than to the agent: the same agent can act on its own for repeat orders from known customers and only advise on a first order from a new one. Movement also goes both ways. An agent that starts making mistakes after a price change or a new product line should drop a level until it's fixed, which is routine maintenance for any agent in production.

Most of the hours a smaller business can recover sit at levels 2 and 3. The agent does the reading, matching, and assembling, which is where the time goes, and a person makes the call that carries the risk.

A 90-Person Supply Distributor's Order-Entry Agent

Consider a 90-person plumbing and HVAC supply distributor with four branches (an illustrative composite). Contractors send roughly 250 purchase orders a day by email, as PDFs, photos of handwritten lists, and plain text, and eight inside sales reps key them into the ERP at about six minutes each. That's around 25 hours of keying a day, and it's the work that keeps reps from calling back customers.

The first attempt was the full end-to-end agent from a vendor demo. It read each email, matched the lines to catalog items, posted the order in the ERP, and emailed the contractor a confirmation, all without a person in the loop. On clean, typed orders it was nearly flawless. On the rest, it guessed. A line reading "same fittings as the Henderson job" became a similar part number, and a case of the wrong valves went out on a truck to a job site. Within a month, the inside sales manager had reps checking every posted order after the fact, which took longer than keying them had. The usage bill ran well above the quote too, because the agent retried the messiest emails several times before giving up. The company paused it in month three.

The second attempt kept the same underlying technology and changed nearly everything around it. The agent now works at the advise level: it reads the email, matches each line against the catalog and that customer's purchase history, builds an unposted draft order in the ERP, and marks each line as matched or uncertain. A rep reviews the draft next to the original email, fixes the uncertain lines, and posts it. The agent can read the catalog and customer history and can write draft orders, and that's all; it can't post, change prices, or email anyone. Uncertain lines and anything it can't read go to a shared queue that one rep owns each shift. The inside sales manager owns the agent itself, reviews match accuracy and minutes per order every Friday, and has a short list of things that send it back a level, starting with any price-list update.

After ten weeks, review time per order had settled well under the old keying time, and the manager moved one narrow slice up to act-with-approval: repeat orders from the 40 longest-standing accounts, where every line matched the customer's own history, which the agent now posts once a rep clicks approve. After another three months of clean results on that slice, those orders post on their own, with a daily sample audit by the manager. Everything else still goes through a rep.

The second version changed four things: how many steps the agent took unchecked, what it could reach, where its uncertain cases went, and who owned it.

Eight Questions to Answer Before You Approve an Agent

These questions settle most of the seven failure reasons before any money moves. If a vendor or partner can't answer them in writing, the project isn't ready to build.

1. Does this job need judgment, or is it a fixed sequence? If it's a fixed sequence, ordinary automation is likely cheaper and more predictable.
2. Which steps will the agent take, and which stay with people at the start? Fewer is better for the first release.
3. What level of autonomy will each task run at on day one? It should usually be advise or act-with-approval, with the conditions for moving up written down.
4. Exactly what systems and actions can it reach? List the permissions, and remove anything the task doesn't need.
5. What happens when it's unsure? Name the handoff rule, where those cases land, and who works them.
6. What does one completed task cost to run at our volume? Include retries and messy cases, and compare it with today's cost.
7. How will we know it's working? Record today's baseline (time per task, error rate) before go-live, and set a regular review.
8. Who inside the business owns it, and what do they need to know to change it? Make knowledge transfer and access to the build part of the agreement.

Large organizations face these questions with extra layers of committees and integration on top, which our page on why most enterprise AI projects fail covers. If the agent will touch patient, client, or credit data, the stakes on questions 4 and 7 rise sharply, and our page on why AI projects fail in regulated industries explains what the rules require.

Where Avolis Fits

Agentic projects fail mostly on scope and sequence, and those are what our diagnostic is designed to settle. Every engagement starts with it. Working through each workflow with the people who run it, we measure how it runs (volumes, time per task, where the work changes hands and where it stalls), and we rank the workflows by what each would return against what it would take to build. For agentic candidates, we also mark which steps need judgment and which are fixed sequences, and what level of autonomy each step can safely start at. The ranked result is yours to keep, whatever you do next.

From there, we build the first workflow inside the systems your team already uses, usually with the agent preparing the work and your people approving it, and with permissions, the exception queue, and logging built in from the start. We stay embedded while it settles and while the next one is built. Because the agent lives in your own systems and the person who owns it is trained to change its rules, permissions, and autonomy levels, it doesn't depend on us to keep up with your prices, products, and customers. Our guide to AI readiness assessment services explains what a good diagnostic produces, and our guide to which AI consulting company to choose covers how to evaluate any partner proposing an agent.

Frequently Asked Questions

Why do agentic AI projects fail?

Agentic AI projects usually fail because the agent is given too many steps, too much system access, and too much independence before its error record supports it. Small per-step error rates compound across a multi-step task, approvals decay into rubber stamps, and nobody owns the agent after launch. Gartner predicts over 40% of agentic projects will be canceled by the end of 2027.

Why do generative AI projects fail?

Generative AI projects most often fail quietly: the tool produces drafts, but people stop using it because it doesn't fit their work or learn from their corrections. MIT's 2025 research found workers preferred AI for quick tasks such as drafting emails but preferred humans roughly nine to one for complex, multi-week work, citing tools that don't retain feedback.

Why do agentic AI projects fail in production when the pilot worked?

Pilots usually run on clean examples with someone watching closely. Production brings messy inputs, cases nobody planned for, higher volumes, and running costs billed per step. An agent that looked flawless on twenty demo orders meets the ambiguous ones, guesses, and acts on the guess. Narrow scope, an exception queue, and approval steps close most of that gap.

What percentage of agentic AI projects fail?

There's no measured failure rate yet. The most cited figure is a forecast: Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. In May 2026 it predicted 40% of enterprises will demote or decommission autonomous agents by 2027.

Should a small business use AI agents?

Yes, for the right tasks and at the right level of autonomy. Start where the agent reads, matches, and prepares work that a person approves, such as draft orders or invoice matching, since that's where most of the time goes. Let it act on its own only for narrow task types after months of clean results.

Continue Learning

Agents fail at the points where a draft becomes an action: too many unchecked steps, too much access, an approval nobody really performs, and no owner once the build is done. All four are decisions you can make before the project starts.

Before you sign for an agent, decide which single task it will do first, at what level of autonomy, and who in your business answers for it.

Why projects fail, in detail:

• Why do AI projects fail? Common reasons and statistics
• Why AI projects fail in regulated industries
• Why do most enterprise AI projects fail?
• Why are AI projects failing? What's different about the current wave

When a project is already in trouble:

• Why your AI consulting project isn't reaching production
• How to recover a failing AI project

Starting the right way:

• AI readiness assessment services
• Which AI consulting company should I choose?

---

Sources

• Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," 25 June 2025 (includes the estimate of about 130 genuine agentic AI vendors), retrieved 2026-10-26: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
• Gartner, "Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure," 26 May 2026, retrieved 2026-10-26: https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure
• Gartner, "Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028," 29 September 2026, retrieved 2026-10-26: https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028
• CIO Dive, "Advances in AI capabilities to outpace cost savings," 17 August 2026, reporting Gartner research by Will Sommer, retrieved 2026-10-26: https://www.ciodive.com/news/ai-systems-outpace-cost-savings/828068/
• Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025, pp. 12–14, retrieved 2026-10-26: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
• Frank F. Xu et al., "TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks," Carnegie Mellon University, arXiv:2412.14161, first submitted 18 December 2024, revised 10 September 2025, retrieved 2026-10-26: https://arxiv.org/abs/2412.14161
• Kung-Hsiang Huang et al., "CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions," Salesforce AI Research, arXiv:2505.18878, 24 May 2025, retrieved 2026-10-26: https://arxiv.org/abs/2505.18878
• OWASP Foundation, "LLM06:2025 Excessive Agency," OWASP Top 10 for LLM Applications 2025, retrieved 2026-10-26: https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
• Tom's Hardware, "AI coding platform goes rogue during code freeze and deletes entire company database," July 2025, on Jason Lemkin's Replit test, retrieved 2026-10-26: https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-coding-platform-goes-rogue-during-code-freeze-and-deletes-entire-company-database-replit-ceo-apologizes-after-ai-engine-says-it-made-a-catastrophic-error-in-judgment-and-destroyed-all-production-data
• Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, decided 14 February 2024 (Tribunal Member Christopher C. Rivers; $650.88 in damages, $36.14 in pre-judgment interest, and $125 in fees), retrieved 2026-10-26: https://decisions.civilresolutionbc.ca/crt/crtd/en/item/525448/index.do

On the compounding calculation. The 36%, 82%, and 12% figures are arithmetic, not measurements: per-step accuracy raised to the power of 20, treating each step as independent and assuming no step checks another. Real agents can catch and correct some of their own errors, and some steps are much easier than others, so the curve is a way to see the effect of chaining unchecked steps rather than a prediction for any particular agent.

On the benchmarks. Both benchmarks tested models available in 2024 and 2025, and newer models score differently. TheAgentCompany's 30% is the share of tasks the best-performing agent completed fully and autonomously in a simulated software company. CRMArena-Pro's 58% and 35% are approximate success rates for the leading agents tested on synthetic data in a Salesforce environment; the 83% figure covers workflow execution tasks in a single turn only.

On the Gartner figures. The 40% cancellation figure and both 2026 figures are predictions, not measured outcomes. The 2026 governance and forward-deployed engineering predictions refer to enterprises.

On the Replit incident. The account of the agent's behavior comes from Jason Lemkin's own posts and the chat logs he published, as reported by Tom's Hardware, along with the response from Replit's chief executive. We found no independent forensic report.

On the distributor. It's an illustrative composite, not a client. Its figures (90 people, four branches, about 250 orders a day, eight reps, six minutes an order, three months, ten weeks, 40 accounts, three further months) are representative examples, not measurements or benchmarks.

---

About Avolis Research Group

Avolis Research Group is Avolis's in-house research practice, focused on how operations-heavy small and mid-sized businesses actually adopt AI. It synthesizes primary economic research, government survey data, and results from real implementations into practical, vendor-neutral guidance.

More about Avolis and how we work · Get in touch