--- title: Why AI Automation Projects Fail (2026 Data + Pre-Mortem Scorecard) description: Most AI automation projects stall before they pay back. What MIT, McKinsey, Gartner and RAND actually measured, six failure modes, and a 10-question pre-mortem. image: https://peppereffect.com/hubfs/blog/images/aifail-hero-20260624-051010.webp --- [Skip Navigation or Skip to Content](https://peppereffect.com/blog/why-ai-projects-fail#skipToMain) [![logo](https://peppereffect.com/hubfs/PeperEffect_Theme_2026/Images/logo.png)](https://peppereffect.com) - [4 Pillars Playbook](https://peppereffect.com/hubfs/lead-magnets/4-pillars-playbook-interactive.html) - [About](https://peppereffect.com/about-ai-consultancy) - [Services](https://peppereffect.com#services) - [Citation Engine](https://peppereffect.com/citation-engine) - [Lead Generation](https://peppereffect.com/lead-generation) - [Sales Administration](https://peppereffect.com/sales-administration) - [Project Management](https://peppereffect.com/project-management-operations) - [Marketing Infrastructure](https://peppereffect.com/marketing-infrastructure) - Resources - [Blog](https://peppereffect.com/blog) - [AI Visibility Check](https://peppereffect.com/hubfs/lead-magnets/ai-visibility-check-landing.html) - [AEO/GEO Readiness Assessment](https://peppereffect.com/hubfs/lead-magnets/aeo-readiness-assessment-landing.html) - [AI Visibility Value Pipeline](https://peppereffect.com/hubfs/lead-magnets/ai-visibility-value-report-landing.html) - [CAC Calculator B2B SaaS](https://peppereffect.com/hubfs/cac-calculator-b2b-saas.html) - [Agency Partner Program](https://peppereffect.com/agency-partner-program) - [Agency comparison](https://peppereffect.com/geo-aeo-agency-comparison) - [B2b Search Traffic Study 2026](https://peppereffect.com/blog/b2b-search-traffic-study-2026) - [Case Studies](https://peppereffect.com#case-study) - [Ranke Consulting](https://peppereffect.com/case-study-ranke-consulting) - [Whitehat](https://peppereffect.com/case-study-whitehat) - [SPRUNG.RAUM](https://peppereffect.com/case-study-sprungraum) - [Helium42](https://peppereffect.com/case-study-helium42) [Contact us](https://peppereffect.com/contact-us) [Contact us](https://peppereffect.com/contact-us) ![Leadership team reviewing an AI project status dashboard with most initiatives flagged red as failing or at risk](https://peppereffect.com/hs-fs/hubfs/blog/images/aifail-hero-20260624-051010.webp?width=1280&height=550&name=aifail-hero-20260624-051010.webp) #### Table of Contents #### Published by ![Picture of Peter Vogel](https://app.hubspot.com/settings/avatar/cd43f2b31a6cdeead12c6b4274a27967) Peter Vogel Founder of peppereffect. Systems architect for B2B founders who want to scale revenue without scaling headcount. Peter has built AI-powered growth operating systems for companies across SaaS, executive search, and high-ticket consulting, engineering the infrastructure that decouples growth from manual effort. [B2B Growth](https://peppereffect.com/blog/tag/b2b-growth) [AI Automation](https://peppereffect.com/blog/tag/ai-automation) 02 Nov. 2026 # Why AI Automation Projects Fail: What the Data Shows and a Pre-Mortem to Run First ## Why do AI automation projects fail? **Mostly for reasons you can see before the build starts: no measurable outcome, no baseline, nobody on the business side who owns the result, data that was never tested on real cases, and a workflow nobody redesigned. The model is rarely the problem. MIT's 2025 study found 60% of organisations evaluated enterprise-grade AI tools and only 5% got them into production, and McKinsey's 2026 survey found just 37% of companies see any EBIT impact from AI.** Both numbers get quoted badly, so this page starts with what each study actually measured. Then it covers the six failure modes that decide a project before the first line of code, a ten-question pre-mortem that scores your own project against them, why agentic projects fail more often, the buy-or-build question, and what the minority that ships does differently. It's written for a COO, CMO or founder at a 20 to 250 person B2B company who's about to sign off an automation budget. 5% Organisations that got enterprise-grade AI tools into production MIT NANDA, Jul 2025 (52 orgs interviewed) 37% Companies attributing any EBIT impact to AI McKinsey, Aug 2026 (n=1,719) 42% Companies that abandoned most AI initiatives before production S&P Global 451, 2025 (n=1,006) 23% Leaders who say accountability is unclear when AI causes a bad outcome Deloitte, Oct 2026 (n=3,500) Key takeaway Most AI projects are lost in the funding meeting. If you can't name the number that should move, the person who owns it, the date you'll stop, and the 50 real cases you tested on, you aren't ready to build. Run the pre-mortem below before you sign. ![Stalled AI automation pilot on a laptop next to a project retrospective listing data quality, scope creep and unclear ownership as red flags](https://peppereffect.com/hs-fs/hubfs/blog/images/aifail-body1-20260624-051009.webp?width=1200&height=675&name=aifail-body1-20260624-051009.webp) ## How often do AI projects actually fail? **It depends on what you call failure, and the famous numbers measure different things. MIT's 95% is the share of organisations getting no measurable P&L return from generative AI. RAND's "more than 80%" is quoted from other estimates. The cleanest reading is S&P Global's: 42% of companies abandoned most of their AI initiatives before production, up from 17% a year earlier.** The MIT figure needs care because it's in every pitch deck. [The GenAI Divide](https://www.marefa.org/w/images/c/c3/MIT_NANDA_v0.1_State_of_AI_in_Business_2025_Report.pdf) (Challapally et al., MIT NANDA, July 2025, PDF copy) rests on a review of 300+ public AI initiatives, interviews with 52 organisations and survey answers from 153 senior leaders collected at four conferences. The authors call their figures "directionally accurate" and based on interviews rather than company reporting. The number I'd take into a budget meeting is the funnel: 60% of organisations evaluated enterprise-grade AI tools, 20% reached a pilot, and 5% reached production. RAND's [root-cause study](https://www.rand.org/pubs/research_reports/RRA2680-1.html) (Ryseff, De Bruhl and Newberry, August 2024) interviewed 65 data scientists and engineers. Its 80% line opens with "by some estimates". Its value is the list of causes, which we use below. Gartner has moved from forecast to observation. In July 2024 it predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. In [January 2026](https://www.gartner.com/en/articles/genai-project-failure) it put the actual figure at at least 50%, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. | Study | What it measured | Figure | Sample | Read it as | | --- | --- | --- | --- | --- | | MIT NANDA, Jul 2025 | Enterprise-grade AI tools evaluated, piloted, in production | 60% / 20% / 5% | 300+ initiatives, 52 orgs, 153 leaders | Directional, interview-based | | S&P Global 451, 2025 | Companies abandoning most AI initiatives before production | 42% (17% a year earlier) | 1,006 respondents | Measured survey result | | Gartner, Jan 2026 | Generative AI projects abandoned after proof of concept by end 2025 | At least 50% | Not disclosed | Analyst observation | | IBM IBV CEO Study, May 2025 | AI initiatives that delivered expected ROI; scaled enterprise-wide | 25% / 16% | 2,000 CEOs | CEO self-report | | McKinsey State of AI, Aug 2026 | Companies with any EBIT impact from AI; high performers | 37% / about 6% | 1,719 respondents | Flat on 2025 (39%) | | RAND, Aug 2024 | AI projects that fail | "More than 80%" | Cited estimate | Not RAND's own measurement | Sources linked in the text and listed at the end of the page. One more pair of numbers explains the pattern. In McKinsey's [2026 survey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), 80% of respondents say AI has improved their individual productivity, while 37% see any effect on EBIT. People get faster. The P&L doesn't move. Closing that gap is a scoping and operations job, which is what the rest of this page is about. ![Bar chart of AI project outcomes: MIT NANDA shows 60% of organisations evaluated enterprise-grade AI tools, 20% piloted and 5% reached production; McKinsey 2026 shows 37% with any EBIT impact and 6% high performers; IBM shows 25% of AI initiatives delivered expected ROI and 16% scaled](https://peppereffect.com/hs-fs/hubfs/blog/images/aifail-data-chart-20261102.webp?width=1200&height=675&name=aifail-data-chart-20261102.webp) ## What makes AI automation projects fail? ![Operations team comparing messy source data with cleaned, AI-ready data on two monitors before an automation build](https://peppereffect.com/hs-fs/hubfs/blog/images/aifail-floatleft-20260624-051019.webp?width=600&height=600&name=aifail-floatleft-20260624-051019.webp) **Six failure modes, and five of them are fixed before the build: an outcome nobody can measure, no baseline, no business owner, untested data and an unchanged workflow. The sixth, running cost per outcome, shows up after launch, but you can predict it from the first two weeks of a pilot.** RAND's interviews produced five root causes: misunderstanding the problem to be solved, not having enough data, chasing new technology instead of a real user problem, missing infrastructure for data and deployment, and applying AI to problems too hard for it. Every one of those is a scoping failure. Here's how they look inside a mid-sized B2B company, with the test that catches each one before kickoff. | Failure mode | How it shows up | Evidence | Test before kickoff | | --- | --- | --- | --- | | No measurable outcome | The brief says "use AI in support". Six months later nobody can say if it worked, so it's never killed and never scaled. | RAND's first root cause; "unclear business value" in Gartner's list | One sentence: which number, by how much, by what date | | No baseline | The pilot "saves time", but nobody timed the old process, so the saving is a guess the CFO won't accept. | IBM: CEOs say 25% of AI initiatives delivered expected ROI | Time, cost and error rate of today's process, on real volume | | No business owner | IT builds it, the team that should use it never adopts it, and when it misfires nobody is accountable. | Deloitte, Oct 2026: 23% say accountability is unclear when AI causes a bad outcome | A named line manager who runs it after launch and whose target it moves | | Untested data | The demo runs on 20 clean examples. Production meets half-filled CRM fields and PDFs in six formats. | Gartner: 63% lack, or aren't sure they have, AI-ready data practices. Informatica: 43% of data leaders name data readiness a top obstacle | Run the logic on 50 real, messy cases before funding the build | | Unchanged workflow | AI is bolted onto the old process, so people do the old work plus checking the AI. | McKinsey 2026: nearly three quarters of high performers redesigned workflows, against a quarter of others | Draw the new workflow, including who approves what | | Unpriced running cost | The build fits the budget. The API calls, human review and exception handling don't. | Deloitte, Oct 2026: 71% track total AI inference and API spend, 44% track cost per unit of inference | Cost per accepted output, review time included, below the value it creates | The "test before kickoff" column is our judgement. Data sources: [Gartner, Feb 2025](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk) (1,203 data management leaders), [Informatica CDO Insights 2025](https://www.informatica.com/about-us/news/news-releases/2025/01/20250128-global-data-leaders-seek-to-harness-the-power-of-genai-for-ai-driven-success.html) (600 data leaders at firms above $500M revenue), [Deloitte, Oct 2026](https://www.deloitte.com/us/en/about/press-room/enterprise-ai-enters-an-roi-reckoning.html). Gartner puts a number on the data risk: through 2026, organisations will abandon 60% of AI projects that aren't supported by AI-ready data. That's a forecast, but it matches what we see. The pilot dataset is almost always cleaner than the real one, because someone picked it. Two patterns sit underneath the table. The first is budget skew. MIT found about half of generative AI budgets go to sales and marketing, while some of the largest savings it documented came from back-office work: lower BPO spend and less external agency use. The visible use case gets the money. The valuable one often doesn't. If you're choosing between a customer-facing chatbot and invoice matching, the second is usually where the [return on the automation](https://peppereffect.com/blog/ai-roi-b2b) sits. The second is incentives, and no survey measures it well, so treat this as our observation from projects rather than a statistic. The team that builds is rewarded for shipping. The team that uses it is measured on something else. Nobody is paid to make it stick, so it doesn't. Want the scope checked before you commit budget? A workflow audit maps the process and prices each handoff before anything gets built. [See how a workflow audit works](https://peppereffect.com/blog/workflow-audit) ## How do you run a pre-mortem on an AI project? **Assume the project has already failed and list why. Gary Klein described the method in Harvard Business Review in 2007. The scorecard below turns the six failure modes into ten questions. Three are hard stops: if you can't name the outcome, the baseline or the business owner, don't fund the build yet.** The weights are our judgement of how often each gap decides the result, informed by the studies above. Answer honestly. "Partly" means it exists but isn't written down or agreed by the people who'll be held to it. AI project pre-mortem scorecard Ten questions, about three minutes. Nothing leaves your browser. 1\. Can you state the outcome in one sentence, with a number and a date? Yes, written and agreed Partly, we know roughly what we want No 2. Have you measured today's process: time, cost and error rate? Yes, on real volume Partly, we have estimates No 3. Is there a named business owner (not IT) who runs it after launch and whose target it moves? Yes Partly, a sponsor but no day-to-day owner No 4. Is there a written kill criterion: what result, by what date, stops the funding? Yes Partly, a review date but no threshold No 5. Have you run the logic on at least 50 real, messy cases? Yes Partly, on a cleaned sample No 6. Is the new workflow drawn, including where a person approves or takes over? Yes Partly, the happy path only No 7. Is the scope one workflow and one team? Yes Two or three workflows No, it's a platform or a company-wide rollout 8. Are the connections to your systems of record (CRM, ERP, helpdesk) scoped and costed? Yes Partly, listed but not costed No 9. Do you know the running cost per accepted output, human review included? Yes Partly, licence and API cost only No 10. Buy, partner or build: is the choice deliberate? Buying or partnering, or building because the workflow is our edge and someone owns maintenance Building because we always build Not decided Does the system act without a human approving it (sends, books, pays, updates records)? No, a person approves every action Yes, within written limits and with an audit log Yes, limits not defined yet Answer the questions to see your readiness. How to read the result. Below 55, treat the next few weeks as scoping and keep the build budget closed. Between 55 and 79, fix the listed gaps first; most take days of work. At 80 or above with no hard stop, fund a pilot with a fixed end date and the kill criterion from question 4. And if the scorecard flags an autonomous system without written limits, that goes to the top of the list regardless of the score. ## Why do agentic AI projects fail more often? ![Business owner and operations lead agreeing an automation plan with defined KPIs and sign-off before an AI agent goes live](https://peppereffect.com/hs-fs/hubfs/blog/images/aifail-floatright-20260624-051022.webp?width=600&height=600&name=aifail-floatright-20260624-051022.webp) **Because every step an agent takes can go wrong, and errors carry forward. On Carnegie Mellon's TheAgentCompany benchmark, the best agent tested completed 30.3% of realistic office tasks on its own. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and Deloitte found 21% of companies have mature governance for autonomous agents.** [TheAgentCompany](https://arxiv.org/abs/2412.14161) simulates a small software company with 175 tasks across engineering, project management, HR, finance and admin. The strongest setup (OpenHands with Gemini 2.5 Pro) fully completed 30.3% and scored 39.3% with partial credit. Agents got stuck in loops, lost track of delegated subtasks and sometimes faked completion with shortcuts. Your workflow will differ from a benchmark. It's still the right starting assumption: an agent that runs ten steps end to end needs every one of them to hold. Adding agents makes it worse unless something checks the work. A Google Research and MIT study of 180 agent configurations ([Kim et al., December 2025](https://arxiv.org/abs/2512.08296)) found independent multi-agent setups amplified errors 17.2 times against a single agent. Central coordination cut that to 4.4 times, which is still worse than one agent. Adding agents only pays when something checks the work between steps. Governance trails deployment. Deloitte's January 2026 survey of 3,235 leaders found 21% had a mature governance model for autonomous agents. Its October 2026 follow-up (3,500 leaders) found 40% of enterprises already let AI execute decisions without human approval, at least to a moderate degree. Gartner's [June 2025 forecast](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) names the reasons projects get cancelled: escalating costs, unclear business value and inadequate risk controls. Don't buy an agent like a software licence An agent dropped into an undefined process, with no approval rules and no owner, bills you for every call while it makes mistakes nobody is watching. Map the workflow first, decide which steps need a human sign-off, then put the agent into it. Our guide to [agent handoff protocols](https://peppereffect.com/blog/agent-handoff-protocols) covers where those approval points go, and [agentic workflows](https://peppereffect.com/blog/agentic-workflows) covers the architecture. ## Should you buy or build AI automation? **For most 20 to 250 person B2B firms, buy or partner. In MIT's interview sample, external partnerships with customised tools reached deployment about 67% of the time, against roughly 33% for internal builds. Build only when the workflow itself is what makes you different and someone owns it after launch.** Read that figure with its caveats, because the report states them. It comes from 52 organisations, it's self-reported, and the authors warn the gap may reflect organisational capability rather than the sourcing choice. Companies that buy well may simply be better at scoping. Still, the direction matches what we see. An internal build has to solve integration, monitoring and maintenance with people who also have day jobs. A specialist has done the integration before. Buying isn't free either. You take on lock-in and a vendor roadmap that isn't yours. Our rule: keep the parts you'd regret outsourcing (the workflow design, the data, and the decision about what the system may do on its own) and rent the rest. The numbers behind that trade-off are in [agency versus in-house: the 3-year cost of ownership](https://peppereffect.com/blog/ai-automation-agency-vs-in-house) and [what it costs to build an AI agent](https://peppereffect.com/blog/cost-to-build-an-ai-agent). ![Team mapping an AI automation workflow on a glass wall with a human-in-the-loop approval checkpoint between AI processing and final output](https://peppereffect.com/hs-fs/hubfs/blog/images/aifail-body2-20260624-051015.webp?width=1200&height=675&name=aifail-body2-20260624-051015.webp) ## What do successful AI projects do differently? **They work in a fixed order: outcome, data, workflow, then the tool, then measurement. They keep scope to one workflow, put a line manager in charge and decide in advance what result stops the project. MIT's top-performing mid-market firms went from pilot to full implementation in about 90 days; large enterprises took nine months or longer.** 1 Write the outcome in one sentence "Cut first response on inbound support tickets from six hours to 30 minutes by 31 March." If you can't fill in the number and the date, stop here. Everything after this step is measured against that sentence. 2 Prove the data on real cases Pull 50 recent, real examples, including the ugly ones, and run the logic by hand or with a rough prototype. If more than a handful break, fix the inputs before anyone builds anything. 3 Redesign the workflow Draw who does what after launch, where the system hands over to a person, and who approves anything that touches a customer or money. This is the step most projects skip, and McKinsey's data says it's the one that separates high performers. 4 Pick the tool last With the outcome, data and workflow fixed, the tool choice gets much easier, and vendor demos stop deciding the scope. Buy or partner unless the workflow is your edge. 5 Measure weekly against the baseline Track the outcome number, cost per accepted output and the share of outputs a person had to fix. Scale when the kill criterion is beaten. Stop when it isn't. BCG's [2025 study of 1,250 companies](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap) supports the order. Only 5% of the sample qualified as "future-built", and BCG's rule for them is a 10-20-70 split of effort: 10% on algorithms, 20% on technology and data, 70% on people and process. More than 60% of future-built firms rigorously track AI value, against 17% of stagnating companies. Across the full sample, 60% report little or no value from AI. Narrow scope is the part people resist, because one workflow feels small next to the board's AI ambitions. It's also how the 90-day firms in MIT's sample got there. One process that provably pays funds the next one, and that's how automation starts to [decouple revenue from headcount](https://peppereffect.com/blog/decouple-revenue-from-headcount). For the build sequence across several processes, see [AI workflow automation](https://peppereffect.com/blog/ai-workflow-automation). ## How do you know if an AI automation project is working? **Compare it with the baseline you measured before the build, on three numbers: the outcome metric you named, the fully loaded cost per accepted output, and the share of outputs a person had to correct. Logins and usage counts don't tell you anything about value. If the outcome hasn't moved by the kill date, stop.** Cost per accepted output is the number most teams don't have. It includes model and API calls, the human review minutes, rework and the exceptions that fall back to a person. Deloitte's October 2026 survey found 71% of enterprises track total AI inference and API spend, but only 44% track cost per unit of inference. Without a unit cost, you can't tell whether one output pays for itself. Use the leading indicators weekly (accuracy on live cases, correction rate, throughput) and the lagging ones monthly (the outcome metric, cost per output against the baseline). The first set lets you correct course. The second set is what finance will accept. Our [automate versus hire](https://peppereffect.com/blog/automate-vs-hire) comparison shows how to set that unit cost against the cost of a person doing the same work. Find out if the project is fundable before you fund it peppereffect runs the pre-mortem with you: we map the workflow, test your data on real cases, set the baseline and the kill criterion, then build the automation in the order your data can support. If the answer is "not yet", you hear it before you spend. [Book a Growth Mapping Call](https://peppereffect.com/growth-mapping-call) [See our AI automation services](https://peppereffect.com/blog/ai-automation-services) ## Frequently asked questions ### What percentage of AI projects fail? It depends on the definition. S&P Global found 42% of companies abandoned most of their AI initiatives before production in 2025, up from 17%. Gartner reported that at least 50% of generative AI projects were abandoned after proof of concept by the end of 2025. MIT's widely quoted 95% is the share of organisations with no measurable P&L return from generative AI, based on interviews with 52 organisations, and its authors call it directional. RAND's "more than 80%" is cited from other estimates. ### Why do AI automation projects fail? Mostly poor scoping: no measurable outcome, no baseline, no business owner, data never tested on real cases, an unchanged workflow and running costs nobody priced. RAND's interviews with 65 practitioners put misunderstanding the problem to be solved first. Gartner lists poor data quality, inadequate risk controls, escalating costs and unclear business value. ### Why do agentic AI projects fail more often? Errors compound across steps. On Carnegie Mellon's TheAgentCompany benchmark the best agent completed 30.3% of realistic office tasks, and a 2025 Google Research study found independent multi-agent setups amplified errors 17.2 times. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and Deloitte found 21% of companies have mature governance for autonomous agents. ### Is it better to buy or build AI automation? For most mid-sized B2B companies, buy or partner. MIT found external partnerships reached deployment about 67% of the time against roughly 33% for internal builds, though the sample was 52 organisations and self-reported. Build when the workflow is your competitive edge and someone owns maintenance after launch. ### How do you de-risk an AI automation project? Run a pre-mortem before funding. Write the outcome as one sentence with a number and a date, measure today's process, name a business owner, write a kill criterion, test the logic on 50 real cases, and redraw the workflow with human approval points. Keep the first project to one workflow and one team. ### How do you measure if an AI project is succeeding? Against the baseline you measured before the build, on three numbers: the outcome metric, the fully loaded cost per accepted output, and the share of outputs a person had to correct. Review the leading indicators weekly and the outcome monthly. If the kill criterion isn't beaten by the agreed date, stop or rescope. ## Sources MIT NANDA, Challapally, Pease, Raskar and Chari, [The GenAI Divide: State of AI in Business 2025](https://www.marefa.org/w/images/c/c3/MIT_NANDA_v0.1_State_of_AI_in_Business_2025_Report.pdf) (July 2025; 300+ initiatives, 52 organisations, 153 leaders). RAND, Ryseff, De Bruhl and Newberry, [The Root Causes of Failure for Artificial Intelligence Projects](https://www.rand.org/pubs/research_reports/RRA2680-1.html) (August 2024, 65 interviews). S&P Global Market Intelligence, [Voice of the Enterprise: AI and Machine Learning, Use Cases 2025](https://www.spglobal.com/market-intelligence/en/news-insights/research/2025/10/generative-ai-shows-rapid-growth-but-yields-mixed-results) (n=1,006). Gartner, [Why 50% of GenAI projects fail](https://www.gartner.com/en/articles/genai-project-failure) (26 January 2026). Gartner, [Lack of AI-ready data puts AI projects at risk](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk) (26 February 2025, n=1,203). Gartner, [Over 40% of agentic AI projects will be cancelled by end 2027](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) (25 June 2025). IBM Institute for Business Value, [2025 CEO Study](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles) (n=2,000). McKinsey, [The state of AI in 2026: On the road to ROI](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) (25 August 2026, n=1,719). Deloitte, [State of AI in the Enterprise 2026](https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html) (January 2026, n=3,235) and [Enterprise AI enters an ROI reckoning](https://www.deloitte.com/us/en/about/press-room/enterprise-ai-enters-an-roi-reckoning.html) (29 October 2026, n=3,500). Informatica, [CDO Insights 2025](https://www.informatica.com/about-us/news/news-releases/2025/01/20250128-global-data-leaders-seek-to-harness-the-power-of-genai-for-ai-driven-success.html) (n=600). BCG, [The Widening AI Value Gap](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap) (September 2025, n=1,250). Xu et al., [TheAgentCompany](https://arxiv.org/abs/2412.14161) (Carnegie Mellon, 2025). Kim et al., [Towards a Science of Scaling Agent Systems](https://arxiv.org/abs/2512.08296) (December 2025). Klein, [Performing a Project Premortem](https://hbr.org/2007/09/performing-a-project-premortem) (Harvard Business Review, September 2007). ## Related blog ![SaaS revenue leader reviewing an automated revenue operations dashboard with leads flowing from marketing to sales to customer success](https://peppereffect.com/hs-fs/hubfs/blog/images/saas-revops-hero-20260630-055625.webp?width=406&height=320&name=saas-revops-hero-20260630-055625.webp) ###### 30 Okt. [Lead Generation](https://peppereffect.com/blog/tag/lead-generation) [AI Automation](https://peppereffect.com/blog/tag/ai-automation) [Sales Automation](https://peppereffect.com/blog/tag/sales-automation) #### AI Automation for SaaS Revenue Operations: What to Automate First [Peter Vogel](https://peppereffect.com/blog/author/peter-vogel) ![Recruiter reviewing an automated recruiting workflow dashboard](https://peppereffect.com/hs-fs/hubfs/blog/images/recauto-hero-20260624-055528.webp?width=406&height=320&name=recauto-hero-20260624-055528.webp) ###### 28 Okt. [B2B Growth](https://peppereffect.com/blog/tag/b2b-growth) [AI Automation](https://peppereffect.com/blog/tag/ai-automation) #### Recruitment Automation in 2026: What to Automate, the Real ROI, and Where to Start [Peter Vogel](https://peppereffect.com/blog/author/peter-vogel) ![Executive team weighing an AI automation agency against an in-house team with a cost comparison on screen](https://peppereffect.com/hs-fs/hubfs/blog/images/ai-automation-agency-vs-inhouse-hero-20260411-094128.webp?width=406&height=320&name=ai-automation-agency-vs-inhouse-hero-20260411-094128.webp) ###### 26 Okt. [Lead Generation](https://peppereffect.com/blog/tag/lead-generation) [Agentic AI](https://peppereffect.com/blog/tag/agentic-ai) [AI Automation](https://peppereffect.com/blog/tag/ai-automation) #### AI Automation Agency vs In-House Team: The 3-Year Cost of Ownership [Peter Vogel](https://peppereffect.com/blog/author/peter-vogel) THE NEXT STEP ## Stop Renting Leverage. Install It. Together we can achieve great things. Send us your request. We will get back to you within 24 hours. [The Audit](https://peppereffect.com/hubfs/lead-magnets/technicians-trap-scorecard.html) [Contact us](https://peppereffect.com/contact-us) ![Group 1000005311-1](https://peppereffect.com/hs-fs/hubfs/PeperEffect_Theme_2026/Images/Group%201000005311-1.png?width=465&height=485&name=Group%201000005311-1.png) [![logo](https://peppereffect.com/hs-fs/hubfs/PeperEffect_Theme_2026/Images/logo.png?width=223&height=38&name=logo.png)](https://peppereffect.com) Growth architects for medium-sized businesses. We help companies become scalable and independent. - ##### Our Resources - [4 Pillars Playbook](https://peppereffect.com/hubfs/lead-magnets/4-pillars-playbook-interactive.html) - [About Us](https://peppereffect.com/about-ai-consultancy) - [Blog](https://peppereffect.com/blog) - [Case Studies](https://peppereffect.com#case-study) - [Contact us](https://peppereffect.com/contact-us) ##### Our Services - [Lead Generation](https://peppereffect.com/lead-generation) - [Sales Administration](https://peppereffect.com/sales-administration) - [Project Management](https://peppereffect.com/project-management-operations) - [Marketing Classics](https://peppereffect.com/marketing-infrastructure) - [Citation Engine](https://peppereffect.com/citation-engine) ##### Contact Info - [+49 163 7096754](tel:+491637096754) - [info@peppereffect.com](mailto:info@peppereffect.com) - Provinzialstrasse 41, 46499 Hamminkeln, Germany - [Imprint](https://peppereffect.com/imprint) - [Privacy Policy](https://peppereffect.com/privacy-policy) © 2026 peppereffect | All Rights Reserved Cookie-Einstellungen ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Peter Vogel", "url" : "https://peppereffect.com/blog/author/peter-vogel" }, "datePublished" : "2026-11-02T04:15:00.000Z", "headline" : "Why AI Automation Projects Fail (2026 Data + Pre-Mortem Scorecard)", "image" : [ "https://peppereffect.com/hubfs/blog/images/aifail-hero-20260624-051010.webp" ], "mainEntityOfPage" : { "@id" : "https://peppereffect.com/blog/why-ai-projects-fail", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://peppereffect.com/hubfs/20201010_peppereffect_Logo.png" }, "name" : "peppereffect" } } ``` ```json { "@context" : "https://schema.org", "@type" : "Article", "author" : { "@type" : "Person", "name" : "Peter Vogel", "sameAs" : [ "https://www.linkedin.com/company/pepper-effect/" ], "url" : "https://peppereffect.com/" }, "dateModified" : "2026-11-02", "datePublished" : "2026-06-24", "description" : "Most AI automation projects stall before they pay back. What MIT, McKinsey, Gartner and RAND actually measured, six failure modes, and a 10-question pre-mortem.", "headline" : "Why AI Automation Projects Fail: What the Data Shows and a Pre-Mortem to Run First", "image" : "https://peppereffect.com/hubfs/blog/images/aifail-hero-20260624-051010.webp", "keywords" : "why AI projects fail, why AI automation projects fail, AI project failure rate, AI pre-mortem, agentic AI failure, buy vs build AI, AI project success factors", "mainEntityOfPage" : { "@id" : "https://peppereffect.com/blog/why-ai-projects-fail", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://peppereffect.com/hubfs/peppereffect-logo.svg" }, "name" : "peppereffect", "url" : "https://peppereffect.com" } } ``` ```json { "@context" : "https://schema.org", "@type" : "FAQPage", "mainEntity" : [ { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "It depends on the definition. S&P Global found 42% of companies abandoned most of their AI initiatives before production in 2025, up from 17%. Gartner reported that at least 50% of generative AI projects were abandoned after proof of concept by the end of 2025. MIT's widely quoted 95% is the share of organisations with no measurable P&L return from generative AI, based on interviews with 52 organisations, and its authors call it directional. RAND's \"more than 80%\" is cited from other estimates." }, "name" : "What percentage of AI projects fail?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Mostly poor scoping: no measurable outcome, no baseline, no business owner, data never tested on real cases, an unchanged workflow and running costs nobody priced. RAND's interviews with 65 practitioners put misunderstanding the problem to be solved first. Gartner lists poor data quality, inadequate risk controls, escalating costs and unclear business value." }, "name" : "Why do AI automation projects fail?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Errors compound across steps. On Carnegie Mellon's TheAgentCompany benchmark the best agent completed 30.3% of realistic office tasks, and a 2025 Google Research study found independent multi-agent setups amplified errors 17.2 times. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and Deloitte found 21% of companies have mature governance for autonomous agents." }, "name" : "Why do agentic AI projects fail more often?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "For most mid-sized B2B companies, buy or partner. MIT found external partnerships reached deployment about 67% of the time against roughly 33% for internal builds, though the sample was 52 organisations and self-reported. Build when the workflow is your competitive edge and someone owns maintenance after launch." }, "name" : "Is it better to buy or build AI automation?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Run a pre-mortem before funding. Write the outcome as one sentence with a number and a date, measure today's process, name a business owner, write a kill criterion, test the logic on 50 real cases, and redraw the workflow with human approval points. Keep the first project to one workflow and one team." }, "name" : "How do you de-risk an AI automation project?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Against the baseline you measured before the build, on three numbers: the outcome metric, the fully loaded cost per accepted output, and the share of outputs a person had to correct. Review the leading indicators weekly and the outcome monthly. If the kill criterion isn't beaten by the agreed date, stop or rescope." }, "name" : "How do you measure if an AI project is succeeding?" } ] } ``` ```json { "@context" : "https://schema.org", "@type" : "BreadcrumbList", "itemListElement" : [ { "@type" : "ListItem", "item" : "https://peppereffect.com", "name" : "Home", "position" : 1 }, { "@type" : "ListItem", "item" : "https://peppereffect.com/blog", "name" : "Blog", "position" : 2 }, { "@type" : "ListItem", "item" : "https://peppereffect.com/blog/why-ai-projects-fail", "name" : "Why AI Automation Projects Fail", "position" : 3 } ] } ```