The pre-deployment AI audit: why 80% of projects fail and what it looks like done right

Contents

1. What a pre-deployment AI audit is, a definition, how it differs from an IT audit and from strategy consulting
2. Why 80% of AI projects fail, data from 6 independent studies with links to the originals
3. What a thorough audit covers, the four pillars of the YESFOR Framework
4. What the process looks like step by step, 6 phases, 4-6 weeks
5. How much an AI audit costs, price ranges, comparison with Big4
6. When you need an audit and when you do not, a decision framework
7. How to choose an audit provider, 8 questions you have to ask
8. Case study: a construction company, a no-audit scenario vs an audit scenario
9. FAQ, the 10 most common questions

What a pre-deployment AI audit is

A pre-deployment AI audit is a structured diagnosis of an organization's readiness to deploy solutions based on artificial intelligence. It covers four dimensions: the quality and availability of data, the maturity of business processes, the team's competencies, and the fit-gap between available technologies and the company's real problems. The result is a written report with a list of priority processes, ROI estimates, and a deployment roadmap, independent of any specific technology provider.

This is not a technical audit in the style of "let's check the GPU infrastructure". It is a business audit that answers three questions the leadership team needs an answer to before signing a contract with a provider:

1. Does what we want to do actually require AI, or will plain process automation (RPA, workflow) solve it more cheaply and more reliably
2. Do we have the data the model will learn from or will process, clean, complete, accessible, GDPR-compliant
3. What exactly will change in the P&L, who will stop performing which tasks, what it costs today, what it will cost tomorrow, and over what horizon the investment pays back

If an organization answers any of these with "I do not know" or "broadly yes", the deployment will fall apart. This is not an opinion. It is data, which I discuss below.

How a pre-deployment AI audit differs from an IT audit

| Dimension | IT audit | Pre-deployment AI audit | |---|---|---| | Main question | Does the infrastructure work | Is the organization ready to change | | Duration | 1-3 weeks | 2-8 weeks | | Who it talks to | CTO, IT department | C-level, operational managers, end users | | What it analyzes | Logs, configurations, security | Business processes, KPIs, data quality, competencies | | Output | Technical report | Transformation plan + business case per use case | | Provider dependence | Often tied to a specific technology | Vendor-agnostic |

How a pre-deployment AI audit differs from strategy consulting

Big4 (Deloitte, PwC, EY, KPMG) and MBB (McKinsey, BCG, Bain) sell AI strategy consulting, it costs PLN 500,000 to several million, takes 3-6 months, and ends with a 200-page slide deck about a "transformation journey". The pre-deployment audit is one operational level lower: it does not answer whether to go into AI, but where exactly and in what order. It takes weeks, not months. It costs an order of magnitude less. And it gives you a plan you can deploy in the first quarter after the report, not in 2028.

It is the difference between a medical diagnosis before surgery and a textbook on general medicine. Both have value. But if surgery is about to happen, the second one will not help.

Why 80% of AI projects in enterprises fail

The numbers from the last 18 months are merciless. Six independent sources, different methodologies, the same conclusion:

| Study | Sample | Result | Source | |---|---|---|---| | RAND Corporation 2024/2025 | 65 ML engineers and data scientists | 80% of AI projects fail to reach production, 2x more than non-AI IT projects | RAND RRA2680-1 | | MIT Project NANDA 2025 | 300+ GenAI deployments, 150 interviews | 95% of GenAI pilots with no measurable P&L impact | MIT NANDA Report | | S&P Global Market Intelligence 2025 | 1000+ firms in the US and Europe | 42% of firms abandoned most AI initiatives in 2025 (vs 17% in 2024) | S&P Global | | BCG "Widening AI Value Gap" 09/2025 | 1250 respondents | 60% of firms with no material value despite continued investment | BCG | | McKinsey Global AI Survey 11/2025 | Global enterprises | 80%+ of organizations with no significant AI impact on EBIT | McKinsey | | Gartner Q1 2026 | 782 Infrastructure & Operations leaders | Only 28% of AI projects meet ROI expectations | Gartner |

Five recurring patterns of failure

Data from RAND (Ryseff, Narayanan 2024/2025), MIT NANDA, McKinsey, and Gartner forms five patterns that recur in 70-80% of failed projects. In order of frequency:

Pattern 1: No success metrics defined before the start (73% of projects)

The project kicks off because the CEO came back from a conference in Davos or from an NVIDIA webinar. No one wrote down what is supposed to change in the numbers after 6, 12, 24 months. Without a baseline there is no way to judge whether the project works. The CFO cuts the budget after two quarters. The CEO stops showing up to the review.

This is the most common and cheapest cause of failure to fix, one strategy workshop before the start is enough.

Pattern 2: The data is not "AI-ready" (43% of responses in the Informatica CDO Insights 2025 survey)

Gartner defined AI-ready data as data that is:

• matched to a specific use case (not "general company data")
• actively managed at the asset level (who owns it, who has access)
• supported by automated pipelines with quality gates
• with metadata updated live
• continuously checked for quality

In most Polish enterprises none of these conditions is met. The ERP has 4-year gaps, the CRM does not integrate with operational systems, the sales team keeps key data in spreadsheets. AI trained on such data returns garbage, just faster than a human.

Pattern 3: Loss of C-level sponsorship in the first 6 months (56% of projects)

Projects with sustained board engagement have a 68% chance of success. Projects where the C-level withdraws after the first phase, 11% (RAND 2024). A sixfold difference.

Sponsorship does not mean "the CEO said yes at the board meeting". Sponsorship means: the CEO looks at four project indicators every month, asks questions, defends the budget against the CFO, and intervenes when operational departments sabotage the deployment.

Pattern 4: Treating AI as an IT project, not a business transformation (61% of projects)

The classic pattern: the board delegates "this AI thing" to the CTO. The CTO picks the technology. The operational department is handed a finished tool with an order to "use it". No one redesigned the processes. No one changed the KPIs. No one updated the compensation system.

McKinsey 2025 shows this unambiguously: organizations reporting "significant financial returns" from AI are 2x more likely to redesign end-to-end business processes before choosing an AI model. That reverses the typical order of deployments, and it pays off.

Pattern 5: Expectations too ambitious, the horizon too short (57% per Gartner I&O 2025)

I&O leaders in the Gartner survey say it plainly: projects fell apart because "we expected too much, too fast". They assumed AI would instantly automate complex processes, cut costs, and fix long-standing operational problems. After 90 days with no effect, a loss of confidence.

The real horizon for measurable ROI from AI in an enterprise: 2-4 years, not the 7-12 months vendors promise.

In conversations with the boards of large Polish companies I see the same thing RAND described in 65 interviews: the technology is ready, the organization is not. Five out of six projects we enter fall apart before the first line of code, because no one defined what is supposed to change in the P&L and who will use the finished tool. A pre-deployment audit is not bureaucracy. It is the only stage at which you can avoid sinking PLN 2-8 million.

What a thorough AI audit covers

A thorough pre-deployment AI audit covers four areas. Each has its own questions, its own deliverables, and its own risks. All four are necessary, skipping any one of them significantly increases the chance that you will join the 80% statistic.

At yesfor.ai we work with our own YESFOR Framework (AI Readiness Score, YESFOR). Four pillars, twelve dimensions, one measurable score from 0 to 100. The pillar weights come from empirical data on the causes of AI project failures (RAND, MIT, McKinsey), not from thin air.

Pillar 1: Data audit (weight in OPS: 30%)

Questions the audit answers:

• What data the company has in which system and who owns it
• The quality of the data across five dimensions: completeness, consistency, accuracy, timeliness, uniqueness
• Whether the data can be linked across systems without manual intervention
• Whether we have consent (GDPR) to use this data to train models
• Where the IT department ends and shadow IT begins (spreadsheets in sales, marketing, HR)

Deliverables:

• A map of data sources with a quality rating per asset
• A list of gaps between the current state and the AI-ready state
• An estimate of the cost and time to bring the data to readiness
• A list of 5-10 use cases possible today (on data in its current state) and 5-10 possible after raising quality

Red flags we look for:

• "We have a data warehouse", most often means an ETL store not updated for 2 years with 60% of tables unused
• "We have an ERP", most often means 30% of key fields are empty or filled with "n/a"
• "We have GDPR consent for everything", most often untrue

Pillar 2: Process audit (weight in OPS: 30%)

Questions:

• Which processes in the company are the most costly (time + money)
• Which are repeatable, structured, and measurable
• Which are actually carried out the way they are described in the documentation (most often the difference is 40-60%)
• Where the bottlenecks are, where errors arise, where people wait on other people

Deliverables:

• An end-to-end process map for 5-10 priority areas
• A classification of each process in an AI potential x organizational readiness matrix
• A list of quick wins (processes to automate in under 90 days) and long-term plays
• A baseline KPI for each process: time, cost, quality, user satisfaction

A critical note: Not every process needs AI. Some can be handled with classic RPA (Robotic Process Automation), some with a simple workflow in Microsoft Power Automate, some with a change to a work instruction. A thorough audit says "AI makes sense here, not there", it does not force AI everywhere.

Pillar 3: Competency and organization audit (weight in OPS: 25%)

Questions:

• Whether the board genuinely knows what AI is (vs a marketing buzzword)
• Whether there are people in the organization who will sustain the deployment after the consultant leaves (capability transfer)
• Whether the compensation structure rewards working with AI or blocking it
• What the sentiment toward AI is in operational teams, the Writer/Workplace Intelligence 2025 study found that 31% of employees deliberately sabotage AI deployments

Deliverables:

• An AI Literacy Score for the board, middle management, and operational teams
• A change readiness analysis: who will be an ally, who a blocker
• A capability transfer plan (so that after the project ends the company can maintain and develop the solution itself)
• Recommendations for changes to KPIs and the compensation system

Pillar 4: Technology and fit-gap audit (weight in OPS: 15%)

Questions:

• What infrastructure the company has today, what it needs for a given use case
• Cloud (AWS, Azure, GCP) vs private cloud vs on-premise
• Build vs buy: build our own, buy off-the-shelf (Salesforce Einstein, Microsoft Copilot), or hybrid
• What the costs of maintenance, model retention, and monitoring are

Deliverables:

• A target architecture map
• A comparison of 3-5 technology options with a 3-year TCO
• A list of technical risks (vendor lock-in, data residency, AI Act compliance)

Why only 15% weight: Because it is the least important pillar. RAND and MIT show unambiguously that technology is rarely the cause of failure in 2026. The organization breaks, the data breaks, the sponsorship breaks. The technology is ready, if the rest is ready.

What the audit process looks like step by step

A standard yesfor.ai audit for a mid-sized Polish enterprise (200-500 people) takes 4-6 weeks and runs in 6 phases:

Phase 1: Discovery call (30 minutes, online, free)

An introductory conversation with a decision-maker. Two goals: to understand the business context and to check whether the audit is even worth doing. Sometimes the client needs something else, deploying a specific tool, training the board, RPA instead of AI. We say so directly. An audit does not always make sense, the question is not "whether", but "where the greatest potential is".

Phase 2: Scope and pricing (week 1)

We define: how many processes we cover, how many people we talk to, which systems we analyze, which deliverables we provide. We sign a short contract (3 pages) with a fixed price. No surprises on the invoice.

Phase 3: Board workshops (week 2)

Four 2-hour sessions:

• Strategy session, what you want to achieve and over what horizon
• Operations session, which processes hurt the most
• Data session, what you have, what you do not have, where the shadow IT is
• KPI session, how we will measure success

Phase 4: Operational interviews (weeks 2-3)

This is the heart of the audit. Conversations not with the CEO, but with the people who actually do the work every day. Department managers, key users, analysts, controllers. We map each process step by step: who does what, in what order, in which tool, how long it takes, where the errors are.

Typically 15-25 interviews of 45-60 minutes.

Phase 5: Analysis and scoring (weeks 3-4)

Work on our side. We assemble the data from interviews, documentation, systems. We calculate the AI Readiness Score for each of the 12 dimensions. We identify gaps, risks, quick wins. We estimate ROI for 5-10 priority use cases.

Phase 6: Report and presentation (weeks 4-6)

The written report (40-80 pages) contains:

• An executive summary (1 page for the board)
• The company's AI Readiness Score + benchmarking against the Polish market
• A process map highlighting priorities
• A deployment roadmap (12, 24, 36 months)
• A business case for 3-5 quick wins with numbers
• A list of risks and recommendations
• Technical appendices

We present the report to the board (a 2-3 hour session). The client receives an editable version and may pass it to any implementer, including a yesfor.ai competitor. The knowledge stays with you, not with the consultant.

How much an AI audit costs in a Polish company

Full price transparency, which you will not get at the Big4:

| Company size | Audit scope | Price | Duration | |---|---|---|---| | SME (10-100 people) | 1-2 processes, 5-8 interviews | PLN 8,000 - 15,000 | 2 weeks | | Mid-market (100-500 people) | 3-5 processes, 10-15 interviews | PLN 25,000 - 50,000 | 4 weeks | | Enterprise (500+ people) | 5+ processes, 20-30 interviews | PLN 80,000 - 200,000 | 6-8 weeks |

Why ranges rather than fixed prices: the difference between the low and high end comes from the complexity of the systems, the number of interviews, the availability of historical data, and the required depth of the report. The quote after the discovery call is a fixed price.

What raises the price:

• Processes across multiple branches / locations
• Many source systems to analyze
• Compliance requirements (financial, medical, public sector)
• English-language deliverables (for international boards)
• NDAs and security audits on the client side

What lowers the price:

• A clearly defined scope at the start
• Processes mapped in advance
• A ready team on the client side to collaborate
• A repeat order (client loyalty)

Comparison with the alternatives

| Option | Time | Cost | What you get | |---|---|---|---| | Audit at yesfor.ai | 4-6 wks | PLN 25-200k | An operational plan, vendor-agnostic, capability transfer | | AI Strategy at the Big4 (Deloitte/PwC/EY/KPMG) | 3-6 mo | PLN 500k - 2M | A strategic slide deck, usually vendor-biased | | "Free audit" from an AI vendor | 1-2 wks | PLN 0 | A deployment plan for their tool | | Deployment without an audit | 6-24 mo | PLN 1-10M | An 80% chance of joining the RAND statistic |

The audit is the cheapest element of the entire AI value chain. It is 2-5% of the budget that secures the other 95-98%.

When you need an external audit, and when you do not

An audit probably makes sense if:

• The company has 50+ people and wants to deploy AI in a key process (not an HR experiment)
• The budget planned for the deployment is PLN 500,000+
• You have 2+ proposals from different providers and do not know which to choose
• The company's first AI project did not pan out and you do not know why
• The board is divided on whether to go into AI at all
• You are preparing for AI Act compliance (high-risk sector)

An audit probably does NOT make sense if:

• You have a specific, well-described problem and are looking for an implementer, then it is better to start with a direct quote
• The company has fewer than 30 people and a single owner-decision-maker, it is faster to talk directly about the tool
• The deployment budget is below PLN 100,000, the audit would eat 20-30% of it
• You have an internal AI/ML team that can do it itself

In these cases we tell the client plainly: "you do not need us, here is what to do instead". Yes, we lose the project. But we get a reference.

How to choose an AI audit provider

Eight questions to ask an audit provider before you sign:

1. Does the report become my property when it ends?

It should. If the provider wants to keep rights to the report or refuses to hand over an editable version, that is a red flag. The knowledge is to stay with you.

2. Are you independent of any specific technology provider?

Vendor-agnostic is not a slogan. Ask directly: "Do you have partnerships with AI providers? Do you receive commissions for referrals?". A clear answer = an honest partner. An evasive answer = the audit will end with a recommendation for their partner.

3. How will you talk to our people?

An audit without conversations with key users is not an audit. You should hear: a minimum of 10-15 interviews for a mid-sized company, with people from different levels of the organization, not only the C-level. Without conversations with the people who carry out the processes every day, there is no way to assess what really does not work.

4. What deliverables exactly will I receive?

A list of concrete documents: a process map, an AI Readiness Score (or another measurable indicator), a business case with numbers, a roadmap, a written report. No specifics = no process.

5. Do you have your own methodology?

If the answer is "we use Gartner/McKinsey frameworks", keep asking. A proprietary, documented, publicly described methodology (such as the yesfor.ai YESFOR Framework) is a sign of a mature practice. Without a methodology, every audit is improvisation.

6. Show me 2 anonymized reports from previous audits

Any serious provider can show anonymized excerpts of reports. If they answer "that is confidential, we will show you nothing", they are probably ashamed of the quality.

7. How long does it take to implement the audit's recommendations?

An audit that ends with a recommendation to "transform over 3 years" is useless. You should receive quick wins to deploy in 90 days + a long-term plan. Without quick wins, the board loses interest.

8. What happens if I decide not to deploy after the audit?

A good answer: "The report stays with you, you can do whatever you want with it, including choosing a competitor.". A bad answer: "We have a non-compete clause in the contract.".

Case study: a hypothetical deployment in a construction company

Context: A Polish general contractor, 3,000 people, PLN 2B in turnover. After a conference in Davos the board makes a decision: "we are deploying AI". The CTO gets a PLN 8M budget and 12 months.

Scenario A, without an audit (the classic failure)

Months 1-3: The CTO selects a provider through an RFP process. They go with a predictive maintenance solution for construction machinery. The vendor promises a 30% reduction in breakdowns.

Months 4-9: Deployment. It turns out that:

• Telematics data from the machines is incomplete (60% of vehicles have older GPS units)
• Drivers do not want to log extra events (it does not affect their bonus)
• Site managers ignore the alerts, because "they know their machines better than the system"

Months 10-12: The system works technically, but only 12% of the people it was meant for use it. The CFO looks at the invoices: PLN 6.8M spent, PLN 0.3M of documented savings. The project is closed. Employees say: "AI does not work".

What actually did not work: people, data, and processes, in that order. The technology was fine.

Scenario B, with an audit (which is how it looks with us)

Weeks 1-6: A yesfor.ai audit. The AI Readiness Score comes out at 38/100. Main findings:

• Board sponsorship is strong (pillar O: 72/100)
• Processes are not mapped, every site has its own (pillar P: 28/100)
• Data is fragmented, 4 systems with no integration (pillar S: 31/100)
• No AI literacy at the level of site managers (pillar Skills: 22/100)

Recommendation: do not go into predictive maintenance now. Instead, three quick wins:

1. AI-assisted cost estimation for bids (3-month deployment, ROI in 6 months), the data is already in SAP, 12 users in the bidding department, easy quick wins
2. Classification of project documents (BIM + AI), high value, low risk
3. Prediction of schedule delays based on the history of projects

Equipment maintenance, in phase 2, in 12-18 months, after bringing the telematics to AI-ready.

Months 4-12: Deployment of the three quick wins on a PLN 2M budget (out of 8M). The remaining 6M stays for phase 2.

Result after 12 months:

• Bid cost estimation 60% faster
• Bid win rate up from 14% to 19%
• Time saved in the bidding department: 2 full-time roles
• Documented ROI: PLN 2.1M/year against a cost of PLN 2M
• AI Readiness Score after 12 months: 58/100

Year 2: the predictive maintenance phase, but this time on cleaned data, with mapped processes, with driver bonuses tied to using the system.

The difference: PLN 4.7M difference in the P&L in the first year. That is the price you pay for skipping the audit.

Frequently asked questions

• What is a pre-deployment AI audit?+
• How long does an AI audit take?+
• How much does an AI audit cost in a Polish company?+
• Is our company ready for AI?+
• Is an audit needed if I already have an AI provider?+
• Which companies most often order an AI audit?+
• Does the report stay with us after the audit ends?+
• Do we have to deploy with you if we do the audit?+
• What is the YESFOR Framework?+
• How long is the wait for an audit slot?+

The next step

If you have read this to the end, you are probably wondering whether your company is in the 80% (that fail) or the 20% (that deliver). There are three ways to check:

1. AI Readiness Self-Assessment, 30 minutes, free, online. You get a preliminary AI Readiness Score and a list of areas that need attention. No consultant, no email to a sales department.

2. Discovery call with yesfor.ai, 30 minutes, free, online. A conversation about your specific business context. You will leave knowing whether an audit makes sense in your case, or whether you need something else.

3. Full pre-deployment audit, 4-6 weeks, fixed price after discovery, the report stays with you. The decision comes after the discovery session, not before.

We do not push the audit. Pushing an audit onto someone who does not need it ends in a bad reference. Bad references cost more than any single project.

Sources and further reading

Research this article is based on:

• Ryseff, J., De Bruhl, B. F., Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation. RAND RRA2680-1
• Challapally, A. et al. (2025). The GenAI Divide: State of AI in Business 2025. MIT Project NANDA.
• S&P Global Market Intelligence (2025). AI Initiative Abandonment Survey.
• BCG (2025). The Widening AI Value Gap.
• McKinsey & Company (2025). Global AI Survey: November 2025.
• Gartner (2026). AI Projects in I&O Stall Ahead of Meaningful ROI Returns. Gartner press release April 2026
• Goedecke, S. (2025). Is it worrying that 95% of AI enterprise projects fail?, a critical analysis of the MIT NANDA methodology

Further articles on yesfor.ai:

• YESFOR Framework, a proprietary method for assessing AI readiness
• Why 80% of AI projects in companies fail, the full analysis
• AI vs process automation, when to use which