Why AI projects fail, and the practical fix for each cause

Why AI projects fail: what does the research say?

The research agrees that most projects stall between pilot and production, and that the model is seldom the reason. MIT NANDA reported in July 2025 that 95% of organizations saw no measurable P&L return from generative AI, and pointed to brittle workflows, tools that do not learn from context and poor fit with daily work.

BCG's October 2025 report on the widening AI value gap found that only a small minority of companies get value from AI at scale. According to an S&P Global Market Intelligence survey, a growing share of companies abandoned most of their AI initiatives in 2025, citing cost, data privacy and security. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027 because of cost, unclear value or weak risk controls, and warned about agent washing.

Read together, those findings point to seven causes. The sections below take them one at a time, each with the fix we apply.

Cause 1: Nobody measured the starting point

A project cannot show value if nobody recorded how the work ran before it. Gartner's unclear value is often exactly this: the system works, but there is no before figure to compare it with, so the budget review has nothing to point to.

The fix is to measure before building. For the chosen workflow, record minutes per task, tasks per month, the error or rework rate and who does it. Agree the target in the same units (for example, invoice entry time and the share of invoices needing correction) and put the comparison in a monthly report from launch onward.

If you cannot measure the baseline yet, spend two weeks logging it by hand before funding the build. It is cheap, and it often shows that the real bottleneck is somewhere other than where everyone assumed.

Cause 2: The workflow breaks on real inputs

Brittle workflows, the first cause MIT NANDA named, usually come from testing on tidy samples. Real inputs include scanned PDFs, forwarded email chains, callers who change topic halfway through, and records that already exist in the CRM under a slightly different name.

The fix is an eval set built from real, messy examples with known correct answers, rerun after every change to prompts, retrieval or model. Add field-level validation (totals that must add up, dates that must be in range, IDs that must exist in NetSuite or QuickBooks) and a confidence threshold that sends doubtful items to a human-review queue instead of guessing.

• Eval set drawn from the last few months of actual inputs, including the failures.
• Validation rules on every field written to a system of record.
• Human-review queue for anything under the confidence threshold.

Cause 3: The tool sits outside the daily work

Poor fit with daily work, MIT NANDA's third cause, shows up as a separate chat window or dashboard that people have to remember to open. Within weeks they stop opening it, and usage quietly drops to zero.

The fix is to deliver output where the work already happens: a draft reply inside the Zendesk ticket, a filled record in HubSpot or Salesforce, a vendor bill waiting for approval in the ERP, a summary posted in the team's existing channel. One of our case studies shows the same principle without a chatbot at all: a therapy services provider moved clinicians to a guided AppSheet portal with validation and billing sync, data entry fell from 10-15 minutes a day to under 3, and errors were stopped at source.

Cause 4: The system does not learn from context

A system that answers from a frozen prompt keeps making the same mistakes, which is why tools that do not learn from context made MIT NANDA's list. Policies change, price lists update and staff correct the same error every week with no effect.

The fix has two parts. First, ground answers in current company documents through retrieval with citations, so updating a file in SharePoint or Google Drive updates the answer and a reviewer can see the source. Second, close the loop: every correction made in the review queue is logged, added to the eval set and used to adjust prompts or rules on a regular schedule.

For voice agents the same idea applies to transcripts: the calls where the agent transferred, misunderstood or dropped the caller are the best training material for the next round of prompt and routing changes.

Cause 5: Costs grow faster than value

Cost is the reason that appears in both the S&P Global and Gartner findings. Usage-based bills for model tokens and voice minutes rise with volume, and an agent stuck in a loop can spend more in an afternoon than its task was worth.

The fix is to engineer the budget into the system. Set a cost cap per task that stops the run and alerts someone, log tokens and minutes per task from the first test, route simple classification to a smaller model and keep the larger one for real reasoning, and review cost per task against value per task every month. Our AI implementation cost guide lists the running cost components to track.

Cause 6: Privacy and security concerns stop the rollout

Data privacy and security are the other reasons S&P Global's survey cited, and they often surface late, when legal or IT sees the design for the first time. A pilot built on a consumer chat account cannot move to production with customer or patient data.

The fix is to settle data handling in the first weeks. Use API or enterprise tiers whose terms fit your data, get a signed Business Associate Agreement from every vendor that touches protected health information, give the system only the narrowest access it needs, and log every prompt and output. Clients own the accounts, so the data stays under contracts they control.

Bring IT and legal into the review before the build, not after it. A one-page data flow showing which systems are read, which are written, which vendors see the data and where logs are stored answers most of their questions early.

Cause 7: Agents act without enough control

Weak risk controls are the third reason in Gartner's cancellation forecast, and agents raise the stakes because they take actions, not just write text. Agent washing adds to the problem: products sold as autonomous agents that are really simple scripts, or that act with far more freedom than the buyer realized.

The fix is to bound what an agent may do. List its tools explicitly, give each tool the minimum permission, require human approval for anything that moves money, contacts a customer or changes regulated records, and keep a rollback path that can switch the agent off and return work to people. When evaluating a vendor, ask them to show the tool calls and logs from a real run, not only the final answer.

Should you build in-house or with a specialist partner?

MIT NANDA reported that projects bought from or built with specialist partners succeeded more often than internal builds. The same research found the biggest returns came from back-office automation, even though budgets leaned toward sales and marketing.

Whichever route you take, the fixes above are what matter. We scope each project in a free review, price it as a fixed build, and hand over the code, prompts and accounts so your team can run it. Our implementation roadmap guide sets out the phases in order.