Why AI Projects Fail on Old Systems: A Practical AI-Readiness Audit
Why AI Projects Fail on Old Systems: A Practical AI-Readiness Audit
A powerful AI model cannot repair missing records, undocumented business rules, disconnected applications or unsafe permissions by itself. Before building an AI agent, inspect the systems it must depend on.
Capgemini says companies are beginning a multi-year modernisation cycle as they prepare to use AI at scale. Legacy applications, fragmented data and complex technology environments can prevent AI systems from completing real business work reliably.
- A seven-part AI-readiness scorecard.
- A 60-minute audit that can be completed before buying an AI platform.
- Practical SQL checks for missing and duplicated data.
- A decision guide for wrapping, refactoring, replacing or retiring old software.
- A 90-day roadmap for preparing one real process instead of attempting a company-wide transformation.
What does “AI-ready” actually mean?
An AI-ready organisation is not simply one that owns cloud servers or has purchased access to an advanced model.
The organisation must be able to provide the AI system with reliable information, clearly defined permissions and dependable ways to read or update business systems.
A process is AI-ready when its data, rules, software connections, security controls and human responsibilities are clear enough for an automated system to operate predictably.
Readiness should be assessed for one process at a time. A company may be ready to use AI for document search while being completely unprepared to let an agent approve payments or modify customer records.
Seven reasons AI projects fail before the model is tested
A customer-support agent may produce an excellent written reply but still fail operationally because it reads an outdated delivery address, cannot identify the latest order or has no secure method to issue a refund.
The 60-minute AI-readiness audit
Choose one small process. Do not begin by auditing the entire organisation. A useful example is “answering a customer’s order-status question.”
At the end of the hour, the team should understand whether its largest problem is the model, data, integration, security or the business process itself.
The seven-part AI-readiness scorecard
Give each area a score of 0, 1 or 2. The maximum total is 14.
| Area | 0 points | 1 point | 2 points |
|---|---|---|---|
| Process ownership | No responsible owner. | Informal owner with unclear authority. | Named owner responsible for outcomes and exceptions. |
| Data quality | Frequent missing or conflicting records. | Usable after manual correction. | Measured, monitored and sufficiently reliable. |
| Shared identifiers | Systems cannot reliably match the same customer or item. | Matching requires rules or manual review. | Stable identifiers connect records across systems. |
| System access | No dependable integration method. | Manual export or fragile connector. | Documented API, controlled database view or stable integration. |
| Security | Requires broad administrator access. | Some restrictions exist but are incomplete. | Task-specific identity with minimum required permissions. |
| Observability | Actions cannot be reconstructed. | Partial technical logs exist. | Inputs, decisions, actions, errors and approvals are traceable. |
| Failure handling | No defined response to errors. | Staff intervene informally. | Clear escalation, rollback and human-approval procedures exist. |
Practical data-quality checks
Before connecting a model, measure the condition of the records it will use. The following generic SQL examples can be adapted to a customer table.
SELECT
COUNT(*) AS total_records,
SUM(CASE
WHEN customer_id IS NULL THEN 1
ELSE 0
END) AS missing_customer_ids,
SUM(CASE
WHEN email IS NULL OR TRIM(email) = '' THEN 1
ELSE 0
END) AS missing_emails
FROM customers;SELECT
customer_id,
COUNT(*) AS record_count
FROM customers
WHERE customer_id IS NOT NULL
GROUP BY customer_id
HAVING COUNT(*) > 1
ORDER BY record_count DESC;SELECT
customer_id,
last_updated
FROM customers
WHERE last_updated IS NULL
OR last_updated < CURRENT_DATE - INTERVAL '365 days';These queries do not repair the data. They convert an unclear concern—“our data may be messy”—into measurable findings that can be assigned and monitored.
An AI system may suggest that two records belong to the same customer, but merging them automatically can create privacy, billing or legal problems. High-impact corrections should be reviewed and logged.
You do not need to replace every old system
Modernisation should begin with the smallest change that creates a reliable and secure interface for the required process.
| Approach | Use when | Main risk |
|---|---|---|
| Wrap | The old system still works, but it needs a controlled API or service layer. | The wrapper can hide underlying data-quality and performance problems. |
| Refactor | Important code is maintainable but needs restructuring, testing or improved interfaces. | Hidden business rules may be changed accidentally. |
| Replatform | The application can move to newer infrastructure without redesigning every business function. | Moving the same design may preserve old limitations. |
| Replace | The software is unsupported, unsafe or too expensive to maintain. | Data migration and employee adoption can be more difficult than expected. |
| Retire | The process is duplicated, rarely used or no longer creates business value. | Important historical information may be lost without an archive plan. |
When wrapping is enough
A university may have an old student-record system that remains accurate but provides no modern integration. A read-only service can expose approved information such as course registration and examination status without allowing the AI assistant to modify official records.
When replacement is justified
Replacement becomes more reasonable when the system cannot receive security updates, depends on unavailable specialists, regularly corrupts data or prevents the organisation from meeting important operational requirements.
A 90-day modernisation roadmap
Days 1–30: Understand and measure
- Select one narrow and valuable process.
- Name the business owner and technical owner.
- Document every data source and software dependency.
- Measure missing, duplicated and outdated records.
- List common exceptions and manual decisions.
- Record the current time, cost and error rate.
Days 31–60: Build the foundation
- Create a controlled API, database view or document index.
- Introduce shared identifiers where possible.
- Clean only the data required for the selected process.
- Create a task-specific service account.
- Add logging for reads, decisions, actions and failures.
- Prepare test cases using normal and exceptional situations.
Days 61–90: Run a controlled pilot
- Start with read-only recommendations.
- Require human approval for external or irreversible actions.
- Compare results with the existing process.
- Measure accuracy, response time, cost and escalation rate.
- Investigate failures instead of hiding them.
- Expand only after defined success conditions are achieved.
AI readiness for a small business
A small organisation does not need an expensive transformation programme. It can begin by placing product, customer and transaction information in one controlled system, removing duplicated spreadsheets and documenting how common decisions are made.
A simple, accurate dataset with clear ownership is more useful than a large collection of disconnected tools.
What AI cannot repair by itself
Original analysis: AI is exposing technical debt that companies learned to tolerate
Employees can often work around poor systems. They know which spreadsheet is current, which field is unreliable and which colleague understands an unusual transaction.
An AI agent does not automatically possess this informal organisational knowledge. It follows the data, interfaces and permissions it is given.
This is why AI adoption can make old weaknesses suddenly visible. The AI project may receive the blame, even when the actual failure is an undocumented rule, duplicated database or unstable integration that existed for years.
The strongest modernisation programmes will not begin with the question, “Where can we add an AI agent?” They will begin with, “Which important process has clear value, measurable problems and an owner willing to improve it?”
AI then becomes one component of process improvement rather than an expensive layer placed over unresolved technical debt.
Student project: conduct an AI-readiness audit
Choose a familiar process such as library-book renewal, student registration, studio order tracking or customer appointment scheduling.
- Write the intended AI task in one sentence.
- Draw the current information flow.
- List every database, spreadsheet and document involved.
- Create 20 sample records and introduce realistic errors.
- Run missing-value and duplicate checks.
- Score the process using the seven-part scorecard.
- Select wrap, refactor, replace, replatform or retire.
- Design a human-approval and failure-handling process.
- Present the improved architecture.
- Explain which problem AI solves and which problems require organisational decisions.
This creates a stronger portfolio project than a chatbot demonstration because it shows data analysis, system design, security and business reasoning.
Frequently asked questions
Does a company need to move everything to the cloud before using AI?
No. A secure hybrid approach may be appropriate when important systems must remain on-premises. The key requirement is controlled and reliable access to the necessary data and functions.
Can AI clean legacy data automatically?
AI can suggest duplicates, missing values and possible corrections. High-impact changes should still follow defined rules, review and audit procedures.
Should every legacy application be replaced?
No. Some systems remain stable and valuable. Wrapping or refactoring may create sufficient access without the risk and cost of complete replacement.
What should the first AI pilot do?
Begin with a narrow, measurable and reversible task. Read-only search, classification or recommendation is usually safer than autonomous financial or administrative action.
What is the biggest warning sign?
The strongest warning is when no one can clearly explain which data is authoritative, who owns the process or what should happen when the AI is uncertain.
Final takeaway
AI projects usually need more than a better model. They need accurate data, stable integrations, limited permissions, visible logs, documented rules and a responsible process owner. Modernise the process that matters before automating it.
Reuters — Capgemini sees multi-year IT modernisation boom as firms prepare for AI:
Read the Reuters report
Capgemini — Data and artificial intelligence services:
Review Capgemini’s modernisation overview
Google Cloud — Organisational readiness for AI adoption and scale:
Review Google Cloud’s AI-readiness guidance
NIST — Supporting digital transformation with legacy components:
Review NIST’s legacy-system guidance
Join the conversation