AI Opportunity Assessment: Why We Built Diagnostik

Ask a leadership team where the business loses time, and the answers will come quickly.

Sales waits for quotations. Operations rebuilds the production schedule whenever an order changes. Finance chases information before it can release an invoice. Experienced employees spend hours answering questions that the company has answered before.

Then ask which problem deserves investment first.

The conversation becomes harder. Each department has a reasonable case. The costs sit in different budgets. Some problems need better information; others need a clearer process. Several could benefit from AI, but nobody has established which ones are ready for it.

Buying a tool at that point leaves the most consequential decision unresolved.

An AI opportunity assessment helps a business identify where AI could create value, test whether the opportunity is feasible, and decide what to pursue first. It connects the operational problem to the investment decision before implementation begins.

That is why we built Diagnostik, with an AI interview agent called Iris at its centre. Diagnostik is designed to map business processes, surface where time and money are being lost, and establish priorities for transformation.

The proposition is focused: 10 days. Six dimensions. One fundable case.

01 AI should not be your starting point

There is a difference between knowing that AI is powerful and knowing where it belongs in your business.

A model may extract information from documents accurately while the underlying approval process still takes a week. A forecasting tool may improve predictions while planners continue working from conflicting versions of the order book. A useful assistant may save employees time without changing the constraint that limits output.

The business case depends on what changes after the technology produces its answer.

McKinsey’s 2026 global AI survey illustrates that distinction. Eight in ten respondents said AI had improved their own productivity, while 37% reported a positive contribution to their organisation’s earnings before interest and taxes. These are self-reported survey findings, not proof that any particular assessment method produces returns. They show why individual time savings and enterprise financial value need to be evaluated separately. Source: McKinsey.

For leadership, the starting questions are practical. Where does work stall? What does that delay cost? Who owns the outcome? What would have to change for an improvement to reach the customer, production line or financial result?

AI can help investigate those questions, as Iris does within Diagnostik. But the decision to deploy AI in an operational process should follow an understanding of the problem it is meant to solve.

02 What should an AI opportunity assessment establish?

A useful assessment gives leadership enough evidence to compare possible investments. It connects three questions:

• Where is the opportunity? A recurring problem, affected users, and a measurable consequence.
• Can we act on it? Suitable data, workable integration, operational ownership, and acceptable risk.
• Is it worth doing first? Expected benefits, full costs, dependencies, and the evidence needed to proceed.

An AI readiness assessment generally examines whether an organisation has the foundations to adopt AI. An opportunity assessment also asks which business problem makes those foundations worth investing in. In practice, the two overlap: an attractive opportunity may require data preparation, while a technically ready organisation may still lack a worthwhile use case.

Microsoft’s readiness assessment, for example, examines business strategy, governance, data, organisational culture, infrastructure and other capabilities. Those are useful foundations, but a funding decision also needs a specific workflow and an expected result. Source: Microsoft AI Readiness Assessment.

The assessment should leave room for several answers. Start an AI project. Improve the data first. Simplify an approval. Connect existing systems. Or leave a process alone because the likely gain does not justify the disruption.

That flexibility protects the quality of the decision.

03 Start with the work people actually do

The organisation chart tells you who reports to whom. It rarely explains why a customer order waits three days between sales and production.

To identify AI opportunities in your business, follow a piece of work from its trigger to its completed outcome. Examine what people receive, what they check, which systems they use, and what happens when something is missing or unusual.

An invoice might require information from a purchase order, a delivery record and an email. The visible task is data entry. The expensive part may be resolving disagreements between those records. Automating entry alone would leave the difficult work intact.

Useful discovery questions include:

• Which recurring task requires someone to chase information?
• Where is the same information entered or checked more than once?
• Which exceptions depend on one experienced employee?
• What causes completed work to return for correction?
• Which delay affects revenue, service, throughput or cash collection?

These questions help turn frustration into a problem that can be investigated. “Reporting is slow” becomes “the team spends every Monday reconciling incompatible order records before production planning can begin.” The second statement identifies a workflow, a dependency and a place to measure.

RAND’s 2024 research, based on interviews with 65 experienced data scientists and engineers, identified misunderstanding the business problem and inadequate data among recurring causes of AI-project failure. The study concerned AI and machine-learning projects; it excluded projects that simply used pretrained language models. Its findings support a disciplined approach to problem definition, rather than a universal failure-rate claim about every AI tool. Source: RAND.

04 Why we built Iris into Diagnostik

People close to a process often know exactly where it breaks. What they lack is a consistent way to connect that knowledge to the wider business decision.

A leader may describe margin pressure. A manager may describe repeated approvals. A frontline employee may describe the spreadsheet that keeps the process moving. Understanding the relationship between those accounts is where useful diagnosis begins.

Iris runs the interviews inside Diagnostik. Her purpose is to uncover where money leaks out, where time is wasted, what concerns leaders, and whether the available data can support the AI opportunities under consideration.

This gives the launch proposition a practical starting point: conversations about how the business operates. Technology selection comes after the problem has been made clear enough to assess.

Interviews also need an evidence standard. When someone says a task takes two hours, a sound assessment asks how often it occurs and what those two hours include. When a team describes lost revenue, the business case needs a way to distinguish missed demand from work merely completed later.

Those are principles for evaluating any assessment, including ours. A useful finding should be specific enough that the people responsible for the process recognise it and leadership can question its assumptions.

Diagnostik brings this work into a defined proposition: a ten-day assessment across six dimensions, directed towards a fundable case and an order of priorities. The aim is to help leaders make the next investment decision with a clearer view of the business behind it.

05 Check data readiness for the actual use case

“We have years of data” is a starting observation. It does not establish whether that data can support the proposed workflow.

Consider a maintenance assistant. Service records may exist, but equipment identifiers might differ between systems. Manuals may include obsolete revisions. Technicians may record symptoms in free text without recording the eventual cause. The assistant’s usefulness depends on those details.

Data readiness for AI should therefore be assessed against the intended task:

• Coverage: Are the necessary records and outcomes available?
• Meaning: Can someone explain the fields, terminology and exceptions?
• Reliability: Are records sufficiently complete, current and consistent?
• Access: Can the intended users and systems retrieve the right information with appropriate permissions?
• Evaluation: Is there representative material against which to check performance?

This is a working checklist, rather than a description of Diagnostik’s six dimensions.

A missing foundation does not automatically invalidate an opportunity. It changes its sequence and cost. The next step might be to standardise identifiers or capture outcomes for a defined period before testing a model.

This is also why data preparation should be scoped to a decision. Trying to repair every dataset in the business before starting can create a large programme with no clear measure of completion. Establish what the priority workflow needs, then evaluate that requirement properly.

06 How to prioritise AI use cases

A useful shortlist compares opportunities consistently while keeping uncertainty visible.

For each candidate, examine the expected business effect, the effort required to deliver it, the readiness of users and data, and the consequences of error. Microsoft applies a related approach in its guidance for software providers, evaluating business viability, user desirability and technical feasibility before prioritising use cases. Source: Microsoft business envisioning.

For a leadership team, that comparison can support four practical decisions:

• Test now. The problem matters, the evidence is accessible, and a bounded test can resolve the main uncertainty.
• Prepare first. The opportunity is worthwhile, but a specific data, process or ownership gap blocks progress.
• Investigate further. Potential value is significant, but the assumptions are too uncertain to justify implementation.
• Defer. The expected benefit is limited or the effort and consequences outweigh it.

A high score should never hide a critical dependency. If a proposed application needs records the company cannot access, that is a prerequisite to resolve. Averaging it together with an attractive benefit estimate does not make it disappear.

The option to stop matters too. NIST’s AI Risk Management Framework describes understanding context and impacts as a basis for an initial decision on whether to develop or deploy an AI system. That supports assessing suitability before committing to a build. It does not constitute NIST approval of a particular product or methodology. Source: NIST AI RMF Core.

07 An example: turning quotation delays into a decision

The following scenario is illustrative. It is not an Embiggen customer result or a sample generated by Diagnostik.

Suppose a manufacturer processes 400 quotation requests each month. Preparing each quotation requires an average of 45 minutes of staff effort, including finding specifications, checking previous orders and confirming prices.

That represents 300 staff hours per month: 400 requests multiplied by 0.75 hours.

Leadership initially proposes an AI quotation generator. Discovery reveals that much of the work involves searching for approved specifications, while pricing exceptions still require commercial judgement.

A narrower first experiment could help staff retrieve the relevant information and prepare a draft for review. Pricing authority would remain with the responsible team.

Assume a test targets a 15-minute reduction in average handling time. If achieved across the same monthly volume, that would release 100 hours of capacity per month. At an illustrative loaded labour cost of €40 per hour, the capacity has a gross labour-value equivalent of €4,000 per month.

That figure is not a cash-saving forecast. Salaries may remain unchanged. The value depends on whether the company uses the capacity to process more quotations, reduce overtime, avoid additional hiring or improve another outcome. Any revenue benefit would need separate evidence, and the same gain should not be counted twice.

Implementation, data preparation, integration, review, training and ongoing operating costs also belong in the calculation.

The next decision becomes concrete: test whether the proposed workflow reduces handling time while preserving quotation quality. Record corrections, turnaround time and actual use. Then compare the result with the cost of continuing.

An assessment has done useful work when it makes that decision possible, including identifying when a simpler process change would achieve the same result.

08 What makes an AI business case fundable?

A fundable case gives a decision-maker enough evidence to authorise a defined next step. It does not imply guaranteed funding or guaranteed returns.

For an AI initiative, leadership should be able to see:

• The baseline: what happens today, how often, and with what measurable consequence.
• The proposed change: where the workflow changes and what remains a human responsibility.
• The benefit assumptions: the expected improvement, its value, and what must be true to realise it.
• The full cost: preparation, delivery, integration, adoption and ongoing support.
• The prerequisites and uncertainties: what is known, what is estimated, and what needs testing.
• The owner and decision point: who is accountable and what evidence will justify continuing, changing direction or stopping.

These are general criteria for reviewing a business case, not an itemised specification of Diagnostik’s deliverable.

The quality of the case lies in how well the argument survives scrutiny. A modest improvement with a measured baseline and a committed owner may be more investable than a much larger claim with no credible route to delivery.

09 From diagnosis to the first investment

Diagnostik puts the decision about priorities at the front of transformation. Iris helps surface the operational knowledge; the assessment focuses attention on what to fix first and what that improvement could be worth.

This reflects the business-first approach behind our method: understand the operating problem before committing to the technology intended to address it.

For leaders facing competing proposals, the useful next step is to make the choice explicit. Which constraint matters most? What is ready to change? What evidence would justify spending more?

10 Frequently asked questions

What is an AI opportunity assessment?

It is a structured examination of where AI could improve business outcomes, whether the required conditions exist, and which opportunities deserve priority. It connects processes, data and implementation feasibility to an investment decision.

How does it differ from an AI readiness assessment?

Readiness examines the foundations for adopting AI. Opportunity assessment connects those foundations to particular business problems and their potential value. A sound investment decision needs both perspectives.

Should a company fix all its data before assessing AI opportunities?

Begin by identifying the intended workflow and the data it requires. Assess the relevant gaps, then include the necessary preparation in the sequence and cost. An assessment can reveal that preparation is the first investment needed.

What does Iris do inside Diagnostik?

Iris is the AI agent that runs Diagnostik’s interviews. She is designed to surface wasted time, financial leakage, leadership concerns and questions about whether the available data can support AI.

What does Diagnostik aim to deliver in ten days?

Diagnostik’s proposition is a ten-day assessment across six dimensions, aimed at prioritising transformation opportunities and producing a fundable case. That timeframe concerns the assessment; deploying a subsequent operational solution is a separate undertaking.

Can an assessment recommend a change that does not involve AI?

A credible assessment should preserve that possibility. A process simplification, conventional automation or data improvement may address the constraint more directly. The recommendation should follow the business evidence.

11 Sources

• McKinsey, The state of AI in 2026: On the road to ROI. Self-reported survey findings; not evidence of Diagnostik’s performance.
• RAND, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024. Qualitative interviews with 65 experienced practitioners.
• Microsoft, AI Readiness Assessment.
• Microsoft, Evaluating and prioritizing an AI use case with ISV business envisioning. Guidance written for software providers.
• NIST, AI Risk Management Framework Core.

Talk to Embiggen about Diagnostik and establish where your transformation should begin.