Short answer

An AI project fails when the model is expected to settle decisions the organisation has not yet defined: what each entity is, who may see it, which data item is the official source and which indicator counts as valid. Before automating, it is worth mapping out the operating model: people, projects, documents, statuses, roles, permissions, reporting outputs and exceptions. That map reduces risk and gives the AI a verifiable foundation to work on.

Key points

  • An AI project fails when the model is expected to settle decisions the organisation has not yet defined, such as which data item is the official source or which indicator is valid.
  • Before automating, it is worth mapping the operating model (people, projects, documents, statuses, roles, permissions, reporting outputs and exceptions) so that AI works on a verifiable foundation.
  • A data model is also a model of responsibility: it makes explicit who creates records, who validates, who can see, who corrects and who decides on the output.
  • AI works best as a support layer, for classifying, summarising or spotting inconsistencies, and worst when it is used to cover up a lack of process.
  • The sign of maturity is not having lots of prompts but knowing which data can be used, within what limits, for which decision and under what human oversight.

AI is no substitute for internal definitions

Many pilots start with an attractive question: what can an assistant do with the organisation’s documents, emails or spreadsheets? The useful question usually comes earlier: what does each data item mean, and who has the authority to change it?

If the same person appears under different names, if a project has several codes, if dates do not mean the same thing in every department or if nobody knows which version of a spreadsheet is the right one, the AI model may produce fluent answers that are hard to defend.

The data model is also a model of responsibility

A data model is not just a table. It defines what things exist in the organisation and how they relate to one another: people, activities, case files, invoices, indicators, documents, permissions and decisions.

Once those relationships are made explicit, responsibilities emerge too: who creates a record, who validates it, who can see it, who corrects it and who uses the output to make a decision. Without that layer, AI is operating in an ambiguous environment.

Where AI fits best

AI works best as a support layer: classifying documents, summarising information, spotting inconsistencies, generating drafts, helping to write transformations or suggesting labels for human review.

It works less well when it is used to cover up a lack of process. If there is no agreed definition of beneficiary, project, intervention, status or outcome, AI cannot create an operational truth on its own.

A sign of maturity

A good AI diagnostic usually produces less immediate enthusiasm and better decisions. Some tasks are automated, others are redesigned and others are dropped because the risk or the cost does not pay off.

The sign of maturity is not having lots of prompts. It is knowing which data can be used, within what limits, for which decision and under what human oversight.

Frequently asked questions

Why does an AI assistant give convincing but unreliable answers about our organisation's data?

Usually because the underlying data is ambiguous. If the same person appears under different names, a project has several codes, dates mean different things in different departments or nobody knows which version of a spreadsheet is the right one, the model can produce fluent answers that are hard to defend. AI does not, on its own, fix ambiguous data, blurred responsibilities or poorly defined processes.

What should we review before launching an AI project on our organisation's data?

First map out the operating model: people, projects, documents, statuses, roles, permissions, reporting outputs and exceptions. You also need to define what each entity is, who may see it, which data item is the official source and which indicator counts as valid. That map reduces risk and gives the AI a verifiable foundation to work on.

Which tasks is AI genuinely useful for in an organisation, and which is it not?

AI works best as a support layer: classifying documents, summarising information, spotting inconsistencies, generating drafts, helping to write transformations or suggesting labels for human review. It works less well when it is used to cover up a lack of process. If there is no agreed definition of concepts such as beneficiary, project or outcome, AI cannot create an operational truth on its own.

What should we expect from an AI diagnostic?

A good AI diagnostic usually produces less immediate enthusiasm and better decisions. Some tasks are automated, others are redesigned and others are dropped because the risk or the cost does not pay off. The sign of maturity is knowing which data can be used, within what limits, for which decision and under what human oversight.