Consulting services

Improve what isn't working

If people are spending too much time correcting or working around an AI tool, another prompt may not be the answer. Let's look at where the problem actually begins.

Your starting point

We already use AI, but something isn't working well.

This service fits an existing workflow that people can show and explain. It could involve poor answers, inconsistent quality, a slow handover or a tool that worked in a demonstration but struggles with everyday tasks. You don't have to know the cause before asking for help.

Questions we work through

Which failures keep happening, and under what conditions?
Are instructions clear enough for the task?
Does the source information support the expected answers?
Where does checking or rework remove the benefit?
Should the team fix, narrow, replace or stop using this approach?

How we work together

1. Define the problem

We agree which workflow to inspect and what is going wrong. You share examples of successful and unsuccessful work, along with what the team expected. We identify immediate limits on use if a problem needs attention before the review is complete.

2. Trace the failure

I examine the available inputs, instructions, source information, outputs and handovers. Where access permits, we try to reproduce the problem under agreed conditions. A result we cannot reproduce remains an open question, not a confidently named cause.

3. Compare possible changes

We look at changes that address the cause: clearer information, different instructions, better checking or a narrower task. A tool change is an option, not the default answer. I explain which recommendations are supported and which need a test.

4. Check the improvement

Within the agreed scope, we compare proposed changes against representative work and the original failures. You receive the findings, any remaining limits and recommended next actions. Implementation beyond the review is separately agreed.

What this could look like

A situation you might recognise

If an assistant gives different answers to the same policy question, I'd look at which documents it can access and whether those documents conflict. We could check how it selects sources, whether it identifies the version and what it does when the answer is unclear. A useful change might involve correcting the source information or asking the assistant to pass uncertain questions to a person. We'd then repeat the relevant checks instead of assuming that a more convincing answer proves the problem is fixed.

What you bring

Examples of the problem and what a correct result should look like
Relevant settings, instructions and approved source material
Access to the process owner and someone familiar with the technical setup
Existing test results or logs where available and appropriate to share

We'll agree how information is shared before you provide anything sensitive.

What you receive

A clear account of what was examined and what was observed
Likely or confirmed causes, with the distinction made explicit
A prioritised set of changes and the reasons for them
Test findings where testing is included, plus remaining questions and next steps

Keeping the work clear

This is not an emergency response service or a formal security audit. Findings depend on the evidence and access available. If the issue involves a possible data exposure or a specialist technical fault, your responsible team should handle the immediate response while the appropriate review is arranged.

How pricing works

A defined diagnostic review can be quoted at a fixed fee after an initial discussion. The quote depends on the workflow, access, available evidence and amount of testing. Corrective development is priced separately unless included. You can take the findings to your existing delivery team.

You'll receive a written scope, fee, payment terms and any applicable tax before the engagement is confirmed. Ongoing support is optional and agreed separately.

Let's talk about your work

Tell me what the tool is meant to do and where the results are falling short; don't send confidential logs in an initial message.

Discuss this service

Not the right starting point? See the other three choices.