Example scenario, not a client project. The customer support team at a SaaS company got an AI assistant for answering tickets. After three months, 15% of agents use it. The diagnosis shows the assistant opens in a separate window, and the agents' corrections don't affect later suggestions. The first two fixes: suggestions right where the reply is written, and visible sources from the knowledge base.
Is this for you?
- You rolled out an AI tool, but only a small part of the team uses it.
- Users start, but don't finish the most important task.
- People check everything the AI did, so there are no savings.
- The team has several theories about why "it isn't catching on", and no evidence.
- Before the next investment, you want to know what to fix first.
What I do
- Digital product — expert review, usage data analysis, short conversations with users.
- AI pilot — how people really use it, where they work around it, what blocks it.
- Operational process with AI tools — whether the tool fits the work, or the work goes around the tool.
The AI layer
I check what MIT NANDA identifies as the main reason pilots stall — no learning and no fit with the work:
- Does the system learn from feedback — do people's corrections change anything?
- Do people know when they can trust it — can they see sources and confidence levels?
- Does it fit the workflow — is it where the work happens, or do people have to "go to" it?
- Who is responsible for the result — are people afraid to sign off on AI work?
- Do people prefer their own tools — and why? (MIT NANDA: 90% use personal AI tools.)
What you get
- A diagnosis report — a few pages: what isn't working, why, and what the evidence is.
- A map of where people drop off — along the task path.
- Causes, not symptoms — for every problem: process, interface, trust or data.
- A list of fixes by impact and effort — what to do this week, what this quarter.
- A next-step recommendation — fix, redesign or stop.
How it runs
- Access — the product, usage data, documentation, contact with a few users.
- Expert review — walking through the key tasks against proven criteria (including the Microsoft HAX guidelines for AI).
- Data and conversations — confirming hypotheses in the numbers and with people.
- Synthesis — causes and priorities.
- Presentation — with the team and decision-makers.
Roughly 1–3 weeks [to be confirmed]. A fixed-scope variant: the AI Workflow Opportunity Map for a live pilot.
How we'll know it worked
- The team agrees on the 3 main causes of the problem.
- The first fixes are rolled out within a few weeks.
- After the fixes, the share of people who finish the task or come back to the tool goes up.
What this doesn't cover
I don't assess the AI model's technical quality or the system's security. I look at how people work with it.
Visualizations
Where people drop off (sample artifact)
Sample artifact: where people drop off, and why.
Fixes by impact and effort (sample artifact)
Sample artifact: fixes ordered by impact, not by opinion.