GenAI for mid-career managers: what to learn to lead AI work
In this article (7 sections)
Related programme: the FDE for Professionals course (16 weeks (proposed)). See also open Forward Deployed Engineer jobs in India, updated daily.
A mid-career manager leading teams that build or adopt GenAI needs to learn enough to ask the right questions, not enough to write the code: how these systems fail, how quality is measured, what they cost to run, what data and approval issues they raise, and how to tell a demo from something ready to release. The fastest way to that understanding is a small amount of hands-on use plus a set of review questions you apply to every proposal. Managers who skip the hands-on part tend either to overtrust demos or to block everything.
Three mistakes managers make with GenAI projects
- 1Approving on the demo. A demo shows the system on examples its builder chose. It says little about the hundredth real case.
- 2Asking "is it accurate?" without asking "on what, measured how, compared with what?" The question invites a single reassuring number.
- 3Treating it as an IT rollout. GenAI features change behaviour when models or prompts change, and need ongoing evaluation, not a one-time sign-off.
Each is avoidable with the checklist below.
What to learn, in order
| Topic | What you need to understand | How you'll use it |
|---|---|---|
| How output is produced | Models generate plausible text; they can be confidently wrong | Set expectations with your own leadership; see explaining hallucinations to stakeholders |
| Grounding and retrieval | Answering from company documents, with sources | Ask whether answers can be traced to a source |
| Evaluation | Test sets, denominators, results by segment | Refuse proposals without an evaluation plan |
| Human review | Where people check and decide | Decide which steps keep a person accountable |
| Cost | Per-request cost, hosting, people to operate it | Compare against the value of time saved |
| Data and approvals | Personal data, security, governance review | Plan approval time into the schedule |
| Change | Models and prompts change behaviour | Require re-testing before changes go live |
Hands-on work worth doing yourself
Set aside a few hours across two weeks. Use only tools your organisation allows.
- Break a system. Take an approved assistant and ask it ten questions from your own area where you know the answers, including two it shouldn't be able to answer. Note how often it's wrong and how it sounds when it is.
- Read an evaluation set. Ask your team for the test cases behind a project. Read twenty. Are they realistic? Are the hard cases there?
- Estimate a cost. Ask for the per-request cost and expected volume, and do the multiplication yourself.
- Review a failure. Ask for three examples where the system got it wrong and what the team did about them.
This small investment changes the quality of your questions for months.
The review checklist
Use this when a team brings you a GenAI proposal or a release decision.
| Question | Good answer sounds like | Warning sign |
|---|---|---|
| What problem, for whom, and how is it done today? | A named process, current time or error rate | "Everyone can use it for everything" |
| How did you test it? | "On 200 real past cases, including 40 hard ones" | "We tried it and it looked good" |
| What were the results, by segment? | Numbers with denominators, per language or case type | One overall percentage |
| What does it do when it doesn't know? | Says so, routes to a person | Always answers |
| Where does a person review or decide? | Specific steps named | "Users will check it" with no design |
| What data does it use, and who approved that? | Data sources listed, privacy/security consulted | "We'll sort out approvals later" |
| What does it cost per month at expected use? | An estimate with assumptions | No estimate |
| How will we know if it gets worse? | Monitoring and sampled review | Nothing after launch |
| What happens when the model or prompt changes? | Re-run the evaluation before release | Changes go straight to production |
| What would make us stop? | A threshold agreed in advance | No exit criteria |
A proposal doesn't need perfect answers to every row. It needs honest ones, and a plan for the gaps.
Leading the team, not just approving
- Make evaluation normal. Ask for results in every review. Teams do what their manager asks about.
- Protect time for testing. Estimates that include only build time are incomplete.
- Reward reported failures. An engineer who finds a serious problem before release saved you an incident.
- Keep the decision owner clear. For each AI-assisted process, name the person responsible for the outcome.
- Think about your team's skills. Some routine work may shrink; plan how people will move to the checking, context and exception work that grows.
Talking to your own leadership
Senior leaders often ask either "why aren't we using AI more?" or "is this safe?". The checklist answers both: it shows you are moving, and that you are doing so with evidence. Report projects with three numbers — quality on the test set, time or cost change, and incidents — rather than general enthusiasm.
Should managers take a technical course?
Not necessarily. Many need structured literacy rather than engineering depth. If you have an engineering background and want to lead or deliver client-facing AI work yourself, look at FDE for Professionals, an advanced track still in its proposal stage, and talk it through with a counsellor to judge fit. For managers asked to approve or sponsor a pilot, how to propose a first GenAI pilot shows what a well-scoped proposal contains.
Pankit Kumar has 10 years in Data Science & AI, building and shipping production systems in regulated pharma and clinical environments. He is a freelance trainer at Boston Institute of Analytics, AnalytixLabs and Scaler, and has taught this material to thousands of working professionals.
This article is part of our FDE for Professionals programme — 16 weeks (proposed). An accelerated advanced pathway for IT professionals ready to own enterprise AI delivery.
Explore FDE for Professionals