Your first 90 days as a Head of AI Engineering

Your first 90 days as a Head of AI Engineering

Simor Consulting | 28 Jun, 2026 | 07 Mins read

The first Head of AI Engineering at a company inherits one of three situations. Situation one: there is no AI team, no AI infrastructure, and the mandate is to build from scratch. Situation two: there is a small team running ad-hoc AI projects with no shared infrastructure, and the mandate is to professionalize. Situation three: there is an existing team with established practices, and the mandate is to scale or redirect.

Each situation demands a different 90-day plan, but all three share a common structure. The first thirty days are for listening and assessment. The next thirty are for establishing credibility through quick wins. The final thirty are for setting the strategic direction that will guide the next year. Rushing any phase — jumping to strategy before understanding the current state, or building infrastructure before earning trust — leads to plans that do not survive contact with reality.

This playbook covers all three situations. Adapt the specifics to your context, but follow the phase structure.

Prerequisites

Before day one, clarify three things with your hiring manager or executive sponsor.

First, the mandate. “Build AI capability” is too vague. Get specific: Are you responsible for a specific product’s AI features? For AI infrastructure across the company? For both? The scope of your mandate determines what you can and cannot commit to.

Second, the budget. Not the headcount — the total budget including infrastructure, tooling, vendor contracts, and training. You cannot plan without knowing your constraints.

Third, the success criteria. What will your executive sponsor point to at the end of twelve months to justify the investment? Revenue from AI features? Cost savings from automation? Risk reduction from AI governance? The success criteria shape every decision you make in the first ninety days.

Days 1-30: Listen and assess

Do not build anything in the first thirty days. Do not hire anyone. Do not select tools. Listen.

Week 1: Map the organization

Meet every stakeholder. Not just the people who will report to you. Meet the product managers who want AI features, the data engineers who own the data pipelines, the infrastructure team that provisions compute, the security team that reviews deployments, and the executives who funded your role.

For each stakeholder, ask three questions: What AI work is happening today? What is working well? What is broken? Write down the answers without filtering or judging. You will see patterns by week three that are invisible in week one.

Create a stakeholder map. For each stakeholder, document: their relationship to AI (producer, consumer, blocker, enabler), their priorities, their pain points, and their level of AI literacy. This map guides your communication strategy for the next twelve months.

Week 2: Audit the current state

If there are existing AI projects, audit them. For each project, document: what it does, who uses it, what data it depends on, what infrastructure it runs on, what model it uses, and how it is monitored. Pay special attention to projects that are in production — they are your baseline and your biggest source of technical debt.

Assess the data landscape. Where does the data live? How accessible is it? What quality problems exist? Data readiness is the single biggest predictor of AI success, and it is usually the dimension with the largest gaps.

Assess the infrastructure. Can the team train models? Deploy them? Monitor them? Scale them? Document what exists, what works, and what is missing.

Week 3: Identify pain points and opportunities

By week three, you should see the gap between what the organization wants from AI and what it can currently deliver. Categorize the gaps:

  • People gaps: Skills that are missing, roles that are unfilled, teams that are misaligned.
  • Process gaps: Workflows that are ad-hoc, decisions that are unowned, handoffs that lose context.
  • Technology gaps: Infrastructure that is missing, tools that are inadequate, data that is inaccessible.
  • Strategy gaps: Priorities that are unclear, success metrics that are undefined, roadmaps that do not exist.

Week 4: Deliver the assessment

Write a one-page assessment. Not a fifty-page deck. One page. It should contain:

  • Current state: three sentences describing what exists today.
  • Key gaps: the three to five most critical gaps.
  • Quick win opportunities: two to three things you can fix in the next thirty days.
  • Strategic questions: the questions you need answered before committing to a long-term plan.

Present this to your executive sponsor. Get their reaction. Their feedback tells you which gaps they consider most urgent and which opportunities they find most compelling.

Days 31-60: Establish credibility through quick wins

Quick wins serve two purposes: they deliver tangible value that justifies the investment, and they build trust with stakeholders who are skeptical about AI.

The criteria for a good quick win: it must be completable in four weeks or less, it must address a pain point that a stakeholder named in your assessment, and it must be visible to more than just your team. Fixing a data pipeline that nobody sees is valuable but does not build credibility. Automating a report that three teams manually produce every week builds credibility.

Pick two quick wins

Select two quick wins from your assessment. One should be a process improvement — something that makes the existing team more productive. The other should be a capability delivery — something that gives a stakeholder a capability they did not have before.

Process improvement examples: implementing a model registry so the team can track which models are in production, setting up automated testing for a critical data pipeline, or creating a shared prompt library that prevents duplicate work.

Capability delivery examples: deploying a prototype of a feature a product team has been requesting, building a data quality dashboard that operations has been asking for, or integrating a third-party AI API into an existing workflow.

Execute with transparency

Run both quick wins with weekly progress updates to stakeholders. Use a simple format: what was planned this week, what was completed, what is blocked, and what is planned for next week. Consistent, boring communication builds more trust than impressive demos with no follow-through.

Measure and report

When the quick wins are complete, measure their impact. Quantify it in terms the business cares about: hours saved, errors reduced, time-to-insight improved. Report these numbers to your executive sponsor and to the stakeholders who benefited. This is the evidence that funds your strategic plan.

Days 61-90: Set the strategic direction

With the assessment complete and quick wins delivered, you have the credibility and information to set a strategic direction.

Define the AI engineering vision

Write a one-page vision document. It should answer: What will AI engineering look like at this company in twelve months? What capabilities will exist? What infrastructure will be in place? What will the team look like?

The vision should be ambitious but achievable. “We will have a production AI platform that supports rapid experimentation, safe deployment, and continuous monitoring” is achievable. “We will be an AI-first company” is a slogan, not a vision.

Build the roadmap

Translate the vision into a quarterly roadmap. Each quarter should have three to five deliverables, each with a clear owner and success metric. The roadmap should balance three categories:

  • Platform: Infrastructure and tooling that makes the team more productive. Model registries, feature stores, monitoring dashboards, CI/CD for models.
  • Product: AI capabilities that deliver business value. New features, improved models, automated workflows.
  • People: Hiring, training, and process improvements that grow the team’s capacity.

A roadmap that is all platform produces a great foundation with no visible value. A roadmap that is all product produces visible value that collapses under technical debt. A roadmap that is all people produces a well-trained team with nothing to work on. Balance all three.

Get buy-in

Present the roadmap to stakeholders. Not as a fait accompli — as a draft that needs their input. Ask each stakeholder: Does this roadmap address your priorities? What is missing? What would you change?

Incorporate their feedback. A roadmap that stakeholders helped shape is a roadmap they will support. A roadmap that was imposed on them is a roadmap they will undermine.

Establish the operating rhythm

By day ninety, you should have established the operating rhythm that will sustain the team:

  • Weekly team standups focused on blockers and handoffs.
  • Bi-weekly stakeholder updates focused on progress against the roadmap.
  • Monthly architecture reviews focused on technical decisions.
  • Quarterly roadmap reviews focused on priority adjustments.

Document this rhythm and communicate it. People perform better when they know what to expect.

Situation-specific adaptations

Situation one (building from scratch): Spend more time on the assessment phase. You are not just evaluating existing work — you are deciding what to build first. Your quick wins should be infrastructure foundations: a model serving endpoint, a data pipeline template, a development environment. Do not attempt product features until the infrastructure exists.

Situation two (professionalizing): Your quick wins should focus on the highest-pain process gaps. If the team is manually deploying models, automate that. If there is no version control for prompts, add it. Professionalization quick wins are not glamorous, but they free up engineering time that compounds over the year.

Situation three (scaling or redirecting): Spend the assessment phase understanding why the existing approach is not sufficient. The team may be technically capable but strategically misaligned. Your quick wins should be alignment actions: a shared roadmap, a clarified mandate, a cross-team working agreement.

Common failure modes

Moving too fast in the first thirty days. New leaders want to demonstrate value immediately. Building or hiring before understanding the current state produces solutions to the wrong problems. The first thirty days of listening are not wasted time. They are the investment that makes the next sixty days productive.

Quick wins that are not actually quick. A four-week quick win that takes twelve weeks destroys credibility faster than not attempting one at all. Scope conservatively. Deliver early. A feature delivered in two weeks that does 80% of what was requested builds more trust than a feature delivered in twelve weeks that does 100%.

A roadmap without owners. Every roadmap item needs a named owner. “The team will build a model registry” produces nothing. “Sarah will deliver the model registry MVP by end of Q2” produces accountability.

Ignoring the people who were there before you. The existing team has context you do not have. They know which past initiatives failed and why. They know which stakeholders are difficult. They know where the technical debt is buried. Treat their knowledge as a critical resource, not an obstacle.

Next step

If you are starting a Head of AI Engineering role in the next thirty days, begin your stakeholder meetings before your start date. Ask your hiring manager to set up introductory meetings with the key stakeholders during your first week. The sooner you start listening, the sooner you can act.

Shipping a production AI system?

Find the control gaps before they turn into incidents. Take the AI Production Scorecard for a fast baseline across the seven layers, or book an architecture review and we will turn it into a hardening plan.

Similar Articles

Anatomy of an AI Incident: Post-Mortem of a Model Provider Outage
Anatomy of an AI Incident: Post-Mortem of a Model Provider Outage
19 Jun, 2026 | 09 Mins read

On a Tuesday at 2:14 PM, a major model provider began returning elevated error rates for a specific model endpoint. By 2:31 PM, a customer support platform that depended on that endpoint was producing

AI Rollback Patterns: When to Roll Back a Prompt, a Model, or the Whole Release
AI Rollback Patterns: When to Roll Back a Prompt, a Model, or the Whole Release
27 Jun, 2026 | 11 Mins read

Software rollbacks are well-understood. You deploy a new version, detect an issue, and roll back to the previous version. The rollback is atomic: the entire application reverts to the previous state.

The 7-step vector database selection checklist
The 7-step vector database selection checklist
26 Apr, 2026 | 06 Mins read

Most vector database selection failures come down to one mistake: picking the technology before mapping the workload. Teams benchmark embedding search speed on a curated dataset, pick the fastest opti

Build vs buy: a decision tree for AI infrastructure
Build vs buy: a decision tree for AI infrastructure
03 May, 2026 | 06 Mins read

Every AI infrastructure team eventually faces the same argument. One faction wants to build a custom solution because the commercial options do not handle their specific requirements. The other factio

How to design a prompt ops pipeline from scratch
How to design a prompt ops pipeline from scratch
10 May, 2026 | 06 Mins read

Prompt management in most AI teams starts the same way. One engineer writes a prompt, it works well enough, and the prompt gets committed to a config file. Three months later, there are forty prompts

The data quality scorecard: metrics that actually matter
The data quality scorecard: metrics that actually matter
17 May, 2026 | 06 Mins read

Most data quality initiatives fail not because teams lack tools, but because they measure the wrong things. Teams track hundreds of data quality metrics, generate dashboards full of green indicators,

A cost optimization framework for LLM inference
A cost optimization framework for LLM inference
24 May, 2026 | 06 Mins read

LLM inference costs follow a pattern that catches teams off guard. The first prototype costs almost nothing -- a few hundred dollars a month during development. The pilot scales to a few thousand. Pro

Migration playbook: batch to streaming in 5 phases
Migration playbook: batch to streaming in 5 phases
31 May, 2026 | 06 Mins read

The case for streaming is straightforward: data that arrives in minutes instead of hours enables decisions that were previously impossible. Fraud detection catches transactions before they clear. Pers

How to audit your AI pipeline for bias -- step by step
How to audit your AI pipeline for bias -- step by step
07 Jun, 2026 | 06 Mins read

Bias in AI systems is not a theoretical risk. It is a measurable property that can be detected, quantified, and mitigated at every stage of the pipeline. The teams that treat bias as an audit problem

The 30-day AI readiness assessment
The 30-day AI readiness assessment
14 Jun, 2026 | 07 Mins read

Organizations that skip readiness assessment before investing in AI tend to discover their gaps expensively. A financial services firm spent four months building a customer churn prediction model only

The RAG evaluation framework you'll actually use
The RAG evaluation framework you'll actually use
08 Jul, 2026 | 06 Mins read

Most RAG systems are evaluated with vibes. An engineer runs ten queries, eyeballs the results, and declares the system "working." Three months later, a customer reports that the system confidently ret

AI Enablement Programs: Building Organizational Capability, Not Just Technology
AI Enablement Programs: Building Organizational Capability, Not Just Technology
19 Mar, 2026 | 11 Mins read

A technology company built an impressive AI platform. They had GPU clusters, fine-tuning pipelines, evaluation frameworks, and a growing model registry. They opened access to any team that wanted to u

Building an AI Center of Excellence: Structure, Mandate, and Success Metrics
Building an AI Center of Excellence: Structure, Mandate, and Success Metrics
05 Jul, 2026 | 11 Mins read

Most organizations have attempted some form of AI initiative. Some succeeded and delivered measurable business value. Many failed and produced results that were technically interesting but did not mov

Prompt Engineering as Infrastructure: Version Control, Testing, and Deployment
Prompt Engineering as Infrastructure: Version Control, Testing, and Deployment
22 May, 2026 | 11 Mins read

Prompts are not prompts in the casual sense of suggestions or starting points. They are software. They take inputs, produce outputs, have failure modes that manifest in specific conditions, and requir