Hiring senior engineers in a market that’s split in two

In 2026, hiring senior engineers looks deceptively straightforward from the outside. Thousands of profiles on LinkedIn. Plenty of inbound applications. And yet many engineering managers at startups describe their senior searches as the hardest they have run in years: too slow, wrong candidates, offers that don't land.

The explanation is structural. The engineering market has split, and most hiring processes were built for a single pool.

Two markets, one job board

General software engineering job listings in the US are down roughly 36-49% from their early 2020 baseline, depending on the data source. ML and AI engineering openings over the same period are up 59%. These two curves have been moving in opposite directions since late 2023, and the divergence has widened through Q1 2026.

The result is two candidate pools with almost no overlap in practice:

Pool 1: AI-fluent senior engineers. Engineers who have built production systems with LLMs as load-bearing infrastructure: RAG pipelines, evaluation frameworks, cost management at inference, prompt versioning. Median time to fill through specialist channels: around 17 days, according to placement data from KORE1. Median base compensation: well above $200K.

Pool 2: General senior engineers. Strong backend, distributed systems, database, and infrastructure engineers who have not yet built production LLM features. Many are in the market from the 2021-2022 overhiring correction or from AI-driven restructurings; about 25% of March 2026 layoffs were attributed directly to automation. Larger as a candidate population. Taking 60-90 days to place.

The two pools rarely compete for the same roles. But they frequently compete for the same job descriptions.

What an AI-fluent senior engineer actually looks like

'AI experience preferred' does not describe Pool 1. By 2026, most engineers have used Copilot or a similar tool. That is not the signal.

Pool 1 engineers have shipped something that breaks if the model changes. That means real debugging experience with non-deterministic outputs. Practical judgment about when to add an evaluation step rather than trusting the model. Understanding of chunking and retrieval tradeoffs in RAG pipelines beyond knowing the acronym. And increasingly, cost engineering: the ability to make an LLM product 30-40% cheaper without measurable degradation in output quality.

Compensation reflects the supply-demand gap. Senior AI/ML engineers reached a median base of roughly $236,875 in Q1 2026, according to levels.fyi data tracked by 6figr. The AI premium over general senior engineering roles sits at 12-19%, documented consistently across Robert Half, KORE1, and Pin's 2026 compensation benchmarks. At the total-compensation level (base plus equity plus bonus), the gap is wider at senior and staff levels.

These candidates are almost always currently employed. Cold outreach on LinkedIn has a low response rate. Referral networks and specialist recruiters outperform general job boards significantly. When a Pool 1 candidate does engage, they close fast. The scarcity cuts both ways.

The general senior engineer: where they are and why the pipeline is slow

Pool 2 is larger and slower to fill, but not less capable. Many of the strongest engineers in the market right now sit in this pool, with deep knowledge of Postgres at scale, distributed consensus protocols, event-driven architecture, and database performance at real data volumes. They did not spend the last two years building LLM features. For a large class of engineering problems, that is exactly what you want.

They are in the market partly because the ratio of openings to candidates has shifted. The Bureau of Labor Statistics and private tracking firms both put general software engineering openings well below their 2022 peak. Not because of capability decline, but because demand contracted faster than supply did.

Time to fill for senior general engineers through standard channels runs 60-90 days. A realistic timeline from posting through full interview loop, offer, and notice period is closer to four months. Candidates who move faster are typically accepting contract work or getting meaningfully higher offers elsewhere.

How your job description picks your pool without you noticing

A job description is a filter, not a description. Candidates read it as a signal of what the team actually values, not a literal list of requirements.

'Must have experience with Python and system design, AI/ML experience a plus' sends two conflicting signals simultaneously. To Pool 1 candidates, it reads as: 'This is a general engineering role with an AI wishlist. Not worth my time at their comp band.' To Pool 2 candidates, it reads as: 'They want AI experience and I should probably deprioritise this.' The pipeline fills with candidates from neither pool cleanly.

Specificity is the fix. State the actual scope of AI work in the lead paragraph of the job description. Candidates self-select accurately when they have accurate information.

The compensation gap is real and candidates know it

You cannot offer Pool 2 compensation for Pool 1 requirements and expect Pool 1 candidates to stay. This sounds obvious. It is violated constantly.

The 12-19% AI premium is well established in 2026 and candidates track it. A senior engineer with production LLM experience has seen their offer data on levels.fyi. If your compensation band tops out at $180K base and you have put LLM engineering requirements in the job description, you will attract Pool 1 candidates who assume the band is negotiable, discover it isn't, and decline. Three weeks spent screening someone you were never going to close.

Offering Pool 1 compensation for a role that is genuinely general backend work is less common but also costly. Strong general engineers will accept it, feel under-levelled once the work starts, and leave. Retention suffers within a year.

Calibrate the compensation band to the role, not to the candidate you wish you could hire. Then write the description that matches that calibration.

Interview design for a split market

The interview is where pool mismatch becomes most visible, usually after you've spent four weeks getting to final rounds.

Pool 1 candidates typically underperform on classical systems design questions, not because they cannot think about distributed systems, but because they have been solving a different class of problem. Ask a strong Pool 1 candidate to design a feed ranking system and you will get a competent answer. Ask them to design an evaluation framework for a production LLM feature with a 2% regression budget, and you will get an interesting one.

Pool 2 candidates typically underperform on AI-specific questions, not because they cannot learn the material, but because they have not done it yet. Marking them down for this when the role does not require it wastes a signal that should be positive.

The practical fix: align interview content to the actual role requirements, not to what you'd ideally see. For general backend roles, evaluate systems design depth and operational reliability judgment. For AI-fluent roles, evaluate production LLM judgment, cost awareness, and evaluation design.

The worst interview outcome is not a bad candidate. It is a strong candidate from the wrong pool who leaves the process thinking your team does not know what it wants.

When the timeline goes wrong: three signals

If you are 30 days into an active senior search with no strong candidates in final rounds, the pipeline is telling you something specific. Three failure modes account for most stalled searches:

Candidates you like cannot pass comp. Your compensation band is calibrated for Pool 2, but the requirements in the description are pulling Pool 1 candidates. Fix: either raise the band to match the actual role, or remove the AI requirements that are not genuine.

Candidates who pass comp fall through on role fit. You are attracting Pool 2 candidates, but the role genuinely needs Pool 1 skills. Fix: adjust the comp band and change the sourcing channel before widening the search.

First-screen conversion rate is below 10%. The job description is ambiguous enough that you are drawing from both pools without filtering for either. Fix: rewrite before doing anything else.

Widening the search (posting to more job boards, increasing inbound volume) does not fix pool mismatch. It amplifies it.

What to check before you post next time

The two-pool divide is not closing. If anything, it will widen as AI tooling matures and the skills required to work with it diverge further from general backend engineering.

Before posting a senior engineering role, answer three questions: What percentage of the first 12 months of work involves AI or LLM systems in production? What is the compensation band, and does it match that percentage? What does a strong candidate from each pool look like in the first technical screen, and are you interviewing for the right one?

The companies hiring well in 2026 are the ones that decide which pool they're in before they write the job description. That decision drives everything else: sourcing, comp, interview design, and timeline expectations. Get it right at the start and the rest of the process runs on the logic you set. Get it wrong and you spend 90 days learning the same lesson expensively.