The Head of Applied Science Nobody Needs to Fill Anymore

The Head of Applied Science Nobody Needs to Fill Anymore

A Central European e-grocer — operating across five countries, serving over a million customers, offering twenty thousand SKUs delivered in fifteen-minute windows — is hiring a Head of Applied Science. The job is to unify forecasting and optimization into a single function that decides how much inventory to buy, how many pickers to schedule, where each product sits in the warehouse, and which van carries which bag. The posting proudly notes that there is "no committee between your team and deployment — a solid backtest is all you need to ship."

Read that again. They have removed every human gatekeeper between a data scientist and production. The only remaining friction is the human who decides which model to build, which decision to automate next, and whether the backtest is solid. That is a remarkably thin job description for a human. It is, almost word for word, the system prompt you would give to an orchestration agent.

Let us walk through how AI replaces this role at a fast-commerce company operating exactly this kind of automated decision pipeline.

What the role actually does

The Head of Applied Science at a rapid-delivery grocery company is not doing science. They are doing prioritization and validation. Their team already runs automated production pipelines. Models already make daily decisions across purchasing, fulfillment, workforce scheduling, and commercial optimization. The role's core responsibilities, per the posting, are:

  1. Strategic roadmap and prioritization — deciding which decision to automate next.
  2. Business outcomes ownership — setting targets with cell leads and tracking financial and operational impact.
  3. Technical standards and architecture — presumably choosing frameworks, model families, evaluation protocols.

That is three jobs: a product manager for ML systems, a business analyst with domain expertise, and a chief architect. Two of those three are already automatable today. The third is automatable in a way that is genuinely embarrassing for the person currently drawing the salary.

What automates first: the roadmap

The strategic roadmap question — "which decision should we automate next?" — is an optimization problem itself. An AI agent solves this better than a human because it does not get tired, does not play politics, and does not have a favorite problem.

You build an agent that ingests the company's operational data and produces a ranked list of automation candidates. For each unautomated decision (currently made by a human or a heuristic), the agent estimates: the frequency of the decision, the cost of getting it wrong, the availability of training data, the expected lift from a learned model versus the current baseline, and the engineering effort to deploy. It then ranks by expected value divided by effort. That is the roadmap. The human Head of Applied Science was doing this in their head, poorly, with a spreadsheet, in a meeting.

The prompt is straightforward: "Given this warehouse's operational logs, identify the top five decisions currently made by humans or static rules that would yield the highest financial impact if replaced by a learned model. For each, estimate the data availability, the expected accuracy improvement, and the deployment complexity. Rank by ROI." An agentic workflow with access to the data warehouse, the codebase, and the production model registry executes this end to end. It even writes the JIRA tickets.

What automates second: the backtest gate

The posting says a solid backtest is all you need to ship. That means the entire quality gate — the thing standing between a data scientist's notebook and a production model affecting five countries — is an evaluation procedure. An AI agent runs that procedure more rigorously than any human.

You build an evaluation agent that, upon receiving a pull request containing a new model or a parameter change, automatically: spins up the backtest environment, runs the model against historical data across all five markets, checks for regression on key metrics (shrink, fulfillment cost, on-time delivery, working capital), validates that the backtest protocol itself was not leaked or gamed, runs counterfactual simulations, and either ships the model or returns it with an annotated critique. This is not science fiction. This is a CI/CD pipeline with an LLM in the loop writing the evaluation logic and interpreting the results.

The Head of Applied Science was the person who supposedly reviewed these backtests and gave the nod. But if your quality gate depends on a human glancing at a backtest and saying "looks fine," you do not have a quality gate. You have a ritual. And rituals are the first thing AI eliminates, because they are the easiest thing to specify and the most embarrassing thing to admit you were doing.

The company already says its engineers use AI coding agents daily. They are telling you, in the job posting, that the humans are already the slow part.

The pipeline that replaces the function

Here is the concrete architecture:

A decision registry — a catalog of every operational decision the company makes, from how much sourdough to bake for the Prague warehouse on a Tuesday, to how many vans to route through Bucharest at 5 PM. Each entry includes the current decision-maker (human, heuristic, or model), the frequency, the cost of error, and the data available.

A candidate generator — an LLM-based agent that continuously scans the decision registry for automation opportunities, drafts model specifications, and writes the training code. With tools like the coding agents the company already uses, this agent can produce a candidate model, a backtest harness, and a deployment manifest in a single pass.

An evaluation agent — the automated backtest gate described above. No committee. No human nod. Just metrics, counterfactuals, and a ship-or-reject decision.

A business tracker — an agent that monitors deployed models in production, tracks the financial and operational impact the posting says the Head of Applied Science is responsible for, and flags regressions or drift. When a model starts underperforming — say, the avocado forecast in Vienna starts drifting because a supplier changed — the tracker opens a ticket for the candidate generator to retrain.

The Head of Applied Science is the human who was supposed to sit between these components and make judgment calls. But the judgment calls are themselves specifiable: if expected ROI exceeds threshold and backtest passes, ship. If production metrics regress beyond tolerance, rollback. The human was the policy. The policy can be code.

Where it breaks

Here is where it does not automate cleanly.

The first break is vendor and partner dynamics. A large part of what a senior science leader does at a grocery company is not in the JD: they sit in rooms with the head of purchasing and negotiate whether the forecasting model's recommendation to cut a supplier's order volume by thirty percent is politically survivable. They sit with warehouse operations leads who do not trust the model and want a human to blame. These are trust problems, political problems, and organizational problems. An AI agent does not attend your steering committee and absorb the blame when the model says to discontinue a popular product line and the commercial team revolts.

The second break is novelty. When a pandemic hits, or a supplier goes bankrupt, or a competitor enters a market with aggressive pricing, the distribution shifts in ways that no backtest captures. A human scientist with domain intuition can say "this model is useless now, we need to throw it out and reason from first principles." An AI agent can detect the drift, but knowing what to do about it — knowing that the right response to a supply chain shock is to switch from optimization to rationing, or that the right response to a competitor's promotion is a loss-leader strategy — requires a kind of strategic judgment that does not live in the training data because the event has never happened before.

The third break is the social function of expertise. The Head of Applied Science exists partly so that the CTO can point to someone and say "this is the person responsible for our models being right." That is a liability sponge. When the model over-forecasts perishable inventory and the company writes off six figures in spoiled goods, someone gets blamed. That someone is currently a human with a title. Replacing the blame sponge with a system prompt does not resolve the organizational need for a throat to choke.

The verdict

The company is hiring a human to do prioritization, validation, and stakeholder management — and calling it science. The science is already automated. The pipelines run themselves. The coding agents write the models. The backtests gate the deployments. What remains is a person who picks which problem to solve next and then explains the results to other people.

That is a product manager with a PhD and a tolerance for meetings. And the product management part — the prioritization, the roadmap, the impact tracking — is exactly the part that AI agents are best at, because it is the part that is most specifiable.

The parts that are not specifiable — the politics, the novelty, the blame — those are not science either. Those are the cost of having a human in the loop at all. And the company that figures out how to automate the science and absorb the blame without a human title-holder will have a five-country decision engine that runs with no Head of Applied Science at all.

The job posting is a confession that the last human bottleneck in the model pipeline is the person who decides what to model next. That is a loop. And loops close.