First in a series on what AI is doing to retail planning, where the role is going, and what we are building at Toolio for it.
Two years ago, engineers wrote code. Now they review it.
Open Cursor or Claude Code. Describe what you want, in plain English. The agent writes the code, runs the tests, fixes its own bugs, and comes back when it is done. The engineer reviews. The engineer ships. That is what writing software looks like at most well-resourced engineering organizations today. It is not a future-state vision. It is the current workflow.
Anthropic's CEO has said AI is writing "much of the code" inside Anthropic. OpenAI shipped a model in February that they used to help build itself. The companies closest to this are running their own engineering teams on this loop, not because it is on-trend but because it works.
Here is the part most retail planning leaders have missed.
This did not happen to software engineering by accident. The AI labs aimed at code first, deliberately, because they needed AI that could help build the next version of itself. Once they had it, they moved on to the next set of problems. Most of those problems are knowledge work. Most of them look a lot more like retail planning than they look like writing software.
Why software engineering is the leading indicator
The lazy version of this argument is "AI is general, it will hit everything eventually." A planning leader has heard that pitch enough times to tune it out. The structural version is more specific.
McKinsey recently drew a clean line through what AI does and what it does not. Intelligence is the work of gathering, reading, summarizing, and surfacing patterns. Judgment is the decision that follows. AI is now extremely good at intelligence. Judgment still belongs to humans, and we think that distinction holds. That is the right frame for reading what just happened to software engineering, and the right frame for reading what is about to happen to retail planning.
Software engineering and retail planning rhyme more than people think. Both are knowledge work that builds structured artifacts from messy inputs. Engineers turn a vague product spec into a working system. Planners turn vague trend signals, vendor constraints, and historical performance into a buy. Both jobs have a large intelligence layer (writing the boilerplate, pulling the report, reconciling the data, formatting the output) and a smaller judgment layer (architectural calls, pricing decisions, vendor reads). Engineers spent two years assuming most of their work was judgment. It was not. The intelligence layer was bigger than they realized, and AI absorbed it faster than they expected. There is no good reason to think planners are reading their own job differently right now.
Engineering is roughly twenty-four months ahead of retail planning on this curve. That is the leading indicator. What you saw happen to engineering between 2023 and 2025 is what is starting to happen to planning right now, and we think it will run faster, because the labs already solved the hard part. They built capable models. Building agents on top of those models for retail planning is a smaller engineering problem than building the models was in the first place.
Most planning organizations have not changed at all
This is the part that should be said out loud.
Most retail planning teams look exactly the way they looked three years ago. Excel-heavy. Same team structure. Same cycles. Same Monday morning recap that takes a planner half a day. The big middle of the market has not changed.
What has changed is what is happening at the leading edge.
We are starting to see real work coming out of a small number of customers and prospects who have leaned in personally. A VP of planning who runs his entire weekly exec recap in Claude in two minutes. A senior planner pressure-testing a merchant's buy by asking the AI to model what has to be true for the AURs and sell-through assumptions to hold, and getting back a list of three places the plan is fragile in five minutes. A merchant working through markdown timing and end-of-life choices as plain-English questions against live data, no model build, no data scientist required.
These are not common practice; these are early signals. Three or four people in a market of thousands. It’s happening now, with planners who six months ago had not used AI for anything serious. The thing that took engineers eighteen months to absorb between 2023 and mid-2024 is happening to these planners in eight or ten weeks.
It’s a trickle right now, but it’s happening.
The role shifts from executor to governor
The clearest way to think about what comes next is this. The planner stops executing the plan and starts governing the planning intelligence.
The work that used to define the day, building the plan, pulling the report, reconciling the export, formatting the deck, becomes work an AI does. The work that defines the day in the new model is different. Reviewing what the agent produced. Catching the cases where it is wrong. Making the calls the agent flags as uncertain. Owning the cross-functional conversations no software is going to handle.
The pyramid inverts. Today most planning teams have many execution-level analysts and planners and a few senior leaders. The agent-capable team is mostly senior judgment-level planners with software doing the execution. Smaller, more senior, harder to hire for.
The right analogy is the pilot and the autopilot. A modern commercial pilot does not fly the plane manually for most of a flight. The autopilot does. But you absolutely need the pilot. You need them to know when to disengage, when the autopilot is wrong, when something the autopilot has never seen is happening. That is what governing planning intelligence looks like. The retailers who think the autopilot means they do not need pilots are the ones who are going to have an expensive holiday season at some point.
The timeline, as we see it
Engineering ran a roughly two-year arc from "helpful assistant" to "I have handed over most of my work and my day looks completely different." We think retail planning runs a similar arc, starting now, and we think it runs faster.
Inside six months, agents are generating first-draft plans in production at the most aggressive retailers. Six to twelve months, planning team structures start changing in retail org announcements, and the merchant-with-agent model goes from experimental to operational at multiple mid-market brands. Twelve to eighteen months, retailers who did not start the transition are visibly behind on margin and inventory productivity.
The capability is here. The constraint is org adoption. Adoption usually moves faster than most leaders expect once two or three peer retailers in the same category run a season at materially better margin with materially smaller, more senior planning teams. At that point the boardroom conversation changes, and the timeline compresses on its own.
What we are building at Toolio
We have been building Toolio for the planning function described above. The view we have taken is that a planning interface is table stakes in this world and the intelligence is the product.
The Toolio MCP server is live this quarter. A planner can query Toolio's planning data from Claude or any other MCP-compatible AI tool, with planning-specific context layered on top so the answers come back right, not just confident. The exec recap, the plan pressure-test, the price question, all of those are running on top of the MCP server today, at customers who have leaned in.
The bigger piece is the agentic planning platform. We are building it deliberately and we are investing in it accordingly. The work is in the intelligence layer: the models, the retail-specific logic, the exception handling, the override tracking, the cross-category visibility a smaller, more senior planning team needs to govern a much larger planning operation than it could manage manually. This is where the next several quarters of engineering at Toolio are pointed, and it is what most of the rest of this series will get into.
What a planning leader should do this quarter
Three things, none of them dramatic.
Get yourself, personally, on the paid tier of Claude or ChatGPT. Not the free one. Use it for an hour a day, every day, for two months. The point is not to generate output. The point is to develop instincts about what the current models can and cannot do. Those instincts are the most valuable thing you can have right now, and you cannot get them by reading about AI. You have to use it.
Pick the two reports your team spends the most time pulling and try one of them in Claude this week. Use whatever planning data your team has access to. The first attempt will not be perfect. That is fine. Iterate. Most of what you will learn is in the iteration.
Ask one merchant or one cross-functional partner this question. What would you want from my team in two minutes that you currently wait two days for? The answers will tell you which agent to build first when you are ready to build one.
This is the first post in a series. More on the role, the technology, and what we are building at Toolio in the coming weeks.
