How to use AI for demand forecasting without a data science team
Most marketing directors and heads of ecommerce have the same quiet frustration: they know their demand forecasting is wrong, they know it costs them money, and they have no clear path to fixing it that does not involve hiring a team of specialists they cannot justify or afford.
The mistake most organisations at this stage make is believing that better forecasting requires a data science team first. It does not. What it requires is the right approach to the data you already have, the right tooling applied to the right problems, and someone with the commercial sense to know which forecast errors actually matter.
This article gives you a practical framework for deploying AI-driven demand forecasting without building a data science function from scratch. We will cover where to start, what the AI actually does, what it cannot do without your input, and how to know whether it is working. If you lead marketing, ecommerce or growth at a consumer technology business, this is written directly for you.
Why your current forecasting is probably worse than you think
Most businesses in the £500m–£1.5bn revenue range are forecasting demand using some combination of last year's actuals, a spreadsheet model built by someone who has since left and gut instinct dressed up as commercial judgement.
That is not an insult. It is the natural consequence of growing fast enough that your data infrastructure lags behind your trading complexity. You now have multiple channels, promotional mechanics, seasonal peaks, new product launches and shifting customer behaviour - and you are trying to model all of that in Excel.
The hidden cost is not the forecast error itself. It is the decisions that cascade from it: the excess stock you are sitting on, the out-of-stocks during peak periods, the media spend timed against demand that was not actually there, the markdown cycles you could have avoided.
A fashion ecommerce business turning over £800m, for instance, might carry an average of eight weeks of stock on fast-moving lines when four would suffice - not because of poor buying judgement, but because the forecast told them demand was higher than it was. The capital tied up in that excess stock is not a warehouse problem. It is a forecasting problem.
What AI-driven forecasting actually does differently
Traditional forecasting is essentially extrapolation. You look at what happened, you apply a trend line or a seasonal adjustment, and you project forward. It breaks down the moment the future does not look like the past - which, in consumer markets, is most of the time.
AI-based demand forecasting - specifically machine learning approaches - does something structurally different. It learns from a much wider set of signals simultaneously: historical sales patterns, promotional calendars, pricing changes, web traffic and search intent data, weather, competitor activity, even social sentiment. It does not need to be told what matters. It identifies the relationships itself, and it updates those relationships as new data arrives.
For a head of ecommerce running multiple product categories, this means a few concrete things. First, the model can distinguish between a demand spike caused by a promotion and one caused by an organic shift in customer behaviour - and treat them differently in future forecasts. Second, it can forecast at a much more granular level: SKU by channel by week, rather than category by month. Third, it does not require you to manually rebuild the model every time something changes in the market.
None of that requires you to have a data scientist in-house. It requires you to have clean enough data, the right platform and someone who can translate commercial context into model inputs - a role that often sits better with a senior analyst or a commercial finance manager than with a technical specialist.
What the AI cannot do without you
This is the part that most vendors will not tell you clearly. AI forecasting models are only as good as the signals you give them. The model cannot know that you are launching a new product category in Q3, that you are changing your promotional strategy, that a key wholesale partner is delisting you or that you plan to enter a new market. These are material inputs. Without them, the model will forecast the world as it was, not as it is about to be.
The practical implication is this: AI demand forecasting is not a hands-off system you install and ignore. It is a tool that amplifies your commercial knowledge when that knowledge is fed into it correctly.
The businesses that get the most from AI-driven forecasting treat it as a dialogue. The model generates a baseline. A commercial or marketing lead reviews the output, flags the assumptions it cannot see - "we are running a TV campaign in week six", "we are expecting supply constraints on this SKU" - and those inputs are incorporated. The model then recalibrates.
This is not a failure of the technology. It is the correct way to use it. A growth leader who understands their customer and their market is not made redundant by a forecasting model. They become significantly more effective because they are no longer doing the mechanical parts of the job.
How to start without a data science team
The practical path to AI-driven demand forecasting does not start with a technology procurement exercise. It starts with a data audit.
Before you can use AI to forecast demand, you need to understand what data you actually have, in what state and in what systems. For most businesses at this scale, the answer is: more than they think, less clean than they need.
A useful starting framework:
- Inventory your data sources. Sales by SKU and channel, order data, website sessions, promotional history, returns, marketing spend by channel and week. These are your minimum inputs.
- Assess data quality. How complete is the history? Are there gaps around peak periods? Is promotional data captured consistently, or does it live in someone's inbox?
- Define the decision you are trying to improve. Demand forecasting is not a single use case. Are you forecasting to improve buying? To plan media spend? To time promotions? To manage fulfilment? Each decision has different granularity requirements.
- Start with a single category or channel. Do not try to build an enterprise-wide forecasting model in one go. Start where the cost of forecast error is highest and prove the value there first.
- Get external support for the build, not the strategy. The commercial judgement - what to forecast, why, and how to use the output - should come from inside your business. The model build and deployment can and should come from outside.
That last point matters. The businesses that waste money on data science engagements are usually the ones who outsource the strategy alongside the build. You end up with a model that optimises for the wrong thing, built to solve a problem the vendor defined rather than one you actually have.
Knowing whether it is working
The final trap is measurement. Many businesses that deploy AI forecasting do not know whether it has improved their position because they never defined what improvement looked like before they started.
Set your baseline before you deploy. Measure your current mean absolute percentage error (MAPE) by category. Track the inventory decisions and stockout rates that flow from your current forecasts. Then measure the same metrics after the model has been running for a full trading cycle - including at least one promotional period and one peak.
The signal you are looking for is not forecast accuracy in isolation. It is the downstream commercial impact: reduced excess stock, lower markdown rates, fewer out-of-stocks, better-timed media spend. Forecast accuracy is the input. Margin is the output.
A consumer electronics business that reduces its average MAPE from 28% to 16% across its top 200 SKUs might save eight to twelve weeks of working capital - not because it hired data scientists, but because it applied the right tools to a clearly defined commercial problem.
The cost of waiting
Every trading cycle you run on a broken forecast is a cycle where you are leaving money on the table in ways that are almost invisible - not in a single catastrophic decision but in a thousand small ones that compound over time.
The good news is that you do not need to solve everything at once. A well-scoped initial engagement can define your data position, identify the highest-value forecasting problem in your business and produce a working model for a single category in eight to twelve weeks. That is a reasonable scope, a manageable cost and a proof point you can take to the rest of the business.
Rodan runs paid diagnostic engagements designed to do exactly this: assess your data readiness, define the right forecasting problem and give you a clear path to a deployed solution - without requiring you to hire a team first. If demand forecasting is costing you more than it should, that is a reasonable place to start.
