Key takeaways
- The value sits in unglamorous systems. Product data, search, forecasting and order risk hold most of it, and a chatbot is rarely the best first build.
- At an illustrative $50m online retailer, the seven systems could add about $0.5m to $1.4m of operating profit a year once running.
- Build in the right order and plan around your peak season, and never run more than two builds at once.
- Models are the cheap part: under 1% of first-year cost. People are about 94%.
Most AI playbooks are written for the giants. They assume a data science team, a platform team and a budget with its own line, and they start with the most visible use case, usually a chatbot. Mid-sized online retailers and direct-to-consumer brands have none of that. They have a product and tech team of a few people, thousands of SKUs, customers in several countries and the same shoppers as the giants, who expect relevant search, instant answers and delivery that works.
This playbook shows how a team of three to ten people can close much of that gap with seven AI systems, built in the right order on one shared foundation.
Who it is for
Leaders of online retailers and direct-to-consumer brands with about $10m to $250m of online revenue, 2,000 to 100,000 SKUs and a product and tech team of three to ten people, selling on Shopify, BigCommerce, Adobe Commerce, WooCommerce, Salesforce Commerce Cloud, commercetools or a custom build. The main report is written for the CEO and CFO. The technical annex is written for the CTO.
Five messages for leaders
- The value sits in unglamorous systems. Product data, search, forecasting and order risk hold most of it. A chatbot is rarely the best first build, and at mid-sized volumes it often has the slowest payback.
- A $50m retailer can add $0.5m to $1.4m of operating profit a year. For an illustrative retailer with $50m of online revenue and 30,000 orders a month, the seven systems are worth about $0.8m to $1.7m a year in extra gross profit and costs avoided. After running costs, that lifts operating margin by roughly one to three percentage points.
- Build in the right order, and plan around your peak season. Start with clean product data, because search, AI shopping assistants and the service agent all rely on it. Never run more than two builds at once.
- Models are the cheap part. What you pay for the AI models themselves is under 1% of first-year cost at the reference retailer. People are about 94%.
- Own your data and your tests, and rent the rest. Models, search engines and helpdesks can be swapped whenever your own test sets and controlled comparisons show a new one is better. That is how a small team keeps pace with teams of hundreds.
The seven systems, in the order to build them
| # | System | Domain | Value a year* | Payback after go-live* |
|---|---|---|---|---|
| 1 | Content factory | Be found | $140k to $280k | 4 to 7 months |
| 2 | Search and discovery | Be found | $150k to $300k | 4 to 7 months |
| 3 | AI agent readiness | Be found | Option, not sized | Not applicable |
| 4 | Customer service agent | Serve and retain | $110k to $260k | 5 to 16 months |
| 5 | Demand forecasting | Deliver and plan | $170k to $370k | 4 to 8 months, live after the peak |
| 6 | Order risk | Deliver and plan | $140k to $250k | 5 to 8 months |
| 7 | CRM and retention | Serve and retain | $100k to $200k | 4 to 9 months |
*Illustrative, for a reference retailer with $50m of online revenue. The paper lists every assumption so you can replace it with your own numbers.
The order follows three rules. Build what the other systems rely on first. Time each launch around the peak season. Where there is still a choice, start with the faster payback.
What to do on Monday
- Name one accountable owner and one first use case, and write down the number it must move.
- Measure this month’s baseline for five numbers: the zero result rate on search, the cost per service contact, the days it takes to list a new SKU in every language you sell in, the return rate, and the loss per order from fraud, chargebacks and abuse.
- Build a 200-case golden set from your own search logs and service conversations, in every language you sell in, before you talk to any vendor.
- Start daily stock snapshots today. A forecast cannot learn from history you did not keep.
- Check that your robots.txt file and your CDN do not block the crawlers of AI assistants.
- Find out where consent for email, text and messaging is recorded in each market.
What is inside
Main report: the mid-market gap; where the value is; be found; serve and retain; deliver and plan; the shared foundation; what leaders get wrong; making it stick; and your first year, with a self diagnostic.
Technical annex for your CTO: the reference stack; where models run and how data is routed; model tiers and dated prices; evals and statistical gates; blueprint detail for each system; security and the OWASP LLM Top 10; law and regulation; and the cost and value model.
Download the full 55-page PDF above. It may be shared in full with attribution.