What eCommerce AI Actually Costs in 2026: Real Budget Ranges by Use Case

We reviewed a large number of articles about AI in eCommerce while researching this one. Not a single one published a budget figure. They all reach the same paragraph — implementation costs can be significant — and change the subject.

That is not caution, it is positioning. A vendor cannot publish honest costs without undercutting its own sales motion, and a consultancy cannot without shortening the discovery phase. So here are ours.

Six Projects, Costed

1. Product Data Enrichment

Extracting and normalising attributes across a large catalogue. The unglamorous one, and consistently the highest return, because it unblocks search, filtering, feeds and every subsequent AI project.

• Build: £15,000–£45,000 — pipeline, prompt design, validation, import tooling
• Inference: small. On a catalogue of tens of thousands of SKUs the model cost is usually a rounding error against the engineering
• Ongoing: a modest allocation to run it on new products, plus merchandiser review time
• What moves it: catalogue size much less than catalogue messiness, and the number of distinct attribute sets

2. Semantic or Hybrid Site Search

Replacing keyword search with something that understands intent. High commercial value because search users are your highest-intent segment.

• Buy (a search SaaS): £500–£5,000/month depending on catalogue size and query volume, plus £8,000–£25,000 integration
• Build: £40,000–£120,000, plus embedding and hosting costs
• What moves it: whether your attributes are clean enough for it to work — see project 1, which is a prerequisite rather than an optional precursor
• Honest guidance: buy, unless you have an unusual relevance requirement. This is a well-served market.

3. A Customer Support Agent

The project most often sold and most often disappointing, because scope discipline determines the outcome more than technology does.

• Buy and configure: £300–£3,000/month plus £10,000–£30,000 to integrate properly with order data
• Build: £50,000–£150,000 for something genuinely connected to your systems rather than answering from a help centre
• Running: per-conversation inference cost, which is small individually and worth modelling at your actual volume
• Ongoing: this one needs a permanent owner — evaluation, prompt maintenance, escalation tuning. Budget it as a part-time role, not a project
• What moves it: how many actions it may take. Read-only is a fraction of the cost of anything that changes an order.

The scope framework matters more than the budget here; we set it out in what an AI support agent must never touch.

4. Recommendations and Personalisation

• Platform-native: frequently included, frequently sufficient. Start here.
• Buy (a personalisation SaaS): £1,000–£10,000/month, sometimes revenue-share
• Build: £60,000–£200,000, and only justified with unusual catalogue dynamics or a genuine data advantage
• What moves it: traffic volume, because low-traffic stores cannot generate the interaction data these systems need, and no budget fixes that
• Honest guidance: the most over-bought category in commerce AI. Measure your native recommendations properly before replacing them.

5. Demand Forecasting

• Build: £40,000–£150,000 including the data engineering, which is most of it
• Ongoing: monitoring, retraining and a planner who trusts it enough to use it
• What moves it: data history and quality. Two years of clean sales data is workable; eighteen months of data with a replatform in the middle is not
• Honest guidance: beat a naive baseline first. If a simple moving average with seasonality is not being beaten in a backtest, the project is not ready.

6. Agent-Readiness and AI Search Visibility

Making your catalogue legible to assistants and agents. Newer, cheaper than it sounds, and mostly a data and markup project.

• Audit: £5,000–£15,000
• Remediation: £15,000–£60,000, overlapping heavily with project 1
• Ongoing: monitoring visibility, which is a marketing-operations line rather than an engineering one
• What moves it: how much structured data you already publish correctly

The Costs Nobody Quotes

Four lines that decide whether the project is still alive in eighteen months.

1. Evaluation infrastructure. You need a way to tell whether a change to a prompt or a model made things better or worse. Without it every change is a guess and quality drifts invisibly. Budget £10,000–£30,000 to build, and time to maintain.
2. Model version churn. Providers deprecate models. Your carefully tuned prompt will need revalidating on a schedule you do not control. Assume a re-evaluation cycle at least annually.
3. The human in the loop. Almost every worthwhile commerce AI project has one — a merchandiser reviewing samples, a support lead tuning escalations, an analyst checking forecasts. This is a permanent operational cost and it is left out of most business cases.
4. Peak-load inference. Your token spend scales with traffic. Model Black Friday, not an average Tuesday, and design a degradation path for when it is more expensive than the margin it protects.

A Business Case That Survives a CFO

Four components. Anything missing one of them will be sent back.

1. A baseline you measured before starting. Without it you cannot prove anything afterwards, and you will end up arguing about seasonality.
2. A conservative and an optimistic scenario, with the decision made on the conservative one.
3. Total cost of ownership over three years — build, run, evaluation, human time, and one re-platforming of the model layer.
4. A kill criterion. What result, by what date, would mean stopping? Projects with no kill criterion do not get killed; they get quietly starved, which costs more.

On measurement, the single most valuable discipline is a holdout. Route a proportion of eligible traffic or conversations past the AI system for the first quarter. It costs you a fraction of the upside and it converts revenue went up into the project caused revenue to go up, which is a different sentence in front of a board.

Where the Money Is Actually Wasted

• Building what could be bought. Recommendations and search are well-served markets. Build only with a genuine data advantage.
• Buying before the data is ready. A personalisation platform on a catalogue with 40% attribute coverage will underperform its demo and the vendor will not be at fault.
• Pilots with no production path. A successful pilot with no plan for integration, monitoring or ownership is a cost with no asset at the end.
• Optimising the wrong metric. A support agent that maximises containment while conversion falls is destroying value efficiently.
• Chasing the demo. The impressive capability is rarely the profitable one. The profitable one is usually attribute enrichment, which nobody demos.

If you want the numbers built against your own store rather than against a band, that is what our AI integration practice and AI and machine learning team do before a project starts. The same discipline applied to a platform build is in what a custom Node.js platform really costs.

Get the Numbers for Your Business

We will cost the shortlist against your catalogue, traffic and data, and tell you which projects are worth funding this year. Frequently that is one project, not five. Talk to our team; we reply within a business day.