A private AI consulting engagement for an ecommerce brand usually runs in four stages: a discovery of two to four weeks, a proof of concept on one workflow, a production build with access controls and testing in place, and a handover followed by close monitoring. A focused first system usually reaches production in eight to twelve weeks from kickoff. Brand Media Works runs this kind of engagement for ecommerce brands, and the stages below are the ones we would expect any partner to show you.
What are the stages of a private AI consulting engagement?
The table sets out each stage, how long it usually takes, what the brand should have at the end of it, and the decision it leads to. Durations are the ranges consultancies in the Brand Media Works competitor register publish for their own work.
| Stage | Typical length | What the brand gets | Decision at the end |
|---|---|---|---|
| Discovery | 2 to 4 weeks | Use cases ranked by value, a check of the data behind each one, and a running-cost model set against current AI API spend | Which one workflow goes first |
| Readiness and architecture | 3 to 6 weeks, often alongside discovery | A deployment design for the brand's own cloud account, the identity and access setup, and the integrations needed with Shopify, Klaviyo or the support desk | Where the model runs and which data it can reach |
| Proof of concept | 4 to 8 weeks | A working version of one workflow, tested against a set of the brand's own questions with a pass mark agreed in advance | Go to production, change course or stop |
| Production build | 8 to 16 weeks for a production pilot | A hardened system with single sign-on, permissions, audit logging and monitoring, deployed to the wider team | Which workflow comes next |
| Handover and monitoring | The first 90 days after launch | A runbook, a named owner, the test set and close monitoring of accuracy, usage and cost | Whether to run it in-house or keep a partner on call |
Within the production build, the work typically splits into a prototype of three to four weeks, hardening of four to six weeks and deployment of one to two. For retail and ecommerce, a single AI feature such as product recommendations, search, support replies or content usually ships in eight to twelve weeks. A suite spanning storefront, support and merchandising runs twelve to twenty weeks, with A/B test windows built into the rollout.
How soon can an ecommerce team start using it?
Sooner than the full timeline suggests. A private AI workspace, where staff use a model running in the brand's own environment, can be in use about four weeks after kickoff, including the security sign-off. Typical private AI and private LLM deployments take four to eight weeks, depending on how ready the cloud setup is, how identity is managed and how many systems need connecting. Each regulated workflow added on top usually takes another four to six weeks.
Workflows that need the model to learn the brand's own material take longer. A self-hosted assistant built on a custom knowledge base, such as product data, policies and past support tickets, usually takes eight to twelve weeks, including data preparation, model customization, integration and testing. On-premise or private-cloud LLM explains the deployment choices that drive those timelines.
Why do so many AI pilots never reach production?
Because the production step is larger than it looks. IDC research found that 88% of observed AI proofs of concept never make it to wide-scale deployment: for every 33 a company launched, only four reached production. Deloitte's August 2024 survey of 2,770 director- to C-suite-level leaders across 14 countries found the same pattern: 68% said their organization had moved 30% or fewer of its generative AI experiments fully into production.
One consultancy that builds these systems estimates that most organizations underestimate the production step by a factor of three to five. The gap is data pipelines, security hardening, monitoring and integration, none of which a demo needs. A private AI engagement that is planned to reach production from the start does three things in its first phase:
- Models the running cost. Price the model over a 12-month horizon using expected token volume, response-time needs and support effort. As a rule of thumb, a self-hosted model becomes financially competitive above about 2 million tokens a day, and below 1 million a day the API is usually cheaper.
- Puts the controls in the plan. Single sign-on, permissions and audit logs are scoped in discovery, so they are ready before the first wider rollout. AI data security on AWS covers what that includes.
- Agrees a stop point. Ask for a written point early in the build, around week three, where either side can stop if the proof of concept has not met its pass mark. One firm estimates that a solid discovery phase saves up to 30% of the total budget.
Where does the model run during the engagement?
In the brand's own cloud account, from the proof of concept onward, if the model will see customer data. On Amazon Bedrock, a fine-tuned model is a private copy of the foundation model, and the model providers have no access to the accounts where their models are deployed. An open-weight model running in the brand's account goes further, because no model provider is involved at all. OpenAI's fine-tuning platform is closed to new users, which makes an open-weight model the usual route when a brand wants a model tuned on its own catalog and customers. Amazon Bedrock vs a self-hosted LLM compares the two, and private LLMs on open-weight models describes the build.
What should a brand prepare before the first call?
Discovery moves faster when the brand arrives with five things:
- Three to five candidate workflows, such as support replies, product questions, catalog enrichment or segment analysis, each with the person who owns it.
- The systems behind them, such as Shopify orders, Klaviyo profiles and the support desk, and who can grant read access.
- Current AI spend and usage, including API bills and seat licenses, so the running-cost model has a baseline.
- A sample of real questions from customers and staff, which becomes the start of the test set.
- The person who signs off on security, because their approval is often what holds up launch.
A pilot is typically delivered by a cross-functional team of four to five people: a project manager, a solution lead, AI engineers and specialists such as a data engineer or a security consultant. On the brand's side, plan for the workflow owner, someone who knows the customer data, and the security approver. If the data behind the first workflow is spread across tools that disagree, a marketing data engineering step may come first.
What should the brand own when the engagement ends?
The code, the prompts, the test sets and any model weights trained on its data, plus the data map, access roles, log settings and runbook. How to choose a private LLM partner lists the questions to settle before signing, and what a DTC brand should ask a private AI partner about its customer data covers the data side.
Frequently asked questions
How long does a private AI consulting engagement take?
Discovery takes two to four weeks, and a focused first system usually reaches production in eight to twelve weeks from kickoff. A private workspace for staff can be in use in about four weeks. Larger programs that add several workflows run for months, usually one workflow at a time.
Can we stop after the proof of concept?
You should be able to. Ask for each phase to be agreed separately, with a pass mark for the proof of concept and a written stop point early in the build, so the brand only commits to production once the first workflow has proved itself.
Who from our team needs to be involved?
The owner of the first workflow, someone who understands the customer data and the systems it lives in, and the person who approves security. Their time is heaviest in discovery and at launch.