Most people searching for an AI consultant aren't really looking for an AI consultant, if I'm honest—they're looking at some messy corner of the business, the spreadsheet three people maintain and nobody trusts, or the inbox that never empties, or the thing that breaks quietly every time someone's on leave, and they're sitting there thinking surely this shouldn't still be this manual.
And that's usually the right instinct, which is worth saying because most people second-guess it. The useful version of AI consulting isn't a strategy deck with the word transformation on slide four. You bring the stuck thing, I turn it into a working system, and that's about the whole pitch.
The stuck thing usually has a shape
In ecommerce the same shapes keep turning up, and once you've seen them a few times you honestly stop being surprised by them—customers who can't find the right product so they bounce, a team that can't trust its own numbers so every meeting opens with an argument about which dashboard is lying, product pages that take forever to write so the catalogue's permanently half done, ad accounts quietly leaking spend through junk search terms, and two thousand customer reviews full of the exact language you'd kill for in your copy, just sitting there unread because nobody has a spare week to go digging.
AI is useful when it sits inside one of those and does the job—not the demo where a robot does something clever and everyone nods along, just the job.
- Product finder: a customer describes what they need in plain language and gets the right product back, pulled from verified catalogue data, not from a model's imagination.
- Content engine: a pipeline that writes catalogue pages, SEO fields and internal links in the brand voice, then checks its own work before anything goes live.
- Data copilot: your team asks plain-English questions about Shopify, Meta or Google Ads and gets answers straight from the actual accounts.
- Guardrail agent: a small machine watching one expensive failure mode—junk search terms, creative fatigue—that taps you on the shoulder before the spend gets out of hand.
- Daily briefing: the business read for you by 06:00, so your first decision of the day isn't hunting for the right tab.
Most brands don't need a grand AI transformation, they need one annoying workflow turned into something reliable—narrow enough that you can actually test whether it works, and useful enough that people notice the day it's missing. That last part is the whole quiet test, really.
If this is your problem
The way in is a $2,500 roadmap sprint: we sit down, name the bottleneck, work out the data it needs and the risk it carries, and scope the smallest build that actually helps. From there, bespoke builds are scoped to your data and start at $5,000—a product finder, a content engine, a data copilot, whatever the stuck thing turns out to be. The pieces that run every day and watch the business for you—the 06:00 briefing, the guardrail agents—live under the $3,000/mo Intelligence Retainer. If your problem doesn't fit neatly into a box, good. Naming it properly is the first job anyway. Full pricing is at /pages/ai-implementation.
The build matters more than the model
The model is the engine, but the build is everything around it—the car, the road rules, the brakes, the dashboard, and the small warning light that says no, do not send that one to a customer.
A product finder is the clearest example I've got. You can't just ask a model to be helpful, because helpful is exactly how you end up with a confident, fluent answer about a product that doesn't exist. I learned that the hard way, which is also why I trust the build and not the model. The version that's actually safe to put in front of a customer lets the model understand the words, then forces the answer back through verified product data, and if the data doesn't support the answer then the answer doesn't go out. That's the whole trick, and it's not a clever one.
Content's the same story. I don't want a model freestyle-writing 4,134 product descriptions and wandering off somewhere into vague catalogue mush—I want a pipeline with inputs and voice rules and product facts and SEO structure, plus a step where it checks its own work and a step where it fixes it, all before anything ships. That's the build we've run across 14-plus brands now. It's a bit boring to describe, and honestly that is the point: boring is what survives contact with a real catalogue.
So what do you actually ask for?
You don't need to turn up with a perfect spec, and please don't try to. Most good builds start with a sentence a lot closer to this is the part that keeps eating Thursday, and that's genuinely plenty to work backwards from. From there we figure out who uses it, what data it touches, where it's allowed to write and where it can only suggest, and what actually counts as a good answer. Then the one that matters most, which is what would be genuinely embarrassing if it got it wrong. That last question is usually where the real build reveals itself, because the thing you're most afraid of is the thing worth engineering around. It's also why server-side tracking and verified data end up sitting underneath most of what I build, not because it's exciting but because it's the floor.
If it touches your store, your ads, your CRM, the catalogue, the reviews or the inbox, it can probably become a system.
The best AI work doesn't make a business feel more futuristic, and I'd be a bit suspicious of anyone selling you that. It makes it feel less haunted—fewer tabs, fewer decisions you've already made ten times, fewer people stuck waiting on the one person who happens to know where the answer lives. That's what a good AI consultant should actually hand you: not a speech, but a working piece of the business that runs a little better tomorrow than it did today.