Why most AI consulting fails inside real businesses.

Most AI consultants have never been inside a real business. They've been inside AI. That, in one line, is why their work breaks the day it meets your Tuesday morning.

I've spent the last twelve years not in front of a model card but on the operators' floor — building Magento storefronts for a €340M lingerie group across twelve country sites, running ERP integrations for a Florida beauty manufacturer, debugging the B2B pricing rules that decide whether a defense reseller gets paid, watching a fulfilment workflow go through twelve intermediate status changes before a customer in Ghana sees their package. That work taught me something the OpenAI cookbook can't. It taught me what a real business actually looks like underneath the slide.

A real business is not the diagram in the deck. A real business is the spreadsheet that runs the company that nobody admits is the spreadsheet that runs the company. It's the integration that's been duct-taped for nine years and nobody touches because it just works. It's the CFO who quietly overrides the dashboard because she doesn't trust the numbers, and she's right not to. It's the manager who always ignores the alert and the one who always over-acts on it. It's the fact that the real decisions happen in Slack DMs at 11pm, not in the workflow tool. The data is dirty. The processes are weirder than the docs. The org chart is half-true.

If you've never seen this, you can't build AI for it. You can build AI for the diagram in the deck — that's what a lot of consulting today is — but the deck is not the business. So you ship a pretty proof-of-concept that dies in week seven because the data flow you assumed doesn't actually exist. You ship a recommendation engine whose suggestions get ignored because nobody asked the manager what override authority she needs. You ship an “automated” workflow that secretly requires three humans to keep running. You ship architectures optimised for the slide, not for the Tuesday morning of an actual operator. Then you move on, and the operator inherits a thing that looks impressive and breaks under load.

The flip side of that is what happens when you have been inside.

When a national retail-chain owner walks in and says “tell me what to put on each shelf, in each store, in the country,” you don't go build a perfect multi-objective optimisation engine that takes nine months. You translate the vague pain into a concrete surface — per-store, per-shelf SKU recommendations on a handful of real signals — and you wrap the AI in an approvable dashboard, because you know the store managers will override your model and you know that override is data, not noise. Working proof-of-concept in seven days.

When a US wholesaler is burning days per new vendor onboarded, you don't propose an AI transformation programme. You see the actual bottleneck — humans hand-mapping inconsistent supplier spreadsheets into a 120-column ERP — and you give the LLM the messy work and the humans the approval queue for the edge cases. Effort drops 60 to 75 percent per vendor. Cycle time goes from weeks to hours.

The pattern is the same across both cases, and across every piece of AI we ship:

• Frame the pain correctly. This is the hardest part of the job and the most underrated. The vague brief is the brief. Translating it into a concrete surface is the work.
• Pick the right place for the LLM. And, just as important, pick the wrong places to keep it out of.
• Build for the override, not against it. The human stays in charge. That's not a limitation; it's the design.
• Ship in days, not quarters. So you can learn from the override behaviour while there's still a window to change anything.

You learn this by spending years inside real businesses. You do not learn it by spending a weekend with the OpenAI docs.

And we don't only consult. KCG ships its own AI to production. Skarb.ai is a GraphRAG family-archive app where users converse with their own memories — three languages, web and iOS, real paying users, Stripe billing. Engineered solo, ten weeks start to launch. It exists partly because we wanted it to exist, and partly because I don't trust an AI consultant who hasn't shipped one of these themselves.

Twelve years of Magento and ERP and B2B pricing logic and SAP integrations and warehouse rules and customer-support workflows isn't our legacy baggage. It's the school we went to. It's the school that taught us what AI is for and, just as importantly, what AI is not for. Most of the people pitching you AI right now did not go to that school.

“We don't build AI demos. We build the things operators trust on a Tuesday morning.”

That's the bet we're making with KCG. AI that works inside the business you actually have — not the slide-deck version. Built by people who learned what a real business is the hard way, before the AI conversation started.