AI Consulting: What You Actually Need (And What You Don't)

The AI consulting market is full of companies that will charge you six figures to write a PDF about your "AI transformation roadmap." Then they leave, and you still don't have a working product.

We think about this differently at Fovea. Here's what companies actually need when they say they need "AI consulting."

What Most Companies Actually Want

When someone reaches out about AI consulting, they usually mean one of these:

"We want to add AI to our existing product"

This is the most common one. You have a working app and you want to make it smarter — add a chatbot, generate summaries, automate decisions, surface predictions.

What you need: an engineer who understands both AI and your existing codebase. Not a strategy deck.

"We want to build a new AI product"

You have an idea for something AI-powered and need someone to build it — from architecture to deployment.

What you need: a small team (1-3 people) who can design the system, pick the right models, build the app, and ship it.

"We're using AI but it's not working well"

You built something, it kind of works, but it's slow, expensive, inaccurate, or breaking in production.

What you need: someone to audit what you have, identify the problems, and fix them. Usually this is prompt engineering, architecture changes, or better eval systems.

"We don't know where to start"

You know AI could help your business but you're not sure how.

What you need: a short (1-2 week) discovery engagement to identify the best use case, prototype it, and decide if it's worth building.

What You Don't Need

A 6-month strategy engagement

If someone wants to spend 6 months studying your business before building anything, run. The best way to figure out if AI works for your use case is to build a prototype in 2-4 weeks and test it with real users.

A massive team

Most AI features can be built by 1-3 good engineers. You don't need a team of 10 with separate "AI strategists", "data scientists", "ML engineers", and "AI ethicists." You need people who can do the work.

Custom model training (usually)

Most companies don't need to train their own model. Pre-trained models (GPT-4, Claude, Llama) with good prompting and retrieval-augmented generation (RAG) handle 90% of use cases. Custom training is expensive, slow, and usually unnecessary.

An "AI platform"

You probably don't need to buy an enterprise AI platform. You need someone to integrate AI into your existing stack. The best AI features are invisible — they're just part of the product.

How to Pick an AI Consulting Partner

Here are the questions that matter:

Have they shipped AI in production?

Not demos. Not proofs of concept. Production apps with real users. Ask for examples. If they can only show you slide decks and prototypes, keep looking.

Do they understand your stack?

AI doesn't exist in a vacuum. It needs to integrate with your database, your API, your frontend, your deployment pipeline. If the consulting team only knows Python and Jupyter notebooks, they'll build something that doesn't fit your infrastructure.

Can they build the whole thing?

The best AI consultants are full-stack engineers who also understand AI — not AI researchers who can't deploy a web app. You want someone who can build the UI, the API, the AI layer, and the infrastructure.

Are they honest about what AI can't do?

If someone promises AI will solve all your problems, they're selling you something. Good consultants will tell you when AI isn't the right answer, when a simpler solution works better, and what the limitations are.

What a Good Engagement Looks Like

Week 1-2: Discovery

Understand the problem and the existing system
Identify the best use case for AI
Build a quick prototype to validate the approach
Estimate costs, timeline, and expected accuracy

Week 3-8: Build

Design the architecture
Build the AI features
Write evals and tests
Integrate with existing systems
Deploy to production

Week 9-10: Handoff

Documentation
Knowledge transfer to your team
Monitoring and alerting setup
Runbook for common issues

Total time: 2-3 months for most projects. Not 6-12 months.

The Technology Stack Matters

We build with Go, Kubernetes, and Azure because they work well for AI apps:

Go is fast, deploys easily, and handles the orchestration layer well — managing prompts, calling models, handling retries
Kubernetes handles the bursty nature of AI workloads with autoscaling
Azure OpenAI gives you enterprise-grade model access with data residency controls

But the specific tools matter less than the approach. A good AI consulting partner adapts to your stack rather than forcing their own.

Red Flags

Watch out for:

"We need to build a custom model" — for 90% of use cases, this is wrong
"This will take 12 months" — most AI features can ship in 2-3 months
"We need to set up a data lake first" — you can start with the data you have
"Our proprietary framework..." — you'll be locked in forever
No production examples — they haven't done this before

How We Work

At Fovea, we're a small team that builds AI apps. We've shipped our own products — SignalOdds and IddaaLens are AI-powered sports prediction platforms we built and run ourselves.

We bring the same approach to client work:

Small team, fast delivery
We build the whole thing — not just the AI layer
We use our own stack in production, so we know what works
We're honest about what AI can and can't do

If you're thinking about adding AI to your product or building something new, let's talk. No strategy deck required.