We have talked about data sovereignty time and again.
If you have not read it before, as a business owner who is implementing AI, it is a topic you ought to be aware of. The value you gain out of AI and your valuation both depend on this. There may be reluctant business owners who scaled beyond their imagination; but you can never build cash reserves reluctantly. So, pay attention.
Read these three articles if you have not:
- Moative Essay: Reliable AI vs. Coin Flip - click here
- Moative Essay: Hello CEO! Own Your Data. Own Your AI. Own Your Destiny. - click here
- Moative Essay: The AI Trap Door - click here
Intelligence is costly. Intelligence is not a monolith. Ergo, intelligence can be cheap.
There is only one right way of spending for intelligence in your business. There is a cap and the mileage of intelligence models. The best control measure on AI spend is on where you have the keys to decide which models to use for what jobs.
The problem with using OpenAI or Claude for everything is that they default to the best model that is pricey and may just not be necessary. For most business users, the image that flashes is the Claude or OpenAI chat interface. That’s not what I am talking about. It’s the API. At a certain scale, you are forced to use the API for your org. If I recall correctly, it's beyond 25 users.
Frontier models are like this prankster that steps one leg on the weighing machine when you choose to weigh. You may not like the weight (cost) you see but that’s because there is an extra burden you carry that you may not even be aware of.
The case for hiring ‘forward deployed engineers’
If you are new to this term and are amazed at ‘Big Tech’s’ ability to usurp terms from other industries, I am too. ‘Forward-deployed engineers’ is a term coined by Palantir, the intelligence company that breaks drug cartels and enemy armies through bar charts and histograms. They are so good at it that the seized drugs fire the imagination of their CEO Alex Karp. I am kidding, but am I?
Anyway, the street didn’t know if Palantir was a tech platform or a consulting company, until Palantir brilliantly coined the three-word phase ‘Forward-deployed-engineer’ to essentially avoid saying that they do ‘consulting.’ The rest is glorious history whose recent chapter has a story about Infosys, an Indian IT outsourcing company claiming that they have 300,000 FDEs.
Ignore the caterwaul. There is still merit in the idea that you need engineers to make general AI, your AI. But the ‘FDE' concept is an aircover for AI product companies (the old, meandering one whose contract you cannot get out of, who suddenly calls themselves AI for X, who now has FDEs).
50% of AI product companies that claim to deploy FDEs are saying, we want to learn your business. Our AI works half the time. We don’t know when we can deliver the ROI. But if we just give you keys to the login and move on, you won’t renew. So we are going to hang in with you, make it work. Along the way we take your data, context, and tribal knowhow and make our product better. You get “savings” and we build the next massive platform with IPO aspirations.
The game is self-serving but the play is interesting.
We already established that open-source models are, apples-to-apples, equal or better than OpenAI or Claude. We established that you need to pick the right models for the right workflows.
Here is where engineering matters.
Owning the engineering layer
If you use a cheaper model to do the job but an expensive model to validate the output, you can control cost. But the handover is an engineering job. If you have five software systems from which data has to come out and go back in, there has to be a central repository (data lake) and a common understanding (semantic data models) of what means what across all systems. That is engineering. Then there is the science of distilling all the wisdom that big models have and make them work on your corpus of knowledge. That job involves a lot of engineering.
When you buy an AI product, especially as a midmarket CEO, you should realize that most vendor-sold AI started to address some parts of your business but now have the ambition to be the platform that controls it all.
The tradeoff is that your vendor’s AI strategy is now yours.
What’s wrong?
Perverse incentives is what is wrong. Do you have flexibility to choose models or is the vendor choosing it and on top of that, adding their margins? If you think, RPA will do the job decently with some agentic re-tries, but your vendor is pointing to expensive and unproven ‘computer use’ tools that automate the actions with AI agents, do you have a recourse and an opt-out?
No.
The implied assumptions you are making are:
- My data and tribal knowledge are commodity and it is ok for someone else to own it
- My AI product vendor knows what is the best way to deploy AI for me
- The best way to cut cost is just implementing AI and not doing the due diligence on the architecture
- The vendor may never renegotiate and if they do, there is no cost to starting from scratch.
I am not saying these. Even Satya Nadella, of Microsoft, which is known for its anti-trust sojourns, is recommending not to handover the keys to AI vendors. What he did is interesting. He started an FDE company. Microsoft now consults on how to implement AI the right way. You may stay away but you get the idea.
FDE is a great concept. But there is no need for a lock-in to an AI product vendor’s FDE. Get your own FDEs. Incentivize them to do the right thing and build an architecture that is right for your business. Keep the data within your boundaries. Use local models that are tuned for your business.
You can have AI, get the savings, and spend less.
The cherry on that cake is that you will be far ahead of your peers in your AI approach and that unlocks its own valuation.
Central Takeaway:
- FDEs are great
- Be suspicious of FDEs from parties from vested interests
- Bring your own FDEs
Moative builds FDE pods and hands them over to you. You get a team that, from day one, knows that their interest lies in making you succeed, because you are their future employer. If you choose not to take on a team that you don’t know how to manage, you at least get a team that sets your architecture right, negotiates the build vs. buy with AI vendors for you, and earns their keep by saving you the bills.
Just reply and ask me how we have done this thrice in the last 18 months.