Published 2 September 2026
To get AI working in practice, a company needs exactly four things: a place where AI can safely reach your real business data, data that is findable (not perfect), one decision-maker who uses it every day, and only after that, for the heavy integrations, an engineer. In that order. Almost every company that stalls started in the reverse order.
This article is the complete shopping list: which people are being sold to you right now (and who you genuinely need, when), what has to be true about your data, and the fastest route from nothing to AI that answers real questions about your own business today.
Why "working in practice" is the actual problem
Start with the honest baseline, because the numbers on AI implementation are brutal.
| What goes wrong | The number | Source |
|---|---|---|
| GenAI pilots with no measurable P&L impact | 95% | MIT NANDA |
| GenAI projects abandoned after proof of concept | 30% | Gartner |
| Agentic AI projects predicted to be scrapped by end of 2027 | over 40% | Gartner |
| Companies that abandoned most AI initiatives in 2025 | 42% | S&P Global |
Notice what is not on that list: that the AI was not good enough. The models have been capable for years. What fails is always the same thing: the last mile between an impressive demo and real work, with real data, real people and a real decision riding on the answer. We have taken that gap apart in why AI projects fail and AI pilot to production.
The market has caught on. The Dutch financial daily FD reported in late August that Deloitte, Accenture, Capgemini and McKinsey are all hunting AI specialists whose entire job is closing that last mile. That is the clearest proof yet that "getting AI to work" has become the scarce resource, not the AI itself.
The job titles you see everywhere (and when you actually need each one)
Search "hire an AI specialist" and you get a forest of titles. Here is the translation, stripped of recruiter language.
| Title | What they actually do | When you need them |
|---|---|---|
| Forward deployed engineer (FDE) | Builds AI solutions and sits at your table; carries AI onto the work floor | Complex integrations into core processes; scarce and expensive, see the AI talent auction |
| AI engineer | Builds applications on top of existing AI models: integrations, workflows, agents | Once AI has to talk to your systems (ERP, CRM), not just your documents |
| Machine learning engineer | Trains and tunes custom models | Almost never in a mid-sized company; only with unique data at scale |
| MLOps engineer | Keeps running models reliable, secure and affordable in production | Only once something is in production that a team depends on |
| Data engineer | Connects data sources and keeps them clean | When your data is scattered across many systems and joining them is unavoidable |
| AI architect | Designs how AI fits into your overall IT landscape | Several AI initiatives running in parallel in a larger organisation |
| Prompt engineer | Writes and tests instructions for AI systems | Rarely a separate role; it is a skill embedded in the others |
| AI consultant / AI transformation lead | Advises on strategy, policy and change | Useful for buy-in and governance; usually builds nothing |
| Chief AI Officer | Owns AI at board level | Large organisations; in a mid-sized company, that is you |
Two things stand out when you read that list honestly.
One: most of these roles are only needed after AI is already doing something for you. An MLOps engineer with nothing in production is a gym membership without a gym.
Two: the role everyone is fighting over, the forward deployed engineer, is at its core not a person but a capability: bringing AI to where the work happens, grounded in the company's real data. And for the first and most important slice of your business, that capability is no longer reserved for scarce staff.
What you actually need: the list of four
Cross out the vacancy texts and four things remain.
1. A place where AI can safely reach your real data
The main reason AI answers stay generic and useless: the AI does not know your company. The difference between a toy and a working instrument is access to your own numbers, contracts, minutes and quotes. Which means the first decision is a security decision: that data does not belong pasted into a public chatbot. Your team is already doing it, policy or not: at over 90% of companies, employees use personal AI tools for work, mostly out of sight. So the choice is not whether AI touches your data, but whether that happens in a controlled place you manage, or uncontrolled in the shadow AI economy.
2. Data that is findable, not data that is perfect
The classic advice: first a data warehouse programme, then AI. For the questions an executive team actually has, that is the wrong order. Modern AI simply reads your existing files: the spreadsheet, the PDF, the contract, the report. "AI-ready" for your first year mostly means: the documents that matter, in one findable place. What that does and does not involve is in AI data readiness for executives.
3. One decision-maker who uses it daily
This is the item every procurement list skips, and the one that makes the difference. AI does not land through a rollout; it lands through a habit: someone with a real decision on their desk asking real questions every day. What is moving in my margin? What do these hundred pages actually say? Which customer is quietly slipping away? That is why AI adoption works as rhythm and ownership, not as a project. And why a working AI implementation starts at the top, with the AI-native executive, not with a working group. The labour market rewards the same thing: 66% of leaders say they would not hire someone without AI skills.
4. Only then: the engineer
There comes a point where you genuinely need build capacity: when AI has to enter your order flow, touch your ERP, or run part of a process on its own. Then you hire deliberately, from the table above, and you know exactly what for. That is the big difference from starting the other way round: an executive who first works with AI on their own data can no longer be sold a pilot that dies in a drawer, and buys engineering for a proven need instead of a hope. The per-workflow trade-off is in build vs buy AI.
The fastest route to good AI in practice
Line up the four and the fast route writes itself.
Day one. Pick the place: an AI assistant that runs on your own laptop and can reach your files there, so your data never has to leave your own machine. No procurement, no IT project.
Week one. Feed it: put the documents that carry your decisions (numbers, contracts, reports, quotes) in that one findable place. Imperfect is fine; findable is the bar.
Week two and onward. Build the rhythm: at least one real question every working day that you would otherwise have put to an employee, an adviser or a report. This is where the habit forms that pilots never produce, and where you discover for yourself which integration is worth hiring an engineer for one day.
None of these steps needs an AI team, a year-long data project, or a place in the queue at a consultancy. That is what "fast" means here: days to weeks, because the first implementation step is not a build step but a way-of-working step.
What this route does not solve
Honesty belongs on the shopping list. A personal AI assistant on your laptop does not write custom software, does not automate your production line, and does not replace an integration programme in which AI has to talk to your ERP. That work still needs the engineers in the table, scarce or not. What this route does do: make the thinking and analysis work most executive teams are waiting for available today, and teach you precisely what to hire that scarce engineer for later.
That is the idea behind AI Board:
AI Board is your personal AI assistant that makes you AI-native: it runs on your own laptop, grounded in your company's data, and gets sharper as your knowledge grows. Private by design, reasoning at CEO, CFO and CTO level, so you ride the AI wave instead of swimming behind it.
Frequently asked questions
What do you need to implement AI in a business?
Four things, in this order: a place where AI can safely reach your real company data (not a public chatbot), data that is findable in one place (not perfect), one decision-maker who asks it real questions daily, and only then, for integrations with systems such as your ERP, an AI engineer. Most failed implementations started in the reverse order: technology and staff first, usage last.
How do I get AI working fast in practice?
Start where no building is required: an AI assistant on your own laptop, grounded in your existing documents, used by one decision-maker with a daily question rhythm. That is up in days, not months, because the first step is a way-of-working step rather than an IT project. Integrations and automation come afterwards, for needs you have proven by then.
Which AI specialists exist and which one do I need?
The most requested titles are forward deployed engineer, AI engineer, machine learning engineer, MLOps engineer, data engineer, AI architect, prompt engineer and AI consultant. For the first working use of AI you need none of them. An AI engineer or forward deployed engineer becomes relevant once AI has to talk to your business systems; a machine learning engineer almost only if you train your own models, which rarely applies to mid-sized companies.
Does my data need to be in order before I start with AI?
Findable, not perfect. Modern AI reads your existing files (spreadsheets, PDFs, contracts, reports) without a data warehouse programme. The bar for your first year is that the documents carrying your decisions sit in one place your assistant can reach. A large data project before first use is, for most companies, the most expensive way to never start.
Why do so many AI implementations fail?
Not because the AI falls short, but because the last mile is missing: from demo to real work, with real data and a real user. MIT measured that 95% of GenAI pilots have no measurable impact; Gartner predicts that over 40% of agentic AI projects will be scrapped. The common thread is always building without a daily user who owns a real decision, and without access to the real company data.
The consultancies are fighting over the people who get AI working; you can bring the working part in-house today. See what that looks like on your own laptop →, read how your company data becomes one brain, or start with the news that made this urgent: the hunt for AI specialists.