Enterprise AI Deployment: Beyond the Hype
The gap between a demonstration and a working system
There is a disconnect between what enterprises are told generative AI can do out of the box and the work required to make it useful inside a business.
A model can draft an impressive answer in a demonstration. That does not mean it understands a company's products, brand rules, approvals, data or tolerance for error. A plausible response is a very different thing from an enterprise-ready output.
I think the next stage of enterprise AI will be defined by a more realistic understanding of that gap. The opportunity is still enormous, but achieving it requires sustained adaptation.
Your subject-matter experts are essential
Engineers bring critical skills, but we cannot expect an engineer to become an expert in every business discipline the system supports.
In marketing, someone needs to understand what makes an output strategically sound, creatively appropriate and recognisably on-brand. Those judgements already exist within the organisation. AI deployment needs to bring them into the development process.
That involvement cannot end after the first training session. Products change, guidance changes and new exceptions appear. The people who know the business must help maintain the system as well as define it.
Build knowledge base squads
I use the term “knowledge base squads” for teams that combine engineering, knowledge management and domain expertise.
Their job is to turn a collection of documents and assets into information the AI can use reliably. They establish which sources are authoritative, resolve contradictions, structure material and define what a good answer should contain.
The technical approach may involve retrieval-augmented generation, curated reference material, fine-tuning or several methods together. These are different mechanisms. Giving a system access to a knowledge base is not the same as changing the model's underlying weights.
What matters is choosing the appropriate combination for the task and testing the result against the organisation's real requirements.
Own the knowledge and the operating logic
Third-party platforms can accelerate development. Enterprises still need a clear strategy for the information, permissions and workflows those platforms connect.
I would want to know who can update the knowledge base, how changes are approved, how information can move between systems and what happens if a supplier changes its pricing or product.
Ownership of that operating logic matters. It is difficult to create a distinctive capability if the business cannot inspect, adapt or meaningfully control the knowledge behind it.
This does not mean every enterprise should build its own foundation model. It means being deliberate about which capabilities to buy and which knowledge and decisions to retain.
Agents need shared foundations
A future in which specialised agents coordinate work sounds attractive. It also depends on very practical foundations: compatible information structures, clear identities, defined access and reliable handoffs.
An agent organising work across disconnected systems will encounter the same inconsistencies people do, often faster. Adding a coordinating agent does not make those inconsistencies disappear.
Start with a bounded process. Define its inputs, outputs, exceptions and owner. Establish a baseline, then measure whether the AI improves accuracy, completion time and the amount of human correction required.
The serious work starts here
For me, enterprise AI is an organisational change as much as a technical deployment. It needs people who understand the business, people who understand the technology and a structure that lets them work together.
The end of unrealistic expectations would be a positive development. It would let us focus on the work that creates value: trusted knowledge, tailored systems, useful evaluation and workflows designed around what the technology can actually deliver.