Every AI vendor is currently selling autonomy, while the actual delivered value sits in a much narrower set of patterns. By my count there are six patterns where AI consistently delivers what it promises, and three more that get sold as AI but tend to disappoint in production, regardless of how impressive the demo was. Knowing which is which changes the questions you ask before signing a contract, and it changes how you read what your organisation already owns.
This piece is the taxonomy I have ended up using when I think about AI deployment with clients. It is not exhaustive, but it covers what most organisations actually get value from when they deploy AI well, alongside the patterns that get pitched aggressively and tend to underdeliver in ways that only become visible six months in.
The pattern underneath
Before the list, the principle that organises it. AI augments human leverage on cognitive work, and the value it delivers depends almost entirely on where the human sits in the workflow. Every pattern that delivers maps to a position relative to a human: before them, preparing material; with them, augmenting reasoning; after them, executing intent; around them, managing the periphery. The patterns that disappoint are the ones that try to remove the human from the decision point itself, which is precisely the part of the workflow where judgement, accountability, and the ability to course-correct matter most.
This frame matters because it explains why the patterns look the way they do, and because it gives you a quick test to apply when something is being pitched to you. If the use case removes the human from a decision they would normally own, the pattern is almost certainly being oversold.
The six patterns that deliver
Advisory. Reasoning over a curated knowledge base in conversation, where the agent answers questions from a body of knowledge the organisation has deliberately assembled. This scales because one knowledge base serves many users, and it is where most “AI for your company” pitches deliver real value when they are done properly. The discipline here is in curating and governing the knowledge, not in the AI itself. The organisations that succeed treat the knowledge layer as the asset and the AI as the interface, and the ones that fail treat it the other way around.
User-spreadsheet. AI as a personal data manager, maintaining structured records on behalf of one user, accessed through conversation rather than a user interface. Personal training logs, personal CRM, personal time tracking, personal nutrition. The spreadsheet metaphor is precise: local, flexible, ungoverned, owned by one person. It is powerful for personal productivity and dangerous when mistaken for an enterprise system, because it has none of the controls that an enterprise system requires.
Synthesis. Compression of large input into structured output: meeting transcripts into action items, fifty-page documents into one-page briefs, email threads into summaries. The value is information compression with judgement applied, and it is probably the highest-value AI use case in enterprise today. Most of it is currently delivered through plain ChatGPT or Claude rather than through governed platforms, which is itself a signal worth paying attention to about where the value actually sits.
Generation. Draft production from a brief, for human review. Marketing copy, first-draft contracts, code, slide decks, images. The AI lowers the cost of starting, and the human does the finishing. The pattern only works reliably when the human stays in the loop for final review, and most of the visible generation failures over the last two years have come from organisations that quietly stopped doing the review because the output looked good enough most of the time.
Curation. The agent watches a stream and surfaces items requiring attention. Email triage, calendar conflict detection, code review bots, anomaly flagging in logs. The initiative belongs to the agent rather than the user, which is what makes it powerful and what makes it dangerous, because the risk of alert fatigue is real and often arrives faster than anyone planned for. Where this works well is in replacing the cognitive load of constant low-grade monitoring with a single trustworthy signal that the human acts on.
Creative collaboration. Brainstorming, ideation, exploring variations. This is the use case where non-determinism is the feature rather than the bug. You want surprise, you want multiple plausible answers, you want to be pushed off the obvious path. Traditional ML is the wrong tool for this. Large language models are the right one, provided you remember that breadth of options is not the same as actual creativity, and that recognising what is good still requires a human who knows what good looks like.
Three patterns that disappoint
Agentic AI as currently marketed. Strip away the framing and most agentic products are advisory with tool calls, or generation that writes to systems, with the word “agent” doing more work in the pitch deck than in the code. Every successful agentic product I have seen has a narrow scope, a human at the decision point, and tool access constrained to a specific domain. The pattern is real and often valuable, and it is entirely different from the autonomous behaviour the marketing implies. Buyers who go in expecting autonomy and discover scoped tool use feel mis-sold, regardless of whether the underlying product is good, because the gap between the pitch and the delivery is too wide to ignore.
Decision-making at scale in regulated workflows. Putting an LLM inside a governed flow is a category error, and it is one that organisations are increasingly making because the pitches are compelling. The governed flow exists precisely to remove non-deterministic actors from decisions that need to be consistent, auditable, and verifiable. Replacing the human with an LLM is replacing the weakest link with a similarly weak link. The short version, which deserves its own treatment in a later piece, is that AI belongs over governed workflows, providing interfaces and inference at the edges, never inside them.
Replacement of expertise. AI as the doctor, the lawyer, the financial adviser. This is the AGI-around-the-corner framing, and it consistently overpromises. The closer reality is that AI handles the first ten minutes of those experts: preparing the brief, surfacing the precedent, drafting the outline. The expert continues to do the work that requires expertise, and the AI does the preparatory work that used to fall to junior staff. The shift is real and meaningful for how professional services scale, but it is not replacement.
The question this leaves about junior roles
The replacement-of-expertise pattern leads naturally into a conversation the industry shrugs off too easily, which is what happens to the people who used to do the preparatory work. Junior roles in professional services, in financial services, in legal work, in consulting, have always involved a mix of two different kinds of activity. There is low-value repetitive work that nobody particularly enjoyed doing, and there is apprenticeship work where the doing was how you learned to become senior. AI absorbs the first cleanly, which is mostly welcome. It absorbs the second at the cost of the development pipeline that produces senior expertise in the first place, and that cost is real even when it is not yet visible on a balance sheet.
The organisations that get this right will be the ones that recognise the trade and deliberately reinvest the time AI saves into the development work that has historically been crowded out by document review, meeting notes, and basic research: structured mentoring, real client exposure, deliberate reasoning practice, the kind of conversations between senior and junior staff that happen when there is time to have them. The ones that get it wrong will quietly hollow out their development pipeline and discover the cost in five years, when there is nobody senior coming through to replace the people currently doing the senior work.
This is one of the most important conversations an organisation deploying AI seriously should be having, and it is almost always missing from the deployment plan. The technology question is comparatively easy. The development pipeline question is what separates organisations that will have credible senior teams in 2031 from those that will not.
What this means for buyers
If a product pitch does not map cleanly to one of the six patterns, treat the pitch with scepticism. Most useful AI deployments are unsexy, in the sense that they synthesise, advise, draft, and flag, rather than autonomously running your business. The vendors selling autonomy are either misunderstanding what their own product does or hoping you will, and in either case the gap between what you bought and what you received tends to be larger than anyone planned for.
The harder question to ask is what you have already bought. Most organisations that deployed AI heavily in the last eighteen months bought it on the autonomy story and are now quietly disappointed by what arrived. Mapping what you actually have against the six patterns is a useful diagnostic. The products that sit cleanly inside one pattern tend to be delivering value. The ones that were sold as something more ambitious tend to be the source of the quiet disappointment, and the conversation about what to do with them is one most senior teams have not yet sat down to have.
The provocation
AI delivers scaling benefit on cognitive work. It does not replace humans at decision points. The patterns that work reflect this. The patterns sold as exceptions to it tend to fail in production, regardless of how they performed in the demo.
The question worth sitting with: map the AI products your organisation has bought or is evaluating against the six patterns above. How many sit cleanly inside one? How many were sold as something else? And if the answer to the second question is more than you would like, what does that tell you about where the value will actually come from, how much of your current AI spend is going to deliver on what it promised, and what conversation you need to be having with your team about the development pipeline you may be hollowing out without meaning to?
MultipleWorks helps organisations diagnose where their AI investment is actually delivering value, where it is buying theatre, and what to do about the harder questions AI raises about how senior expertise gets developed. If this article reflects a conversation you have been having internally, get in touch at hello@multipleworks.com.hk.