From asking to orchestrating: four stages of AI adoption

Ask is conversational AI. You type a question, the model answers, and you decide what to do with it. This is where prompting matters, and where most organisations stop.

The gains are real but limited. For an individual, I put the gain from chat tools at around 20% faster, and in some cases 50% faster. The catch is that the gain lives inside one person’s day. Nothing changes in the process, and nothing is left behind when that person goes on leave.

The test for this stage is simple. Are your people asking good questions, and do they know where the model is strong and where it is weak?

Automate means handing a whole, repetitive process with a real thinking load to AI, with a human checking the output.

The example is the quotation process at a small aluminium-casting manufacturer. Clients sent 2D drawings. The engineering team met for 30 to 45 minutes to list queries, waited for the client’s answers, then met again to price. Each quote took two to three weeks. Of 100 to 150 requests a month, around five became orders. No one was laid off. Each quote now has one owner, and there is no meeting. Today AI agents run the quotation process, and the CEO still reviews every quote before it goes out. Cycle time dropped from weeks to days. I tell the longer story of that case in a separate piece.

The key idea at this stage is that you do not need to teach the model your industry from scratch. I don’t have to teach the foundation model manufacturing all over again. A foundation model is the large, general-purpose model, such as Claude or GPT, trained on a vast amount of text. It already knows good practice in most fields. Your job is to show it your process.

Automation hands AI a process. Delegation hands it a goal and the rules, and lets it work.

Delegate is where agents act on their own inside limits a human sets: budgets, thresholds, approvals. They do the work, and they stop and ask before crossing a line.

My example is RR Mart, a rural supermarket chain of about 28 stores where I sit on the board. Each store had one person and a helper. When stock ran low they sent a WhatsApp message, and two people at head office read every message and arranged supply, around the clock. There was no analytics team, and the chief executive did not want to build one.

What replaced that reads the store ERP (the software that records sales and stock) and the store WhatsApp group, reconciles the two, and runs about 50 agents in the background. An agent here is simply an AI given one role and its own instructions: supply chain manager, fleet manager, consumer success. Between them they produce a dead-stock list, a near-expiry radar for the next 30 days, supplier scorecards and a van route plan that is handed to the driver. The agents disagree, a supervisor agent settles it, and the disagreement is logged.

What makes this delegation rather than automation is that nobody tells the agents which stores to look at or which supplier to rate. They are given the goal, keep stock moving and keep it fresh, and the rules they have to work inside. What makes it delegation rather than the stage above is where it stops: it briefs a small head office that still decides.

The Google Ads system I built with my class follows the same rule from the other end. A team of agents researches keywords, writes the copy and prepares the campaign. The agent never clicks Publish. A human does.

Delegate is not about removing people. It is about deciding, deliberately, which decisions need a person and which do not.

Delegation gives a team of agents a goal. Orchestration gives them an organisation.

The case here is Helm, which I built to run marketing and e-commerce for Grashya, a small FMCG brand I operate myself. It is a department expressed as software: 73 agent roles across 16 departments and 131 defined action types, each with its own form, procedure, approval gate and outputs. You add a brand, connect its accounts, set a budget and a level of freedom, and the team plans, writes, publishes, prices and measures.

Helm did not start as Helm. The earlier version was GrowthForge, a marketing system that planned campaigns, wrote the posts, published them and ran paid promotions, and that could spend a set daily budget on its own before it had to come back and ask. It worked, and it showed me the ceiling: one system doing marketing well is still one system. Helm is what it became when the whole department, rather than the campaign, was written down as roles, rules and a ledger.

Two rules were fixed before any code was written. Every action, decision and cost goes on record, so nothing happens off the books. And spending is limited by code, not by a polite request: a prompt can be argued with, a limit enforced by the program cannot.

That is the real difference from the stage below it. RR Mart is a team of agents doing one job well. Helm has an org chart, a rulebook and a ledger. Roles argue with each other and the arguments are kept. Twenty-seven runs are on the books, 39.8 million input tokens, $41.70 of recorded cost. What the business learns stays in the system rather than in the memory of whoever ran it.

I should say where it stops, because this stage is the easiest of the four to oversell. Helm plans, debates and records as designed. Its execution layer never ran end to end, and on my own review the web app was a notebook rather than a cockpit. The shape of the stage is right. The finished thing is not there yet.

The same idea now extends the quotation case from Stage 2. The quotes that did not convert are fed back so the agents can learn what the lost ones had in common. That is capability compounding, and it is the point of the stage: the organisation stops relearning the same lesson.

A quick self-check, one question per stage:

• Ask: Do staff use AI daily, and do they know where it is weak?
• Automate: Can you name one repetitive process, with an owner, that AI now runs end to end with a human check?
• Delegate: Have you written down the thresholds (money, customers, reputation) above which a human must approve?
• Orchestrate: Does a whole function run as an organisation of agents, with its budget enforced in code and the reasoning behind each decision kept where the next person can read it?
• Ownership: Is one leader accountable for moving the business from one stage to the next?

Many firms in Dubai and across the Gulf run on WhatsApp groups, Excel sheets and the memory of a few long-serving staff. That is exactly where RR Mart started. When what the business knows sits with two people who read every message, it is one resignation away from losing it.

Delegation is what moves that knowledge out of their heads and into a system that works to rules. Orchestration is what keeps it: the record of what was decided and why, in a form the next person can read.

One practical note from the evening: Automate is something you can learn and try on your own. Delegate, Orchestrate and running your own on-premise models have a much steeper learning curve.

The thread running through all four stages is a change in how managers think. In management we talk about problem solving. In the AI world it is more about architecting a solution. The question shifts from “how do I fix this?” to “can I design this so the problem never comes back?”

That shift in mindset, more than any single tool, is what separates an organisation that uses AI from one that has changed because of it.