Phase 3: agents and orchestration
Phase 3 finishes each department's high-value automations, then brings its agents together under one orchestrator agent as a multi-agent system. It comes last because it depends on everything before it, and people still make every key decision.
AI agents and multi-agent systems are the right step once a department trusts its brain and its first agents, and its remaining repetitive work is a chain of steps. Phase 3 automates those chains, then adds one orchestrator agent per department to coordinate its agents. Orchestration stays inside a department, and people keep every key decision.
When are AI agents and multi-agent systems the right step?
AI agents and multi-agent systems are the right step when a department already trusts its brain and its first agents, and the repetitive work left is a chain of steps rather than a single task. Before that point they automate confusion, and an assistant and a person do the job better.
Phase 3 is the last build phase of the AI-Native Roadmap and comes after Phase 2. It has two parts: finish the department's remaining high-value automations, then bring its agents together under one orchestrator agent.
Phase 3 completes the picture the roadmap describes on every page: An AI-native business is one where AI assistants, department brains and agents do the repetitive work in every team, with people keeping the judgement.
What are the last high-value automations?
The last automations are the multi-step jobs a department still runs by hand from start to finish. Each needs several agents and systems, which is why they wait until Phase 3.
| Department | The chain the agents run | What a person still decides |
|---|---|---|
| Sales | Enquiry to tailored proposal: research the client, draft the proposal, prepare the pricing options | The offer, the price and whether to bid |
| Customer support | Routine request to resolution: identify the customer, check the order, fix or rebook, confirm | Refunds outside policy and every complaint |
| Finance | Month-end close: match transactions, chase missing receipts, draft the reconciliation | Sign-off of the accounts |
| Operations | Supplier onboarding: collect documents, check them, set up the record, request approval | Whether to work with the supplier |
| HR | New starter: issue the contract draft, set up accounts, book the induction, send the first week's plan | Hiring decisions and anything about pay |
| Programmes (charities) | Referral to allocation: log the referral, check eligibility, match to a caseworker, draft the first contact | Who is supported and how |
| Admissions (schools) | Enquiry to visit: answer questions, book the tour, send the prospectus, log the family | Every admissions decision |
These differ from Phase 2's agents, which each do one thing: analyse, enrich or monitor. Phase 3 joins steps into a chain, so a job moves from start to finish with a person approving at the points that matter.
What is an orchestrator agent, and what is it not?
An orchestrator agent coordinates a department's other agents: it breaks a job into steps, hands each to the right agent, checks what comes back and assembles the result. It does no work itself.
- It is not a manager of people. It directs software, and people direct it.
- It is not a decision-maker. Anything strategic or customer-facing goes to a person.
- It is not all-seeing. It has its own narrow permissions, separate from the agents it coordinates.
- It is not a black box. Every step it assigns, and every result, is logged for review.
An orchestrator agent coordinates one department's agents and does no work itself, so each agent stays simple enough to test. A failure can be traced to one step, rather than buried inside one large system.
What does a department look like as a multi-agent system?
A department run as a multi-agent system, an agent team, has a handful of agents with narrow jobs, one orchestrator agent coordinating them, a department brain they all read from and people at every decision point.
Picture a charity's fundraising department. A research agent prepares a briefing on each trust, a drafting agent writes the first version of the application, a checking agent compares it with the trust's criteria and a records agent logs it. The orchestrator agent runs the sequence. The fundraiser decides what to apply for and signs off every word.
Hartz AI's investment committee demonstration shows the shape on a property deal: seven specialist agents pressure-test it and draft the paperwork behind a recommendation. It is scripted to show the reasoning and uses illustrative figures. In a real build a person signs off the call. The Property AI Lab has more concepts like it.
Agents are early, and much sold as agentic is not. Gartner (opens in a new tab) predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls, and estimates only about 130 of the thousands of agentic AI vendors are real.
Smaller organisations are moving more slowly than large ones. McKinsey's 2026 global survey (opens in a new tab) found that 22% of respondents from organisations with revenues of $1 billion or less report scaling AI agents, flat on the year before, against 40% from organisations above that size.
Why does orchestration stay inside one department?
Orchestration stays inside one department because that is where the data, the ownership and the accountability are clear. One person owns the process, one brain holds the information and one team answers for the result.
Across departments those lines blur. A sales agent that rebooks a delivery is now acting on operations' systems, under nobody's clear authority. Phase 3 at Hartz AI stops at the department so that every agent has an owner who can explain what it did.
Regulation is still forming here. The Information Commissioner's Office (opens in a new tab) expects to publish its final guidance on agentic AI in spring 2027, after a call for evidence (opens in a new tab) in autumn 2026. Building agents inside a department, with clear owners and logs, keeps you well placed whatever it says.
Where does the roadmap lead after Phase 3?
Beyond Phase 3 lies organisation-wide orchestration, where agents coordinate across departments rather than within one. It is where the roadmap leads, not a stage Hartz AI delivers today.
Getting there safely needs things few organisations have yet: shared data standards across teams, clear authority for agents that cross boundaries and regulation that has settled. Until then, each department runs its own multi-agent system, and people coordinate between them.
Phase 3 never ends the work either. Managing AI over time covers the checks, reviews and model swaps every live system needs, and ongoing advisory keeps a senior eye on them.
What does a person still decide in Phase 3?
People make every strategic and customer-facing decision, even in a fully orchestrated department. The agents prepare, check and carry out; people choose, approve and take responsibility.
At a grassroots sports body, a multi-agent system might run the whole registration season: chasing forms, checking eligibility and issuing confirmations. Disputes, safeguarding concerns and exceptions to the rules go to the volunteers who run the league.
Agents and orchestration are mostly bespoke work. The people who design and build them are on the Hartz AI team page, and bespoke AI solutions explains how that work is scoped, tested and handed over.
Common questions
Frequently Asked Questions
Find out whether agents are the right next step
A discovery call checks whether a department is ready for agents, and which multi-step job would prove them.