Skip to main content
Agentic Architecture Skills.
Start hereUse casesArchitectureIntelligence ladderKnowledge & memoryResearchSkills
Operate

How to organise the teams: One Platform, Three Lanes

Who owns the agent platform, who owns each agent and who supervises it: a thin technology office, one platform team, three lanes that build each pattern once and step back, and business teams that own their agents.

By Murali SidLinkedInPublished

Once agents move past a pilot, three questions come up. Who owns the platform the agents run on? Who owns each agent? Who supervises its work? This guide's answer is one platform and three lanes. A thin technology office writes the rules and holds the money. One platform team turns the rules into automatic checks. Three small lanes, for AI, cloud and integration, build each new pattern once and then step back. Business teams own their agents, and their own staff supervise them. Almost every agent needs AI, cloud and integration. So one model covers all three.

The four units

The structure has four units. Each owns a clear part of the work.

The sizes are this guide's judgement. Squad size follows Microsoft's guidance of product teams of six to ten [vendor].

Each part the platform owns is defined on the layer that holds it:

Among the seven capabilities, these are the tool gateway, identity and policy, and observability and audit trail. The registry record itself is set out on the agent lifecycle.

People approve only three things. They approve actions in the four deterministic zones, new patterns, and steps up in autonomy. Everything else passes on evidence from the pipeline.

The lanes coach and hand over. Gartner predicts 70 percent of enterprises will abandon agentic AI built by a vendor's engineers by 2028 (29 September 2026). The clients cannot run or change it alone. So delivery teams own what they build, agents included.

Who does what

Who supervises the agents

Exception supervisors are existing staff, not a new specialist team. In field evidence, supervisors who work inside the team did better than specialists. Size their capacity with the queueing calculation on operating model and supervision (R13). Plan for the busiest ten minutes, never a daily average.

The sponsor answers for the agent's results. The supervisor decides each case that reaches the queue. Neither role is a new job title. Both sit in the business team that owns the work.

Four governing bodies

The risk council copies Telstra's. It meets weekly and covers only risk that is specific to AI (UTS Human Technology Institute, May 2026). The advice forum uses the advice process. Anyone may decide after seeking advice, and the decision is logged. Standard changes pass on pipeline evidence, with no change board. DORA links heavy change approval to worse delivery.

Funding

The office, the platform base and the lanes get a central budget. Charging for the base pushes teams to go around it. That grows shadow AI: AI used outside the governed path (Tian Pan, 2 July 2026).

Tokens, compute and gateway calls are shown back to each business unit and agent from the first day. Chargeback starts in year two. Each agent has an agent budget owned by its sponsor, with hard caps per run. Circuit breakers stop spending at set limits, as EY recommends. Lane help is free for six weeks. After that, the business unit pays. First savings fund the next use cases and the platform, as BCG describes. No firm publishes a chargeback model, so these terms are this guide's judgement.

One front door and three support levels

All work enters one intake and one register. Triage asks five yes-or-no questions.

  1. Does it touch a deterministic zone?
  2. Does an agent write, not just advise?
  3. Is it customer-facing, or on regulated data?
  4. Is it off the standard path?
  5. Is it a new pattern?

The answers set the support level. A support level is how much help a team gets. It is separate from the agent's risk tier. The risk profile sets that tier. Triage stays yes or no. Telstra triages with yes-or-no questions and removed its medium risk tier.

When a lane steps back

The failure to avoid is a central team that becomes a queue. Tian Pan's warning signs are rising approval times and frequent exceptions. Others are stale standards, growing shadow use and senior staff leaving. No major firm publishes a test for when a central team should end. This guide sets five. A lane hands a pattern to self-serve when all five hold for two quarters.

  1. At least 70 percent of the lane's new work goes through self-serve.
  2. Reviews find a material issue in under 1 in 20 cases. Blind samples confirm it. A falling find rate can mean tired reviewers.
  3. Every business team has one person who passed the lane's practical test.
  4. No serious incident, graded severity 1 or 2, is traced to a lane pattern.
  5. The median wait at the lane's gates is under two business days.

If a test fails for a quarter after handover, the pattern returns to the lane. Gregor Hohpe calls this work bouncing back to the centre.

Cloud likely steps back first. Its move to a platform is the most mature. AI goes second, once evaluations and gateways are platform features. Integration goes last. With agents, integration becomes the governance of tools.

Some parts never dissolve. They are the office, the policy owner, the registry owner, the platform team and the zone review. This is the autonomy contract applied to the organisation.

Centre of excellence or enterprise architecture

The evidence shows no move back to heavy architecture methods such as TOGAF. The Open Group's recent output is an AI assistant and a conference, not guidance on operating models. Centres of excellence survive, but as a stage. The label is fading into studio, factory, hub and platform team. Analysts want a lighter architecture function that governs AI. Forrester describes the architect as the one who sets the bounds for autonomous AI (18 September 2026).

The verdict is neither label. Architecture moves from reviewing designs to owning the rules and the record of what runs. That record includes the agent registry. Architecture owns the policy within identity and policy. It owns the context within knowledge and memory. The platform owns enforcement and the audit trail. Start with a central team if you need one. Write its exit tests on the first day.

A worked example: billing exceptions at a water utility

Take a water utility that adds an agent to clear billing exceptions. Today, billing operations runs this work with team leads and QA sampling. The billing operations manager becomes the agent's sponsor. The billing team owns the agent and its evaluation set. Team leads already sample released bills. They work the agent's exception queue.

Releasing a bill is a zone action. So triage answers yes. The support level is reviewed, and the risk council looks at the zone work. A named approver signs off the rule that releases bills. The platform supplies the gateway, the registry record, the agent's identity and the cost meter. The design is on resolve billing exceptions.

The same change in a hospital network can trigger consultation duties under enterprise agreements. Those duties apply when admin roles change. Plan that time before the agent goes live.

Evidence and limits

The shape is well supported. Microsoft treats an AI centre of excellence as a stage that hands delivery to platform teams (26 June 2026) [vendor]. AWS ends its cloud centre of excellence phases in a supporting role [vendor]. BCG keeps a central hub that monitors, enables and governs (25 August 2026). DORA ties results to platform quality, not to a team label.

Every size, threshold, time period and funding term here is this guide's judgement. No firm publishes a chargeback model or exit tests for a central AI team. No controlled study links any team structure to results. Most outcome figures in this field are self-reported by firms that sell the service. The Gartner figure is a prediction, not a measurement. The Telstra case is one organisation's account.

Microsoft's centre of excellence guidance changes often. Check it again each quarter.

Next step

Check where you stand with the readiness assessment. Then run the weekly loop on Operate.

Explore this topic

Cite this page

Written by Murali Sid, Founder and AI Systems Architect, last updated 15 October 2026. Licensed CC BY-SA 4.0, so quote it freely with attribution.

Murali Sid (2026). How to organise the teams: One Platform, Three Lanes. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/operate/organisation

Machine-readable version of this page: /operate/organisation.md

Operate: the weekly loop

The four checks to run every week once an agent is live, the measures the weekly review reads, and the gate before the agent may do more.

On this page

The four unitsWho does whatWho supervises the agentsFour governing bodiesFundingOne front door and three support levelsWhen a lane steps backCentre of excellence or enterprise architectureA worked example: billing exceptions at a water utilityEvidence and limitsNext step